Human-smart rollator interaction for gait analysis and fall prevention using learning methods and the i-Walker

Human-smart rollator interaction for gait analysis and fall prevention using learning methods and the i-Walker Atia Cortés Martínez ADVERTIMENT La consulta d’aquesta tesi queda condicionada a l’acceptació de les següents condicions d'ús: La difusió d’aquesta tesi per mitjà del repositori institucional UPCommons (http://upcommons.upc.edu/tesis) i el repositori cooperatiu TDX (http://www.tdx.cat/) ha estat autoritzada pels titulars dels drets de propietat intel·lectual únicament per a usos privats emmarcats en activitats d’investigació i docència. No s’autoritza la seva reproducció amb finalitats de lucre ni la seva difusió i posada a disposició des d’un lloc aliè al servei UPCommons o TDX. No s’autoritza la presentació del seu contingut en una finestra o marc aliè a UPCommons (framing). Aquesta reserva de drets afecta tant al resum de presentació de la tesi com als seus continguts. En la utilització o cita de parts de la tesi és obligat indicar el nom de la persona autora. ADVERTENCIA La consulta de esta tesis queda condicionada a la aceptación de las siguientes condiciones de uso: La difusión de esta tesis por medio del repositorio institucional UPCommons (http://upcommons.upc.edu/tesis) y el repositorio cooperativo TDR (http://www.tdx.cat/?localeattribute=es) ha sido autorizada por los titulares de los derechos de propiedad intelectual únicamente para usos privados enmarcados en actividades de investigación y docencia. No se autoriza su reproducción con finalidades de lucro ni su difusión y puesta a disposición desde un sitio ajeno al servicio UPCommons No se autoriza la presentación de su contenido en una ventana o marco ajeno a UPCommons (framing). Esta reserva de derechos afecta tanto al resumen de presentación de la tesis como a sus contenidos. En la utilización o cita de partes de la tesis es obligado indicar el nombre de la persona autora. WARNING On having consulted this thesis you’re accepting the following use conditions: Spreading this thesis by the institutional repository UPCommons (http://upcommons.upc.edu/tesis) and the cooperative repository TDX (http://www.tdx.cat/?localeattribute=en) has been authorized by the titular of the intellectual property rights only for private uses placed in investigation and teaching activities. Reproduction with lucrative aims is not authorized neither its spreading nor availability from a site foreign to the UPCommons service. Introducing its content in a window or frame foreign to the UPCommons service is not authorized (framing). These rights affect to the presentation summary of the thesis as well as to its contents. In the using or citation of parts of the thesis it’s obliged to indicate the name of the author. Human - Smart Rollator Interaction for Gait Analysis and Fall Prevention Using Learning Methods and the i-Walker Atia Cort´ es-Mart´ ınez CS Universitat Polit` ecnica de Catalunya - BarcelonaTECH A thesis proposal submitted for the degree of Ph.D. in Artificial Intelligence 2018 2 Abstract The ability to walk is typically related to several bio-mechanical components that are involved in the gait cycle (or stride), including free mobility of joints, particularly in the legs; coordination of muscle activity in terms of timing and intensity; and normal sensory input, such as vision and vestibular system. A walk is composed of the stance and swing phases. The faster we walk, the shorter the stance phase will be. Thus, gait requires input from the brain, spinal cord, peripheral nerves, muscular power and joint and cardiovascular health. Because all of these systems are required to coordinate gait, gait speed is an indicator of the health of many physiological systems. At the same time, a relation between gait and cognition has been widely analysed from the medical point of view, and we can find several reviews in the literature. As people age, they tend to slow their gait speed, and their balance is also affected. Also, the retirement from the working life and the consequent reduction of physical and social activity contribute to the increased incidence of falls in older adults. Moreover, older adults suffer different kinds of cognitive decline, such as dementia or attention problems, which also accentuate gait disorders and its consequences. Assistive technologies (AT) play a key role in today’s’ society, especially when it comes to the older adults. ATs have enabled improvements in their Quality of Life, extending their autonomy and community living. This is important, as they can stay active safely and independently. During the last decade, research has focused on developing ATs with a sensor system integrated with the device or located in the human body. Efforts are focused especially on mobility assistance for different targets of people (visual impairment, frailty, chronic diseases or rehabilitation) and activity recognition, which could be used, for instance, to monitor elderly population living in autonomy and community-dwelling. This thesis proposes a methodology to analyse how do older adults at high risk of falling interact with a smart rollator, the i-Walker, to navigate in indoor, flat environments. The iWalker is equipped with a set of sensors and actuators and can collect data for long periods of time (several hours). It has already been tested in post-stroke rehabilitation and fall prevention clinical trials with successful results. In this work, we present results on our approach from a narrative perspective. Results are promising since we can relate the data obtained from human-rollator interaction to clinical parameters. The machine learning approach uses the data obtained with the force sensors of the i-Walker based on the interaction of individuals of different ages and health conditions. The analysis complements our extracted gait parameters with biological and clinical data to learn new characteristics of human gait at a stride-to-stride level. i We believe that users, caregivers and clinicians would benefit from the new knowledge that the i-Walker can generate from this work. ii Contents 1 Motivation 1 1.1 Scope of the thesis ................................ 3 1.2 Plan of the Work ................................. 6 2 A review of Gait, Cognition and Falls 9 2.1 Gait characteristics in elderly population ..................... 10 2.2 Ageing and Falls ................................. 13 3 A Review on Assistive Technologies 17 3.1 Assistive Devices ................................. 18 3.2 SHARE-it ..................................... 22 3.2.1 CARMEN: an ARW with collaborative control ............. 24 3.3 I-DONT-FALL ................................... 28 3.3.1 Fall Management Service ........................ 29 3.3.2 IDF components ............................. 30 3.3.3 The IDF protocol ............................. 31 4 The i-Walker 35 4.1 Main components ................................. 37 4.2 Reactive control .................................. 40 4.2.1 Applying the collaborative control philosophy to the i-Walker ..... 43 4.3 The i-Walker’s Assistive Environment ...................... 43 4.3.1 The role of the i-Walker in post-stroke rehabilitation .......... 44 4.3.2 I-DONT-FALL Results .......................... 46 4.3.3 Detecting Walking Behaviour Patterns .................. 47 4.4 Summary ..................................... 53 iii CONTENTS 5 Clinical Tests: design and implementation of a pilot protocol 57 5.1 Definition of a protocol .............................. 58 5.2 Protocol design .................................. 59 5.2.1 Baseline Pilot ............................... 59 5.2.2 Target Population ............................. 60 5.2.3 Clinical Scales .............................. 61 5.2.4 Ambulatory Exercises .......................... 63 5.2.4.1 Ten Meter Walking Test .................... 66 5.2.4.2 Timed Walking Tests ..................... 68 5.3 Pilots ....................................... 69 5.3.1 IDF Pilot ................................. 69 5.3.2 MAD Pilot ................................ 71 5.3.3 CVI Pilot ................................. 72 5.4 Summary ..................................... 72 6 Methodology 75 6.1 Gait Analysis based on Human-Rollator Interaction ............... 77 6.1.1 Data preparation ............................. 78 6.1.2 Vocabulary of strides ........................... 80 6.1.3 Clustering Time Series .......................... 83 6.1.4 Exercises as bags-of-strides ....................... 86 6.1.5 Cluster stability .............................. 87 6.2 Spatio-temporal Analysis ............................. 88 6.2.1 Descriptive Gait Parameters ....................... 88 6.2.2 Gait Velocity ............................... 89 6.3 User Driving Skills ................................ 90 6.3.1 Laterality ................................. 91 6.3.2 Directivity ................................ 93 6.4 Modelling Exercises by Spatio-Temporal Gait Characteristics .......... 94 6.5 Modelling Fall Risk Assessment ......................... 96 6.6 Summary ..................................... 97 iv CONTENTS 7 Results 99 7.1 SpatioTemporal Analysis ............................. 99 7.1.1 Descriptive Gait Parameters .......................100 7.1.2 Gait Velocity ...............................101 7.2 Clustering Results of the Gait Analysis ......................108 7.2.1 IDF +MAD pilots ............................108 7.2.2 CVI pilot .................................113 7.2.2.1 First Scenario .........................114 7.2.2.2 Second Scenario ........................115 7.2.2.3 Third Scenario .........................119 7.3 CVI Cluster Explanation from the Spatio-Temporal Gait Characteristics ....127 7.3.1 First Scenario ...............................128 7.3.2 Second Scenario .............................129 7.3.3 Third scenario ..............................131 7.4 Modeling Fall Risk ................................134 8 Conclusions 137 8.1 Discussion .....................................139 A Pilot Protocol 143 A.1 Fondazione Santa Lucia ..............................143 A.2 Residencia Los Nogales ..............................146 B Integrating the i-Walker as an intelligent service in a Social Network 149 B.1 Architecture ....................................150 B.2 Social Network (SN) ...............................151 B.2.1 People ...................................152 B.2.2 Devices, Reports and Messages .....................153 B.2.3 Multi-Agent System ...........................153 B.2.4 Integration ................................155 B.3 Service Implementation ..............................156 v CONTENTS C Research Activity 159 C.1 European Projects .................................159 C.2 National Projects .................................160 C.3 Participation in research courses and/or seminaries ...............160 C.4 Participation in conferences ............................161 C.5 Publications ....................................162 C.6 Research stays and visits .............................163 References 164 vi Chapter 1 Motivation Demographic ageing proceeds apace in all world regions, more rapidly than first anticipated in Nations (2003). The proportion of older people fastly increases as mortality falls and life expectancy increases. Population growth slows as fertility declines to replacement levels. Latin America, China and India are experiencing unprecedentedly rapid demographic ageing. The proportion of the population aged 65 and over is expected to triple in less developed countries over the next 40 years, rising from 5.8 to 15% of the total population, while in the more developed countries this figure is expected to grow from 16 to 26% (an increase of more than 60%), the ISSA report says (Scardino (2009)). In nowadays ageing society, many people require appropriated and personalised assistance and new technologies to offer them an extraordinary opportunity to perform their activities of daily living (ADL) and improve their autonomy. A demographic study conducted by Brault (2010) showed that 56.7M people from the US (18.7% of the population) had some level of disability and 38.3 million (12.6%) had a severe impairment. Older studies by Brault (2005) and Brault (2000) show that these numbers are steadily increasing year after year, but also that the majority of this population is concentrated on more older adults. Of people aged 15 and older, 30.6 Million (12.6%) had difficulty with ambulatory activities of the lower body and 15.2M people (6.3%) had trouble with cognitive, mental or emotional functioning. In the case of the EU25, in 2011 there were more than 80 million people with a disability in the population age group of 16-54 years, and it is estimated that this number increases up to 84 to 107 million people in the European Union (de Pejil et al. (2011)). Of all world regions, Europe has the highest proportion of the population aged 65 or over, a statistic that becomes more pessimistic according to the baseline projection of Eurostat, which shows that 1 1. MOTIVATION this percentage will almost double to more than 25% in the year 2050 (WHO (2012). Besides, life expectancy has continued to rise systematically in all of the EU Member States in recent decades (Kotzeva (2015)). Besides, in this population sector, the frequency of falls increases with age and frailty level and are the leading cause of unintentional injury (WHO (2007)). A combination of biological factors and disease-related conditions are the primary cause of most falls among seniors. This combination has several implications for the Quality of Life (QoL) of the elderly population: as they reduce their activity, they increase their frailty and fear of falling while losing their residual skills. This will represent a challenge for the public health systems that will have to face a substantial socio-economic impact to deal with this demographic situation. This is already not sustainable in some countries and will be a worldwide issue shortly. One of the primary objectives of the H2020 program is to focus on the analysis of the causes and consequences of pathologies to find patterns that will support early detection of a disease or associated risks. Consequently, the care community could take decisions on intervention and educational information to delay the physical or cognitive decline of the elderly and try to keep them independent as long as possible living in the community. The evolution of ICT tools (regarding cost, size or availability) in collaboration with medical knowledge has empowered the design and development of innovative solutions to provide tailored, remote and preventive care of people with special needs. In particular, there is an increasing interest in ambient assisted living technology, where individuals (in this case, elderly non-autonomous persons) and their environment are equipped with a system of sensors (from localization to bio-metric, among others) that will collect different types of measures allowing experts to monitor their activities in real-time, or by reports generated by the system. Ambient assisted living environments are expected to gain particular relevance with the incoming Internet of Things paradigm as a tailored, cost-effective solution to improve health systems (Vermesan and Friess (2013)). Assistive technologies (AT) play a crucial role in the care of challenged individuals, such as older adults or people with physical and/or cognitive dysfunction. Their primary purpose is to maintain or improve individual’s functioning and independence in order to facilitate participation in the society and to enhance overall well-being (WHO (2014)). ATs aim to provide assistance to different sorts of target publics, including low vision devices, hearing aids, augmentative and alternative communication, and especially technologies involving the use of mobile platforms, such as canes, scooters, wheelchairs and rollators or prostheses, such as 2 1.1 Scope of the thesis artificial legs. Cane use has been prevalent among the elderly for years, followed by walkers (Brault (2000)). Lately, research has focused on the robotisation of these devices to assist persons with physical and/or cognitive disabilities in their activities of daily living (ADLs). Independent mobility is one of the most critical factors in maintaining the quality of life for elders, and other clinical populations who need assistive devices, by delaying their institutionalisation. Mobility is crucial for performing ADLs, as well as for maintaining fitness and vitality (Alwan et al. (2007)). While robotised wheelchairs have been mainly studied to assist in the mobility and autonomy of the user (see §3.1), smart walkers can go a step further according to Martins et al. (2012): they are useful to a broader range of users and help in the recovery of ambulatory skills. Recent studies on real individuals have tested the potential of robotised rollators in the rehabilitation process of people hospitalised due to some type of impairment, like a stroke recovery or a car accident (Giuliani et al. (2012) and Morone et al. (2016)). Thus smart walkers not only promote mobility and navigation assistance but also can provide gait monitoring and partial body weight. New trends in research and future model business solutions will combine ATs with different embedded or on-body (wearable) sensors that will provide benefits to different groups of interest: assistance to the end-user, monitoring to relatives or caregivers, activity or clinical reports to specialists. 1.1 Scope of the thesis This thesis work presents a new methodology for analysing the results of the interaction between a smart rollator equipped with different sensors, the i-Walker, and a group of older adults with physical dysfuntions due to the natural ageing progress or due to a recent fall. The i-Walker (see §4) is used to collect data during the execution of different walking tests and exercises performed by a group of volunteer participants from various centres and nationalities. Datasets will be completed with biological information from each individual, such as age, gender, the number of falls during the last year, as well as some cognitive and physical assessment measures. The i-Walker will be tested with two groups: a baseline of elder users performing a short walking test (10 Meter Walk Test) and a group of people from a wider age range performing a longer walking test (3 minutes Walk Test). It is difficult to establish a boundary that defines whether an elder is healthy or not but, since we work with in-patients from hospitals and residences, we must discriminate the target population with whom we want to work. As the test involves walking for several minutes, we 3 1. MOTIVATION seek for elders with ambulatory capabilities and good cognitive status, so they can understand and perform the exercises. Following the clinical advice of doctors and physiotherapists from the Fondazione Santa Lucia (FSL), we have defined a protocol with the inclusion and exclusion criteria along with a description of the test and expected outcomes. In §5we provide a detailed justification of the protocol that can also be consulted in Appendix §Afor the complete version. It is well-known that the walking ability of an individual can be affected by its cognitive and/or physical condition (Jahn et al. (2010)). One of the primary manifestations of gait disturbance in old age is the slow gait pace in comparison with the normal age-related slowing. Qualitative abnormalities of locomotion, such as disturbances in the initiation of locomotion or balance while walking are also indications for qualitative impairment of walking. Therefore, we need to identify which will be the most relevant variables from the human-robot interaction able to determine gait disturbance in a group of elderly individuals so that we can define a set of user profiles. We expect this information to be useful in two-folds: on the one hand, we aim to assess the quality of the information obtained from this human-robot interaction; on the other hand, we assume that this data will be useful to find some patterns on walking habits according to these user profiles. On future works, and with more tests, new control strategies could be developed to tailor the amount of help a user profile will need and improve the communication between the two parties. If we can find some patterns linking walking and medical parameters, the i-Walker could be used as an assistive device that is not only able to assist in elders mobility but is also able to diagnose a possible health decline by monitoring an individual’s activity, or at least to contribute as a decision support system to the clinician. In a first phase, we will use the i-Walker with elder individuals presenting different physical and cognitive conditions, separating initially between fallers and non-fallers (i.e. people that have fallen at least once during the last year). We have also collected other personal characteristics such as the presence of neurodegenerative diseases or prosthesis situated in legs and whether the person is using a traditional assistive walking device. Once the analysis is done, we will test the i-Walker again with a new set of users and check if it is possible to correctly classify the income of new data and give the appropriate assistance. We expect the results of this analysis could help clinicians in the diagnosis of cognitive decline and personalise the aid that the i-Walker could offer to the end-user. As we will see later, along with the text, the research on smart walkers during the last years has been mainly focused on studying the mobility of healthy young people (or challenged healthy individuals) and, in a lower proportion, elderly or blind individuals. Little work has 4 1.1 Scope of the thesis been done in the area of people with physical and/or cognitive disabilities, that may include the old. Moreover, in general, the works presented are performed with a short number of participants, probably due to the difficulties found when asking for a protocol approval in hospitals and care centres. When we think about the opportunities that Assistive Technologies can provide to older adults, several questions or issues are raised regarding navigation support, human-robot interaction and user’s satisfactibility with the whole system. Taking into consideration the components and characteristics of the i-Walker (see §4), there were a set of problems that I considered important to take into account for the future design of an intelligent support system for the i-Walker: •Where are we? Where are we going? : Interpretation of user’s intentions The i-Walker is equipped with a set of onboard sensors (forces, odometry and laser readings) that could be used to study how does each person face different navigation scenarios and define user profiles according to their driving skills. This kind of solutions usually includes a known indoor environment and the agenda of activities of the user (Cort´ es et al. (2010)). •Are you in trouble? : Prediction of user’s intentions Depending on the navigation situation, a person may require an extra amount of help (e.g. when there are many obstacles around his/her path). However, it is important to assist the users only when necessary and with the appropriate amount, otherwise the user gets used to make a minimal effort, and s/he may eventually end up losing residual skills. •Was I helpful? : Getting user’s feedback Studies have shown that if a person does not feel attracted to the assistive device, they tend to stop using it (Martins et al. (2012)). Thus, if users are not satisfied with the assistance, provided by the device, they will tend to refuse it. If the i-Walker can interpret the user’s driving intentions, it could provide some guiding assistance by correcting directionality. In Urdiales (2012), a method for evaluating user’s disagreement is proposed to implement a collaborative control between the individual and the robot (see §3.2.1). •Do you understand me? : Human-Machine Interaction The sensors onboard could also be used to enhance the right interaction between the user and the i-Walker. State-of-the-art of the kind of alarm signs that have been used for blind 5 1. MOTIVATION people or post-stroke recoveries has to be done. When dealing with people with cognitive disabilities, we can not abuse the use of alarms because they might not be understood or remembered. Also, the interaction should be smooth by making these alarms discrete enough to avoid stressing the user. As a result, the i-Walker would be an assistive device providing physical and cognitive support to impaired and/or older adults. Safety and monitoring are included in the solution, making the i-Walker a potential aid that would allow a semi-autonomous life to persons with some disability, allowing them to live longer, with an acceptable QoL, in the community. However, this is the first time that the outcomes of the i-Walker are analysed and assessed from a technical point of view. Until now, the i-Walker has been only used as a mobility or rehabilitation aid used to be compared with traditional assistive devices, and the assessment was done regarding clinical scales regarding the user (see §4.3.1). Therefore, the primary objective of this PhD thesis has become to process the dataset obtained from sensor measurements to provide a methodology for data cleaning and preparation. Since the data has been collected in different sites, there is a need for unification and normalisation methods as a primary step. The second objective is to analyse the data to provide a service that can classify a user according to walking and medical parameters and assist him/her in safe navigation. We hypothesise that a fusion of sensors’ data with biological data will allow identifying characteristics of walking behaviour for different groups of individuals, in this case, elderly people with high risk of falling. We will base our analysis on the evaluation metrics that are traditionally used from a clinical, descriptive gait analysis. Taking leverage of the onboard sensors, we believe we will be able to provide information of the human-robot interaction at step or meter detail, instead of obtaining global metrics. Although other studies have analysed gait characteristics, few are the works found on studying walking behaviour when using a rollator. We expect to define different user profiles according to their driving and walking skills using the i-Walker but also related to their clinical condition. The i-Walker could be then used as a support tool for clinicians when making a diagnose of a new user. 1.2 Plan of the Work The organisation of this PhD Thesis is as follows. In §3a review on assistive devices, and more specifically smart walkers, is given. §2contains a review of gait analysis from a clinical perspective, from observation to the involvement of different sensors that will complete the 6 1.2 Plan of the Work medical assessment. We then provide a technical description of the i-Walker in §4, the smart walker that is used to perform tests with real users in this thesis, as well as the previous projects in which it has been involved. A full description of the clinical trials involved in this work is given in §5. The methodology used in this PhD proposal is described in §6and §7show the results obtained at the moment. Conclusions and Future work are presented in §8. A detailed version of the pilots’ protocol is given in Appendix §A. The Appendix §Ccontains the list of papers published so far, as well as the different conferences where I have assisted. Finally, Appendix §Bcontains an example of a social network managed by a multi-agent system where the i-Walker provides intelligent services to the user, relatives and doctors. Results on this PhD could bring a new level of information that could be used in mHealth solutions like the one proposed, generating activity reports, acting in dangerous situations and serve as a patient remote monitoring tool. For this PhD work, I have studied the gait characteristics that are identified in the clinical literature and the analysis performed by several Authors. I have proposed a methodology using AI-based techniques to translate these clinical concepts into the domain of the i-Walker, to learn how to characterise an individual’s walking behaviour while using a robotised assistive device. This perspective is complemented with a machine learning analysis that provides further knowledge on human motion body and how to categorise it. 7 1. MOTIVATION 8 Chapter 2 A review of Gait, Cognition and Falls The ability to walk normally is related to several bio-mechanical components involved in the gait cycle (also known as stride), including (i) free mobility of joints, particularly in the legs; (ii) coordination of muscle action in terms of timing and intensity; (iii) normal sensory input, such as vision and vestibular system (see Rubenstein (2006)). Thus, gait requires input from the brain, spinal cord, peripheral nerves, muscular power and joint and cardiovascular health. Because all of these systems are needed to coordinate gait, the individual’s walking speed is an indicator of the health of many physiological systems (see Fritz and Lusardi (2009) and Pahor (2006)). The relation between gait and cognition has been widely analysed from the medical point of view, and we can find several reviews in the literature (see Haggard et al. (2000), MonteroOdasso et al. (2012) and Rosso (2013)). As people age, they tend to slow their gait speed, and their balance is also affected. Also, the retirement from the working life and the consequent reduction of physical and social activity contribute to the increased incidence of falls in older adults. Moreover, older adults suffer different kinds of cognitive decline, such as dementia or attention problems, which also accentuate gait disorders and its consequences (Scherder et al. (2007), Yogev-Seligmann et al. (2007) and Plummer-D’Amato et al. (2012)). Also, current concepts in disablement emphasise the importance of identifying mobility impairments in ageing humans to enable timely intervention and, ultimately, prevent disability as stated by McGibbon et al. (2001). It is estimated that after age 70, 35% of the population present gait disorders due to different reasons related to cognitive and/or physical decline. Most of them associated with the ageing process, but other factors, such as education and lifestyle, are also influential (see 9 2. A REVIEW OF GAIT, COGNITION AND FALLS Yogev-Seligmann et al. (2007)). One of the most common and dramatic consequences of gait disorders is falling. Over a third of the population, aged 65+ years fall every year (50% for adults aged 80+ years). As a consequence, 4-15% of falls cause significant injuries, while 23-40% of injury-related deaths in older adults are due to a fall Organization (2015)). 2.1 Gait characteristics in elderly population The capability to get from one place to another and successfully reach the desired destination is essential to every animal and therefore for humans1. Walking is a fundamental part of everyday life and depends on balance, joint motion, endurance, and muscle strength (Graham et al. (2008)). Human locomotion has been studied for decades, although the perspective, as well as the tools used for measurement, have evolved to these days. This field of research includes all ages of the human being, but also to the animal domain. This chapter focuses on gait analysis as the study of the human walking, which analyses the body mechanics and the activity of the muscles involved in the walking process (Whittle (2007)). Data used in this field can both come from clinical assessments (physical or cognitive) and different sorts of measurement tools that will collect some gait characteristics. Gait analysis is a systematic technique for recognising negative deviations in the gait pattern and determining their reason and effects (Prakash et al. (2016)). As mentioned before, several conditions might affect the ability to walk, and it affects mainly to people when ageing. For these reasons, it is essential that clinicians regularly assess their gait to diagnose and plan optimal treatments for each situation. Although it is not possible to provide a general description of gait without including all the singularities given in each pathology, it is well accepted that a normal gait involves the locomotive action of the two legs, alternately to provide both support and propulsion, having always at least one foot in contact with the ground Whittle (2007). The walking activity is composed of the stance and swing phases. The faster we walk, the shorter the stance phase will be. A gait cycle is thus defined as the time interval between two successive occurrences of a repetitive event (e.g., the right leg initial contact with the ground). Figure 2.1 shows the different phases included in a gait cycle starting with the right leg (in grey colour), taken from Whittle (2007). Body equilibrium is a key factor to correctly alternate leg displacements and sustains the body weight. 1For the elderly, walking, standing up from a chair, turning, and leaning are necessary for independent mobility. Gait speed, chair rise time, and the ability to do tandem stance (standing with one foot in front of the other, which 10 Chapter 3 A Review on Assistive Technologies Older adults usually suffer from at least one health condition, from visual or auditive impairment, muscular weakness, to neuro-degenerative disease and others. Most of these situations lead to a decline in the locomotive functions, which is manifested as the difficulty to solve complex ambulatory situations (e.g., avoid obstacles or turns to left or right), or gait disorders (e.g., decreasing gait speed, unsteady gait). The combination of the mentioned factors increases the risk of falling in this population. Assistive Technologies are essential for the recovery or replacement of the mobility functions at all ages and enhance the autonomy and quality of life of patients and relatives. Through this section, I will briefly introduce some research works in the field of assistive technologies designed for the mobility recovery (e.g., exo-skeletons) and assistance (e.g., Autonomous Robotic Wheelchairs or canes). However, rollators and more specifically smart walkers are a potential tool to study, since it can be involved in solutions for both problems, reaching a higher number of possible end-users. This review is needed to understand and synthesise the background of this thesis. It also introduces the work performed within two EU funded projects: (i) SHARE-it, where different assistive technology was developed to enhance autonomy and quality of life of elderly people 1; and (ii) I-DONT-FALL, where the i-Walker is used as a rehabilitation tool for fall prevention. 1The KEMLG group coordinated the project at UPC. 17 3. A REVIEW ON ASSISTIVE TECHNOLOGIES 3.1 Assistive Devices It is interesting to observe the fast evolution of assistive mobility devices over the last two decades. Although we are used to seeing people around us using traditional assistive devices (and, more recently powered wheelchairs or scooters), solutions involving robotic technology and Artificial Intelligence are still under research development. However, the evolution of new technologies (e.g., diminishing in size and cost while growing potential) are helping to create many different solutions that will someday improve our way of living. Different robotised versions of each traditional device have been recently proposed, aiming to help in either diagnosis, mobility or rehabilitation. Assistive devices can be classified into two categories depending on the person’s level of mobility: alternative for people with the total or temporal incapacity of mobility and augmentative for people with remaining mobility capacities. In Martins et al. (2012), authors represent this classification as shown in Figure 3.1. In the field of alternative assistive devices, different models of Autonomous Robotic Wheelchairs (ARW) are proposed in the literature (Weston (Hillman et al. (2002)), Wheelesley (Yanco (1998))), CARMEN (Urdiales et al. (2011)). ARWs provide solutions for autonomous and assistive navigation, but those are normally restricted to people with the total incapacity of mobility. It is recommended that people with residual mobility skills avoid the use of wheelchairs during long periods of time, as it may lead to a loss of capabilities ( see Martins et al. (2012)). During the last decade research has focused on the development of robotized augmentative devices, such as rehabilitation tools for ambulatory-training for people suffering from musculoskeletal or neurological disorders like strokes or spinal injury (HapticWalker, (Schmidt et al. (2007)), KineAssist, (Patton et al. (2008)), LokoHelp, (Swinnen et al. (2010))). The disadvantage for this kind of devices is that they are used in-hospital and require the expertise of a clinician or physiotherapist that ensures that the tools are being correctly applied. One of the most important augmentative devices is the smart walker (see §4) because of its potential, not only in mobility but also in rehabilitation. It is well accepted that smart walkers offer enough weight balance, and thus help to raise self-confidence and autonomy to patients with locomotion problems. During the last decade, many solutions have been proposed, being equipped with different kinds of sensors to provide not only mobility but monitoring as well. Solutions also vary in the number of wheels, going from the two-wheeled models to rollator 18 3.1 Assistive Devices Figure 3.1: Categories of Assistive Devices (Martins et al. (2012)) walkers, with four wheels. Smart walkers are expected to present the following functionalities (Frizera-Neto et al. (2011)): •Physical support: smart rollators should provide better gait stability. •Sensorial assistance: smart rollators should also collect and process data from different onboard sensors to assist in navigation and increase security to the final user (useful for obstacle avoidance or fall prevention) •Cognitive assistance: users having problems related to memory or orientation may need some guidance and localisation system •Health monitoring: used to keep the medical history of the user 19 3. A REVIEW ON ASSISTIVE TECHNOLOGIES •Human-machine interface (HMI): directly or indirectly, HMIs are used to communicate with the user through a system of alerts, alarms and commands. Rollators allow the performance of a natural gait pattern during locomotion. However, they are also considered the most unstable version and the risk of falling while using it increases in the situations that require the full-body weight support of the user (Neto et al. (2015)). Regarding the physical aspect, most of the smart rollators found in the literature are not based on a traditional rollator, so there is a vast variation regarding physical designs. For instance, the number of wheels and handlers varies among the proposed solutions, but also the degree of assistance that each of them offer. There are also different applications of these smart walkers regarding sensorial assistance: each solution is equipped with different sets of sensors, usually focused on object detection or user localisation. Some of these solutions are designed to assist blind people in navigation (Yu et al. (2003b)). In Glover et al. (2004) a walker is provided with navigation guidance for older adults who are cognitive or mentally frail by learning people’s motion behaviours and providing directions through a touch-based interface. Smart walkers empower the mobility of the user, providing support while walking and increasing the confidence and safety perception during ambulation. Regarding driving assistance, there are three strategies of human-robot interaction that are usually found in the literature, depending on which one has a higher decision control. These strategies are more commonly developed for ARW systems since a disagreement in the driving decision shall not have major physical consequences for the end-user. However, in the case of smart walkers, a strong disagreement could lead the user to lose balance, provoking some injury or even a fall. Driving control strategies are described below: •Human full-control: The user takes control of the assistive device to move towards a given goal. This solution is not suitable for people presenting cognitive disabilities, as security issues can be monitored but the robot cannot take any decision. •Robot full-control: The robot is aware of the ADLs and the environment of the user, and takes all decisions but, as stated before, this may lead to a serious loss of users capabilities and frustration. •Shared (or collaborative) control: A combination of the previous strategies in which the robot will provide an amount of assistance according to user’s skills. The idea is to help only when necessary, so the user does not get used to the robot doing all the job. 20 3.1 Assistive Devices This last strategy is the most commonly used in the last decades Cowan et al. (2012) in the field of ARWs (see CARMEN Urdiales et al. (2011), Boy et al. (2002) and Carlson and Demiris (2010)). Research on smart walkers has been more focused on navigation assistance as a guide but not as a way of control. For instance, the EU FP7 funded project DALi developed a portable motion planning using a standard rollator equipped with a Kinect and a tablet to guide the user in crowded environments. The aim is to relieve the stress suffered by people with reduced cognitive or physical ability, especially older adults. The system does not actively assist on the navigation, but it provides a brake control when it detects that the user has deviated significantly from the objective. Some examples that have been developed in the last decade are: •PAMM (Spenko et al. (2006); Yu et al. (2003a)) is a smart robotic walker, designed at the Massachusetts Institute of Technology. It aims to provide support, guidance, and health monitoring to elderly users in order to delay to transition to nursing homes. The latest version of PAMM was based on a four-wheeled structure with two handler (a smart cane version with two wheels and one handler is has also been developed). It also contains a camera for localisation and obstacle avoidance. •COOL Aide, (Wasson et al. (2008)), built on a standard three-wheeled rollator. Force and moment sensors have been added to the handlers, as well as encoders in the wheels to obtain the position, velocity and heading. It also contains a Hokuyo laser to assist the user avoiding obstacles. In Huang et al. (2005), authors propose a shared control strategy to help users dealing with possible collisions and reaching short-term goals in predefined paths. •iWalker (Kulyukin et al. (2008)), developed collaboratively by the Carnegie Mellon University and the University of Pittsburg. It is based on a standard four-wheeled rollator equipped with encoders, RFID sensors and a laser to provide autonomous navigation and self-parking option. •Smart Walker (Wada et al. (2016)) have built a sitting-type walker, which is a solution half-way between the wheelchair, where people will not require any motion force to move, and the rollator, which some challenged adults will not be able to use due to the lack of support force in arms or legs. The Smart Walker is equipped with an active-caster driving system which will help the user moving is given situations. 21 3. A REVIEW ON ASSISTIVE TECHNOLOGIES •UFES (Neto et al. (2015)) was developed under a research project between the University of Espirito Santo (Brazil) and the Univerisity of San Juan (Argentina). It presents a stricture of 3 wheels with encoders and motors, inertial movement sensors and 3D force sensors placed on top of two forearm support platforms. UFES provides a control architecture which enables an emergency braking in unsafe situations. It also collects gait parameters with additional IMU sensors placed on the user’s body. The force sensors detect the guidance intentions •ASBgo++ (Alves et al. (2017)) is a four-wheeled motorized rollator with forearm support platforms, built at the Minho University (Portugal). It provides safety navigation control and information about the user gait pattern. It contains a joystick which captures users movement intentions (Martins et al. (2014)). •Cheng and Wu (2017) have developed a smart rollator with pressure sensors added to the handlers to capture user’s driving intention using a support vector machine and AdaBoost classifier to identify the movement vectors. One of the objectives of SHARE-it and also of I-DONT-FALL and FATE, the EU projects in which the i-Walker has been involved in, was to ensure that the assistive device was user friendly,i.e. a device the user is familiar with, so s/he will be more comfortable using it. The cost of production is also reduced as well, as we started from a standard frame with electrical components embedded instead of building the robot from scratch. A full description of the i-Walker is provided in Chapter §4. 3.2 SHARE-it Within the frame of SHARE-it EU, funded project FP6-0450881, at the Universitat Polit` ecnica de Catalunya, two PhD thesis have been defended by Cristian Barru´ e and Cristina Urdiales (see Barru´ e(2012) and Urdiales (2012) respectively) at the UPC Artificial Intelligence PhD programme. These theses introduced different solutions developing and applying a research 1SHARE-it was a three years project funded by the European Commission whose primary objective was to develop AT which enable older adults to live independently and with the high quality of life as long as possible. The aim was to create scalable, adaptive systems of add-ons to the sensor and AT, mainly focused on supporting autonomous mobility, so that they can be modularly integrated into an intelligent home environment to enhance the individual’s autonomy. 22 3.2 SHARE-it approach to improve the quality of life among individuals suffering some disabilities. That technology was further applied in other two EU funded projects I-DONT-FALL and FATE. The primary goal of the SHARE-it project (Cort´ es et al. (2010)) was to contribute to the development of the next generation of intelligent and semi-autonomous assistive devices for older persons and people with disabilities (both cognitive and motor). When it comes to ageing, usually diseases do not come alone. When losing mobility, people’s ability to be self-dependent in performing their ADLs decreases and they end up hospitalised or in a day-care institution, losing their social environment. If the person also has some cognitive disability, the possibilities of having an autonomous life go reduced. The objective of SHARE-it was to develop a scalable, adaptive system of components (i.e. sensors and ATs) integrated into an intelligent home to enhance individual’s autonomy and thus, its Quality of Life. Regarding the mobility aspect, four different assistive devices were developed and deployed in real environments and with real users, three ARWs (CARMEN (Urdiales et al. (2011)), Spherik (Mart´ ınez et al. (2005))and Rolland (Christian et al. (2008)) and a former version of the i-Walker (Annicchiarico et al. (2008)). The main purpose was to provide mobility assistance for a wide range of users with different capabilities and needs. We tested the mobile platforms in an Ambient Intelligence environment, a house equipped with a set of domotic capabilities and sensors. A multi-agent system controlled all the data collected and managed the interaction between the assistive devices and the intelligent house through a set of cognitive services. In 2009, I developed my Bachelor’s project (PFC) at the Fondazione Santa Lucia (FSL) within the SHARE-it project. The objective was to design and perform a benchmark with real in-patients to test the interaction between CARMEN and the intelligent home through a multiagent system. The agent controlling the wheelchair was responsible for deciding the amount of help that every user needed at any given time. A localisation system allowed the intelligent house to monitor the user’s movements; when the user was driving in a narrow space (like a corridor) or crossing a door. The two agents interacted to increase the amount of help, if required, and decrease the security distance between the wheelchair and walls to facilitate the manoeuvrability. We proved that the users were able to end up their tasks easier and their navigation was smoother thanks to the multi-agent system. Other agents should have been involved in the benchmark, but due to external delays, it was not possible to test it during my stay at FSL. However, they were included in the final tests 23 3. A REVIEW ON ASSISTIVE TECHNOLOGIES of the SHARE-it project and results were very promising as well. Users passed a usability and disagreement tests, and the feedback was also very positive. The successful results of this project led to the further development of the i-Walker, both regarding design (e.g., reduction on electric components, European homologation as the medical device) and services for mobility assistance and recovery. It has since then been involved in three EU funded projects (I-DONT-FALL,FATE and ASSAM) and two Spanish national projects (SiRC and RehAdapta) all of them related with the development of Assistive Technologies. Chapter §4contains a complete description of the i-Walker used in this thesis, along with the main contributions in research developed during the last years. 3.2.1 CARMEN: an ARW with collaborative control Within the SHARE-it project, Urdiales (2012) designed and built a robotised wheelchair, CARMEN, with a collaborative control that aimed to assist a person’s mobility. CARMEN can detect how much help a given user needs depending on his/her abilities and current condition and to provide the required help: not more, not less. The concept behind this approach is to avoid loss of residual skills due to excessive help but to provide, nevertheless, the required assistance to achieve mobility in everyday environments with a powered wheelchair. Following the medical team advice, as a security measure, the wheelchair could not be driven backwards. The wheelchair is only driven by a joystick which translates user’s commands into directionality and provides a navigation aid by adapting to the amount of help required by each user, according to their level of disability. A therapist previously determines the amount of assistance of every user. CARMEN was built to be driven in an intelligent, adapted home following an agenda of ADLs defined by a medical team. It was first tested with a set of in-patients of the FSL using a purely reactive control for benchmarking, which mainly avoids obstacles thanks to the information transmitted by a frontal Hokuyo laser. Users had to test CARMEN in different defined paths (always including a door or corridor crossing), first without any aid (as a conventional powered wheelchair) and then with the reactive control. The system gathered data about the user’s commands, its relative position in the environment and the objects (or potential obstacles) surrounding the wheelchair along with other variables, like the time required by each user to complete the path. This information was used to measure how well did the user perform a given task. Urdiales finally formalised the information into different task metrics, although the most relevant for the rest of this work where: 24 3.2 SHARE-it Figure 3.2: Vectors involved in motion command calculation. •directness: the user drives keeping the goal ahead •smoothness: the user’s navigation (interpreted through the joystick movements) presents sharp direction changes •safety: the user drives with a safe distance to obstacles A person’s efficiency performing a given task according to these metrics is then calculated. Figure 3.2).a depicts vectors involved in collaborative control (VR, VH and VC for robot, human and collaborative); Figure 3.2).b shows the angles involved in estimating smoothness (R), safety (G) and directness (B): local efficiency at a given location can be visually evaluated from its RGB colour. In addition to these metrics, users had to complete different tests regarding medical scales that measure cognitive and physical disabilities, or usability and disagreement questionnaires. Urdiales finally tested it with a collaborative control that gives driving control to the wheelchair when the user does not reach a minimum threshold of success in achieving the goal or avoiding an obstacle. Results were promising, as they showed that users were able to perform better their tasks (without the navigation aid, they were often unable even to accomplish them). The second phase of her work consisted of creating skill-based wheelchair navigation profiles. The objective of using all the data gathered from previous tests is to separate users in profiles according to their navigation performance and the medical scales to predict the amount 25 3. A REVIEW ON ASSISTIVE TECHNOLOGIES of help required at every moment instead of fixing it in advance. Urdiales used navigational information from both healthy people and persons with different kinds of disabilities. According to Minguez et al. (2004), in robot navigation, we can deal with six different situations to achieve collision avoidance in troublesome scenarios, depending on the level of safety of the robot within the environment (security zone), and the reachability of the goal (free walking area). Figure 3.31: •Safety criterion: High Safety (HS) or Low Safety (LS) represents the absence and presence of obstacles respectively. The first criterion is applicable to every situation. The following criteria correspond to HS situations. •Goal within the free walking area criterion: Corresponds to the High Safety Goal in Region (HSGR), which means that we are in HS and the goal location is within the free walking area •Free walking area width criterion: When the goal is not in the free walking area, two new situations are defined depending on whether the area is wide or narrow: High Safety Wide Region (HSWR) and High Safety Narrow Region (HSNR) Finally, they obtain three new situations in LS. •Goal within the free walking area criterion: This criterion is similar to the second one, but this time the free walking area presents some obstacles. The resulting situation is Low Safety Goal in Region (LSGR). •Dangerous obstacle criterion: Two possible situations may occur applying this criterion. The first one, Low Safety 1 Side (LS1), is given when there are obstacles within the security zone, but only on one side of the discontinuity (closest to the goal) of the free walking area. The latter one, Low Safety 2 Sides (LS2) means there are obstacles within the security zone on the two sides of the discontinuity In Urdiales et al. (2013) authors used all the information gathered during the tests described above, adding the data of healthy people that performed the same tasks, to obtain a sample of 100 people. Their baseline user profile is built in a three-step clustering process. 1This figure is extracted from Minguez et al. (2004) shows these six possible situations which are defined according to the following criteria 26 3.3 I-DONT-FALL need to be stable patients, with at least one month from last acute event. The presence of a caregiver during the sessions is mandatory. On the other hand, persons presenting aphasia and/or neglect or major behavioural disturbances were excluded from the study. Also, people involved in rehabilitative training cannot participate in the pilot due to safety reasons. Participants who satisfied inclusion criteria were assigned randomly to one of these four groups: •MOTOR: 125 participants in walking training that perform a set of exercises with a physiotherapist using the i-Walker. •COGNITIVE: 125 participants in cognitive training that perform a set of exercises using the SOCIABLE platform. •MIXED: 125 participants in the combined training (motor + cognitive training). •PLACEBO: 125 participants as a control group (placebo activity to control the subjectexpectancy effect). The walking training process is executed in 2 sessions per week for 12 weeks (24 sessions). Each training session takes 1h of duration. Volunteers are free to drop out of the study if the subject fails to participate in the training for more than two consecutive weeks (4 sessions), then s/he is considered as a dropout. Each training session is dedicated for 1/2 to balance and 1/2 to gait exercises after a brief session on warm-up exercises. •Walking training session of 30 minutes: –3 of warm-up exercises (exercises selected from warm-up pool) –15 of balance (exercises selected from balance pool) –15 of gait (exercises selected from gait pool) •Walking training session of 60 minutes: –3 of warm-up exercises (exercises selected from warm-up pool) –15 of balance (exercises selected from balance pool) –15 of gait (exercises selected from gait pool) –5 pause –15 of balance (exercises selected from balance pool) 33 3. A REVIEW ON ASSISTIVE TECHNOLOGIES –15 of gait (exercises selected from gait pool) The central objective of this study is to reduce the number of falls in the experimental group respect to the control group. As a consequence, it is also expected to minimise the risk of falling (measured by the Tinetti test) and the fear of falling (measured by the Fall Efficacy Scale FES test). We also aim to evaluate the usability of the system and user satisfaction concerning the applied training program (i.e., if they have performed walking and cognitive training or just one of the two). The secondary outcomes of this study are the improvement of mobility using users balance and gait (measured by Tinetti test, 6 minutes Walking Test and 10 meters Walking Test) (see Tinetti et al. (1986)), as well as the improvement of the QoL and the functional and cognitive abilities. We will also analyse the dynamic relationship between pushing forces and walking dynamics, crossed with the user medical profiles. Outcomes will be evaluated through a multidimensional assessment that will be administered two times: before the training period starts (T0) and after the training period ends (T1). 34 Chapter 4 The i-Walker The assistive device presented in here is based on the prototype developed on EU funded project SHARE-it (see SHAREit and Cort´ es et al. (2010)) at the Universitat Polit` ecnica de Catalunya (UPC). The consolidated version, the one used for this work, is based on a standard 4-wheeled Rollator AD-100 with a set of embedded sensors and actuators, aiming to assist to the mobility and the rehabilitation of persons with physical and/or cognitive disabilities and monitoring their activities (see Annicchiarico et al. (2008)). In the end, the i-Walker looks like a traditional rollator. It is designed to provide potential users with sufficient ambulatory capability in an efficient, cost-effective way. It follows the ISO requirements for walking aids manipulated by both arms (see ISO) and has obtained the EU approval as a medical device for clinical research. The actual version of i-Walker has been used in the I-DONT-FALL ,FATE and ASSAM EU projects funded by the Competitiveness and Innovation Framework Programme (CIP) of the European Union (see I-DONT-FALL,FATE and ASSAM respectively). For this PhD, part of the data generated by the i-Walker in the I-DONT-FALL project has been used, which has been described in §3.3. In this chapter, a full description of the smart walker (the i-Walker, see Figure 4.1) used in this PhD is given, including an introduction to its main components and the role they play in the system; the reactive control that has been developed to provide compensation and safety in the ambulatory activities in indoor and outdoor environments; a possible approach to shared control navigation. This chapter contains also an overview to the i-Walker’s assistive environment, describing its role in different research projects: three master thesis developed at UPC, where data collected from the i-Walker in different scenarios was analysed with different perspectives 35 4. THE I-WALKER and aims. This section also provides a summary of the I-DONT-FALL project results from a clinical point of view. In Appendix §Bintroduces the results of integrating the i-Walker as an intelligent service within a Social Network (Barru´ e et al. (2015)). This work was a part of a preliminary study to assess the plausibility of learning from the interaction of several i-Walkers and their respective users and caregivers. In that work, the i-Walker interacted in the Social Network as one more agent in a Multi-Agent system. 4.1 Main components The i-Walker is a distributed micro-controller architecture which drives the system and records and provides structured information to therapists. All the electronics are embedded inside the handlers (1) and rear wheels (9) of the i-Walker. A box (3) under the seat (2) contains the computing power onboard (a Raspberry Pi) and a set of sensors that will provide information about movement and tilt. For this work, we have also added a frontal Hokuyo laser to detect obstacles and avoid possible collisions. Table 4.1 summarises all the different variables captured and gave a brief description of each one. The names appearing in parenthesis will be the ones used along the rest of this document. Variable Description Left Hand Force X (lhfx) Longitudinal (Forward-Backward) pushing force exerted by the user on the left handlebar Left Hand Force Y (lhfy) Transversal (Left-Right) pushing force exerted by the user on the left handlebar Left Hand Force Z (lhfz) Vertical (Up-Down) pushing force exerted by the user on the left handlebar Right Hand Force X (rhfx) Longitudinal (Forward-Backward) pushing force exerted by the user on the right handlebar Right Hand Force Y (rhfy) Transversal (Left-Right) pushing force exerted by the user on the right handlebar Right Hand Force Z (rhfz) Vertical (Up-Down) pushing force exerted by the user on the right handlebar 36 4.1 Main components Left Normal Force (lnf) Left Rear Wheel Normal Force. This is the force that the floor exerts on the wheel. When the value is below a given positive threshold, it means that the i-Walker is losing its contact with the floor Right Normal Force (rnf) Right Rear Wheel Normal Force. This is the force that the floor exerts on the wheel. When the value is below a given positive threshold, it means that the i-Walker is losing its contact with the floor Tilt (tilt) Angle over the lateral axis of the i-Walker Roll (roll) Angle over the longitudinal axis of the iWalker Hand Brake Left (hbl) State of the break: 1 if blocked, 2 if manually activated Hand Brake Right (hbr) State of the break: 1 if blocked, 2 if manually activated Estimated Pose X (epx) Pose estimated from the beginning of the exercise in Y axis (starting from point [0,0] in a Euclidean space) Estimated Pose Y (epy) Pose estimated from the beginning of the exercise in Y axis (starting from point [0,0] in a Euclidean space) Estimated Pose Orientation (psi) Estimated orientation of the i-Walker regarding the initial orientation (the orientation that the i-Walker had when the exercise has started) Left Wheel Speed (ls) Left wheel speed Right Wheel Speed (rs) Right wheel speed Table 4.1: i-Walker variables and definitions. Handlebars The original handlebars have been replaced by another model also used in standard rollators. These new handlebars (1) measure the user’s force exerted along the longitudinal, lateral and vertical directions (X,Yand Zrespectively) through embedded force sensors. Height graduation is maintained in this new design (8)1. By adding these force sensors in the grips of the handlebars, we can register at any time the force exerted by the user. The handlebars can also monitor the states of the brake levers, having a manual brake (7) with two different states: 1This is useful in post-stroke patients, as they may need different height graduation on each hand 37 4. THE I-WALKER Figure 4.1: The I-DONT-FALL version of the i-Walker. 38 4.1 Main components •Parking brake: a mechanical brake can be operated by pushing down the brake lever (braked state). This action is recommended when the user wants to stop in the middle of a slope or wants to rest in the seat. •Dynamic brake: the brake lever can be pulled up to slow down the speed of the walker. This will increase the pushing forces exerted by the user, giving him/her a higher sense of security. This functionality is useful in downhill, as the user does not have to be pulling the i-Walker and, thus, avoiding this one to go away too fast (a more detailed description is given in §4.2). All the information provided by the handlebars is collected and stored so the medical staff can control if the user is doing a good use of the device. The new handlebars include some features that aim to improve the interaction between the i-Walker and the person using it. On the one hand, a bright multicoloured lighting ring is used to indicate the different states of the i-Walker: calibration status and battery levels. Handlebars also embody a vibrator that can be used as a haptic device to enhance the interaction between the user and the i-Walker, e.g. when the user is getting too close to a wall or an obstacle. This lighting ring can be handy for users with visual or cognitive impairments. In both cases, the intensity and duration of the flashes of lightning/vibrations can be defined for every type of signal. Rear wheels The rear wheels (10) have been modified by adding a motor and electronics in each one. The external structure of the wheel is maintained, but the internal part is redesigned, changing the axis of rotation to couple with the motor. The rear wheels collect odometric information from three Hall-effect sensors integrated inside each engine. The gathered data is useful to calculate the (X,Y) user’s position, as well as the longitudinal and rotational speeds. In the end, these motors work by helping the user to move around safely, especially in up and down slopes. The motors can work in four different operating modes described in Table 4.2. The reactive control (see §4.2) uses these different modes depending on the amount of pushing/pulling help it is required in every moment. Central box The central box contains a Raspberry Pi, the battery and three new sensors that have been added to have a complete Inertial Measurement Unit (IMU): 39 4. THE I-WALKER Operating Mode Description FREE No action applied on motor. The wheel turns freely, like a conventional wheel. It is useful when we want the wheel to turn freely, but it also maintains access to information about kinematics, temperature, etc AS (Active Speed control) The motor is driven by a set point of angular speed. Useful when any user is not using the i-Walker but we want it to behave like an autonomous vehicle. AC (Active Current control) The motor is driven with an actual set point proportional to a torque set point. Useful to compensate the forces required from the user. AB (Active Braking control) The wheel is braked by setting a current set point that opposes to the movement. Table 4.2: i-Walker Motor operating modes. •Gyroscope: detects the three angular speeds of rotation. This information strengthens the data gathered from odometry, having more precise monitoring of the user’s movements. •Accelerometer: detects if the user is walking on a slope and/or the i-Walker is being accelerated. •Magnetometer: detects the projection of the magnetic earth filed on the integrated circuit. 4.2 Reactive control The most innovative feature of the i-Walker is the design of a reactive control, developed at the Automatic Control Department (ESAII-UPC), which provides mobility aid to the user. The amount of helping force and braking force in each hand ought both to be determined previously by a clinician. The i-Walker platform provides four primary services. Three are related to elder/impaired assistance; we use the fourth for data logging. A physiotherapist should plan all the support given to a user. Services provided at this moment are: •Active motor assistance to compensate lack of muscle force on climbs. •Active brake assistance to compensate lack muscle force on descents. 40 4.2 Reactive control Figure 4.2: Example of left-hand force compensation. •Active differential assistance to compensate unbalanced muscle force. •Recording of sensor measurements and actuators activities for later evaluation (left and right-hand forces, normal forces, tilt and odometry) Described strategies are not exclusive: we can have the user pushing the i-Walker going downhill and at the same time the i-Walker relieving him from part of the necessary pulling/pushing force to move around. For safety reasons the i-Walker automatically stops when the user releases the handlers, that is when no forces are detected on them. Hand force compensation The reactive control works with two primary variables: λand ν. The former represents the amount of helping force that the user receives, while the latter one is a helping brake force. The combination of both parameters allows the therapists to create a patient’s tailored configuration. These parameters are set during the initial setup of the i-Walker by a physiotherapist. 41 4. THE I-WALKER Figure 4.3: Hand force compensation strategies in an uphill scenario. When a user is going uphill with the i-Walker, the λparameter will release the user from part of the pushing force s/he has to do. In Figure 4.2, red areas represent the longitudinal force that the user should be exerting in the given scenario. The green areas correspond to the force compensation provided by the i-Walker. During the uphill, the rollator will compensate the lack of pushing force; when going downhill, it will assist the user by sending active break signals to the rear wheels. This compensation will avoid excessive retaining force (i.e., negative longitudinal force), and thus a possible fall. The aid provided in downhill is independent of the value of λ, as it is a safety mechanism that aims to prevent falls. The i-Walker allows to determine different amounts of compensation strategies in each side of the body according to the user’s dysfuncions. Figure 4.3 shows that the configuration of λ and νparameters is independent on each handlebar (e.g., an individual with hemiparesis needs different help on each hand). In this case, the user received helping force on the left hand (the difference between the pink area and the dotted line is the contribution of the i-Walker’s 42 4.3 The i-Walker’s Assistive Environment Figure 4.5: Dendrogram representing the clustering applied to Right Hand Force X (longitudinal force rh f x). similar methodology with a multi-variable approach to identify significant data fusion useful to learn new concepts on an individual’s walking behaviour. Moreover, it would be of particular interest to identify variables that together can define different gait disturbances and associate them with the pathologies affecting elderly people. A similar a methodology was used to classify users into two groups of age. This time, the selection for the study included 42 individuals of ages between 22 and 94, which consisted in performing a 3-minute walking test in an indoor corridor of 40 metres. The work has been developed by Ojeda (2018) as her Master thesis, which I have co-directed, and results will be presented in Ojeda et al. (2018). The methodology uses the Bag-of-SFA-Symbols (BOSS) model for the representation and 49 4. THE I-WALKER Figure 4.6: BOSS Model is used for indexing and representation and transforms the time series into BOSS histograms indexing. This method transforms the numeric time-series into a bag of words representation to later create a k-dimensional matrix. The BOSS model describes time series as an unordered set of substructures using Symbolic Fourier Approximation (SFA) words. The BOSS flow model is represented as follows (see Figure 4.6): parameters definition, windowing, SFA and histograms aggregation and reduction (Sch¨ afer (2015)). Representation and indexing The goal of this step is to transform each time series to a new representation based on a vocabulary extracted from the behaviour of the data in the frequency domain. This transformation has a set of parameters that need to be tuned in order to find the optimal configuration: length of the windows, the number of Fourier coefficients for a window and the number of letters for the word representation. For each time series, a set of fixed-size windows are generated using a sliding window of length w. The first window begins at 0 and ends in the position w, the second one will offset one position ending at position w+1 and so on until the end of the series is reached. The result will be transformed into a vocabulary where each window represents a word. The length of the words are defined by the quantization of the Fourier coefficients of the window as explained in the following paragraphs. The transformation of the windows into a vocabulary uses the SFA method, which aims to simplify information by removing the unnecessary data and keeping only the most repre50 4.3 The i-Walker’s Assistive Environment sentative characteristics. The SFA performs three steps to achieve its goal: approximation, quantization and words computation. The SFA uses a discretization method based on the Momentary Fourier Transform (MFT), for the approximation process. The MFT targets to keep the critical data extracting the Fourier coefficients of the signal (Albrecht et al. (1997)). This method allows to incrementally compute the first fFourier coefficients of a sliding window in a series in efficient manner. The idea behind this discretization is to decompose the time series into two basic type of functions (Sch¨ afer and H¨ ogqvist (2012)). The former consists in identifying slow changes in the data, while the latter identifies rapid changes. For the approximation, those with slow changes are enough for a fair description of the signal. These types of functions also provide a smoother signal with low pass filtering. The decomposition represents a time series by its Fourier coefficient. The magnitude of the coefficient represents the amplitude of the signal. The proposed approximation uses only the first fcoefficients. The first Fourier coefficient can be optionally discarded because it stands for the mean value of the signal, obtaining this way offset invariance. The quantization step reduces the granularity of the data by dividing the values of the Fourier coefficients into a histogram of equal frequency bins (Gurajada and Srivastava (1991)) and mapping each coefficient to its range. To define these histograms, a number of bins is defined as a parameter, representing the letters of the discretization alphabet that will be used for computing the words representing the windows extracted from the time series. Each position of the Fourier coefficients is discretized separately. This is done by computing Multiple Coefficient Binning (MCB), and it aims to minimize the lost information when performing the discretization process. The discretization works as a map that contains intervals of numeric values per each coefficient. The outcome of the SFA is a word of fletters per window and a set of words per time series. In order to avoid a bias due to large periods of stable signal, the BOSS model reduces the length of the vocabulary by numerosity reduction, removing identical consecutive words. The final stage of the BOSS model is to transform the vocabulary for a time series into a histogram of relative frequencies of words. This transforms the series into a vector, so we can compare different series using different similarity measures. The main advantage of this final transformation is to allow the comparison of time series of different lengths. Similarity Measure and Space Embedding The next step is to evaluate how similar are the sets of time series. From the possible similarity 51 4. THE I-WALKER functions that can be defined, the one that better fits the model is the cosine similarity measure because this metric considers only the orientation of a vector and not the magnitude (Steinbach et al. (2000)); this characteristic is useful when working with symbols instead of numbers. The output of this process is an affinity matrix that will be further transformed before clustering the data. For this model, a Spectral Embedding (Strange and Zwiggelaar (2014)) is applied to the affinity matrix to embed the data in a metric space and to enhance the relevant characteristics by using a non linear transformation. This embedding transforms the affinity matrix into a k-dimensional matrix. The number of dimensions kis part of the configuration parameters. For this paper, the transformation was limited at most to three dimensions. Clustering A Bayesian Gaussian Mixture Model (BGMM) with Dirichlet priors (Li and Mihaylova (2017); Sanjay-Gopal and Hebert (1998)) is used for partitioning the data represented by the k-dimensional data matrix. A Gaussian mixture model is a probabilistic model that assumes all the data points are generated from a mixture of a finite number of Gaussian distributions with unknown parameters. Each cluster is formed with a set of points that shape a Gaussian distribution using a Expectation Maximization (EM) algorithm. The use of a Dirichlet prior includes the determination of the number of clusters in the optimization process. The EM for mixture models consists of two steps. The first step calculates the expectation of the component for each data point given by the model parameters. The second phase maximizes the expectations calculated in the previous step concerning the model parameters. Then these two steps are repeated until the result converges. Evaluation To assess the quality of the clustering, it is necessary to apply some evaluation techniques. This approach considers two techniques: (i) the adjusted rand index (ARI, de Vargas and Bedrega (2013)) to measure the stability of the clustering to random initialization and (ii) the Silhouette index (SI, Verma et al. (2015)). Scenarios The work is based on the recognition of the patterns of the forces applied by the users to the i-Walker while walking. From the forces recorded by the i-Walker, the vertical ones seem to be the most related to the individuals0compensation strategies. Therefore, this study examines two scenarios. The first model is provided with all the forces to find if the transversal and longitudinal forces add relevant information. The second one considers only the vertical forces 52 4.4 Summary to study whether the transversal and longitudinal forces add noise instead of contributing to the results. Results This methodology has been tested with different initial setting parameters to create the vocabulary. After finding the suitable combination of parameters, the clustering was applied to the two above-mentioned scenarios. Results show that the proposed approach is able to divide participants by age using the applied forces to the i-Walker. Moreover, it can be observed how the standard deviation of the leaning forces increases with age, which might be an indicator of the loss of balance that people presents as they age. Figure 4.7) shows that the variability is directly related with the magnitude of the vertical force. The results obtained are coherent with the literature: (i) gait velocity is reduced with ageing, implying a higher risk of falling; (ii) older adults with different sorts of pathologies present an abnormal gait that is identified in a particular cluster and, moreover, is related to unbalanced use of forces, which might lead to disfunctional gait. In addition, clustering results were more accurate when combining the the three resulting forces from the handler sensors than using the vertical force as single-variate clustering. 4.4 Summary Research prototypes, like the i-Walker, are beginning to achieve the performance needed to make a difference in the daily life of the elderly society. By the moment, the market still offers only limited solutions to substantially prolonging the time that older adults can live independently at home, but these should be supported with relevant health and social care services in an integrated manner (see Barru´ e(2012), Wasson et al. (2008)). Older adults are becoming a predominant aspect of our societies and are expected to substantially affect the economy at world-wide level, changing the paradigm of public health systems ir order to maintain them as sustainable as possible. As such, solutions both efficacious and cost-effective need to be sought (Sun et al. (2014)). In particular, there is a need to improve more the cost-to-benefit ratio of robot-assisted therapy strategies and their effectiveness for rehabilitation therapy. It is clear that the i-Walker is an assistive device that offers many possibilities to clinicians, patients and relatives, and it could become a powerful tool for the design of the incoming digital health solutions. It has been designed to collect data for long-term periods, which could be useful in solutions involving the remote monitoring of elder people performing their ADLs. 53 4. THE I-WALKER Figure 4.7: Comparison of the vertical force and its variability per cluster and age. 54 4.4 Summary Authors in Barru´ e et al. (2015) have already shown the value of elders’ activity recognition and monitoring since it might be an early symptom of some decline that must be prevented as soon as possible. The i-Walker can be easily integrated with other stakeholders or assistive devices to share information, generate reports or alerts and process specific data (see Appendix §B). It has also been designed to offer constant mobility assistance to the user, providing safety and ensuring compensation on the upper limbs forces. The objective of this PhD is to combine different outputs of the i-Walker’s sensors and biological data obtained by clinicians through cognitive and physical assessments to extract some characteristics of their gait and try to identify different users profiles. This analysis will identify the strides performed by individuals while walking and try to find groups of people who walk similarly. The existence of those profiles could allow the i-Walker to provide tailored assistance while it learns new features of the patient. The main difference with the research presented in this chapter is that instead of analysing exercises as a whole, it will use the force used at each stride performed during an exercise to learn how do people walk and use the iWalker. In §6we first describe the methodology used in this PhD proposal to study the walking behaviour resulting from the interaction between the user and the i-Walker. Then a model of fall risk prediction is presented and tested with a group of participants of the I-DONT-FALL project (see §5.3.1). 55 4. THE I-WALKER 56 Chapter 5 Clinical Tests: design and implementation of a pilot protocol Health research has become an essential mainstay for to improve the actual state of the art in its different fields, such as epidemiology, biomedicine, health services, personalised medicine as well as its socio-economic impact on society. One of the most common forms of health research is the clinical trial, where volunteer individuals participate in studies to assess a new medical product, device or treatment. Most of this research is done through collecting different types of personal data, from biological characteristics to physical and/or cognitive assessments among others. During the last years, it has also included new data coming from sensors that will interact directly with the individual (e.g., through bio-metric or wearable sensors) or environmental sensors. Although health research aims to promote individuals’ health and well being, as well as to improve care and social services, its management might also represent a risk to the society due to the type of data that is stored and processed. Ethics is an essential dimension of human philosophical research, considered both as discipline and practice. Philosophers today usually divide ethical theories into three general subject areas: metaethics, normative ethics, and applied ethics. For clinical research, ethically justified criteria for the design, conduct, and review of clinical investigation can be identified by obligations to both the researcher and human subject (see Aita and Richer (2005); Guraya et al. (2014); World Medical Association (2001)). Informed consent1, confidentiality, privacy, privi1Informed consent refers to an ethical and legal doctrine based on the understanding that all interventions (diagnostic, therapeutic, preventive, or related to scientific studies) in the medical field should only be performed after a participant has been informed about the purpose, nature, consequences, and risks of the intervention and has freely consented to it, see Glickman et al. (2009). 57 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL leged communication, and respect and responsibility are critical elements of ethics in research. The European Commission has published several documents that describe, from an ethical and regulatory point of view, the different aspects that should be respected when working in a clinical trial with human subjects (see CAREGIVERSPRO-MMD (2016)). The data collected and analysed during this PhD work comes from different hospitals and care centres distributed in Spain and Italy. As a result, we built four datasets, although we excluded the first one from the analysis. The respective Ethical Committee approved each pilot test following the Clinical Trial Directive 2001/20/EC (see Council of European Union (2001)), and was supported by clinicians in the participants’ selection and recruitment and the execution of physical and cognitive assessments specified in each protocol. 5.1 Definition of a protocol According to the International Conference on Harmonization (see ICH (1996)) a protocol is a document that describes the objective(s), design, methodology, statistical considerations, and organisation of a trial. A clinical trial protocol should include: •General Information about investigators and sponsors (i.e., names and contacts). •Background Information of the tested product (i.e., its potential risks and benefits, the description of the target population or references to the literature that are relevant to the trial). •Trial Objectives and Purpose. •Trial Design should include: (i) primary endpoints and the secondary endpoints, if any, to be measured during the trial; (ii) a description of the type of trial (e.g., doubleblind, placebo-controlled or parallel design); (iii) a description of the measures taken to minimize/avoid bias (e.g., randomization or blinding); (iv) a description of the trial treatment; (v) the expected length of the subject’s participation and a description of the duration of all trial periods and follow-up. •Selection and Withdrawal of Subjects defines the subject inclusion and exclusion criteria, as well as the withdrawal criteria and procedure. •Assessment of the product includes methods, timing and scales for assessing, recording and analysing the efficacy and satisfaction parameters. 58 5.2 Protocol design Figure 5.1: Map of the Fondazione Santa Lucia environment and the two first driving tests. Figure 5.2: Map of the Fondazione Santa Lucia environment and the two last driving tests. 65 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL Figure 5.3: Map with the three scenarios of MAD pilot. created confusion among some participants. On the other hand, the straight lines were too short to extract any significant gait feature. For this reason, it was decided to improve the ambulatory exercises thanks to the experience obtained during the FSL pilot. This is why we decided to substitute this test with the 10MWT, which is a well-accepted performance measure used to assess walking speed in meters per second (m/s) over a short distance. 10MWT can also be employed to determine functional mobility, gait and vestibular function. As it is a widely used measure, it is easy to find information about gait speed according to different profiles (i.e., healthy adults, hip fracture, stroke, etc.)1. Figure 5.3 shows the updated paths that will be used in future pilots. 5.2.4.1 Ten Meter Walking Test The traditional 10 Meter Walk Test (10MWT) is a performance measure used to assess walking speed in meters per second over a short distance. It can be employed to determine functional mobility, gait and vestibular function (see Adell et al. (2013)). Conclusions are made from an expert’s observation. The intended population is ranged as: preschool children (2-5 years), children (6-12 years), adolescents (13-17 years), adults (18-64 years), elderly adults (65+) with a range of diagnoses including, among others: 1http://www.rehabmeasures.org/ 66 5.2 Protocol design •Acquired Brain Injury •Geriatrics •Hip Fracture •Lower Limb Amputation •Movement Disorders •Multiple Sclerosis •Parkinson’s Disease •Spinal Cord Injury •Stroke •Traumatic Brain Injury The 10MWT is performed by an individual who is instructed to walk without assistance in a distinct pathway where a straight line of 10 meters is marked off. Additional milestones were placed at meters 2 and 8. The clinician indicates to the user when s/he can start walking and will measure the time used to complete the distance. Usually, the clinician will only take into account the time spent during the intermediate marked 6 meters to allow some space for acceleration and deceleration (see Figure 5.4). The Start and Finish points are indicated to the user before the test execution, but no lines are marked on the floor as a guide. The timed zone corresponds to the six central meters of the exercise. The use of assistive devices is permitted but must be kept consistent and documented for each test. The 10MWT only assesses walking speed and does not consider the amount of physical assistance required devices or endurance. The 10MWT is not appropriate if the individual needs physical assistance to ambulate. Figure 5.4: The 10 Meter Walk Test measurement In the 10MWT the subject has to walk 10 meters in a maximal straightway, as above said the only measurement tool is a stopwatch (see Ali and Raad). In addition to this, the i-Walker can measure the travelled distance, the duration time, the maximal speed, the maximum of the lateral deviation and the pressure on the handlers. From the additional post-processing, it is possible to extract new indicators. Other researchers have been using robotic rollators (see for example Wang et al. (2014) and Ballesteros et al. (2015)) and/or other automatic means to measure users’ performance in the 67 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL 10MWT (see Yorozu et al. (2015)). In other cases, kinematic data from infra-red cameras are recorded to study the walking behaviour of individuals (see Nooijen et al. (2009)). The traditional 10MWT is measured by the time duration of the performance and the resulting average speed of the participant. It is considered that gait speed can be used as a useful outcome to assess the physical condition of older adults as well as to predict the risk of falling, other health deterioration or even survival (see Studenski et al. (2011)). There is a general belief that gait speed is related with ageing process (such as muscle weakness or balance loss), although results are uncertain on this theory due to biological and behavioural differences among the population (see Shimada et al. (2010)). 5.2.4.2 Timed Walking Tests Timed walking tests measure the distance an individual can walk during a given time (usually 6 minutes, named 6mWT) on a flat surface at self-paced. The 6mWTfirst applied in frail elderly patients 60-90 years of age referred to a geriatric hospital, and it targets community-dwelling frail elders. However, the test has been used in the study of a variety of chronic disease adults (Annegarn et al. (2012)) or healthy adults (Harada et al. (1999)). In this last case, authors show that active older users obtained better performances of 6mWT than non-active healthy users. Moreover, it could be used to predict morbidity and mortality. The 6mWT was used in the I-DONT-FALL project as an assessment along with other metrics to measure participants risk of falling at pre and post-treatment. The objective of this test was to evaluate the effectiveness of the different treatments (motor, cognitive, mixed, placebo) to decrease the number of falls, risk of falling and fear of falling. For the last phase of data collection, a new group of adults performed a reduced version of the 6mWT, the 3mWT. This exercise was performed in an indoor corridor or 40 meters length as depicted in Figure 5.5. The main instructions given to the participants to perform the test were: •Walk for three minutes along the corridor. •Do not drop the handlers while performing the exercise. •Turn when reaching the end of the hallway to continue with the walking trajectory. •If the time finishes and the person is in the middle of the corridor keep walking until returning to the starting point. 68 5.3 Pilots Figure 5.5: The 6 minutes Walk Test measurement. The Start and Finish points are indicated to the user before the execution of the test, but no lines are marked on the floor as a guide. The timed zone corresponds to the six central meters of the exercise. This test is expected to provide further knowledge on gait variability and walking patterns due to its higher time and effort consumption in comparison with the 10MWT (Hausdorff (2005b)). 5.3 Pilots In this section, we provide some highlight of each pilot site. The anthropometric characteristics of all the participants involved in this study are depicted in Table 5.2. Since we were collecting data from hospitals and care centres, it was hard to find a balance in age, being most of the participants from the baseline aged over 80 years. Besides women were, in general, more willing to participate. Hence their representation is higher. At the time of designing the protocol, we aimed to collect balanced handedness data, although this was almost impossible. It was quickly remarked that at the beginning of the last century, schools tend to correct left-handed children, thus nowadays most of the old age people are right-handed. 5.3.1 IDF Pilot The I-DONT-FALL dataset is composed of three of the pilots involved in the project: FSL, HGG and SERMAS from Italy and Spain. As mentioned before, participants in these pilots share the characteristic of having suffered at least one fall during the year previous to the experimental phase. Participants of the I-DONT-FALL project where assessed physically and cognitively before the beginning of the three-months training (motor, cognitive, mixed or placebo) and at the end of this period (see §3.3.3). The generated dataset includes biological and clinical 69 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL Characteristics IDF (N= 85) N(%) or Mean ±SD Baseline (MAD) (N= 60) N(%) or Mean ±SD Baseline (FSL) (N= 30) N(%) or Mean ±SD Age <80 years ≥80 years Total 32 (37.64%) 53 (62.36%) 82.53(±8.47) 8 (13.33%) 52 (86.67%) 87.58(±5.79) 29 (96.67%) 1 (3.33%) 70.97(±5.53) Gender Male Female 26 (30.58%) 59 (69.41%) 21 (35%) 39 (65%) 13 (43.33%) 17 (56.67%) MMSE 25 (±3) 25.79 (±2.97) 28.5 (±1.59) Tinetti 17 (±4) 16.56 (±3.90) 24.47 (±3.77) Barthel 79 (±19) 81.95 (±14.77) 96.33 (±6.15) Table 5.2: Anthropometric Characteristics of the Study Participants. 70 5.3 Pilots Criteria Variable Mean Std Dev Elderly Age 74.7 7.9 Formal Education Years 9.9 4.2 High Risk of Fall POMA total 19.8 5.5 High Risk of Fall Previous Falls 1.33 1.16 Non-demented MMSE 25.8 4.6 Table 5.3: Inclusion Criteria for the I-DONT-FALL final dataset. Results under 21 for the POMA analysis along with more that one fall in the recent year are criteria used to determine high risk of falling. data along with the i-Walker measurements of the 10MWT exercises performed at pre and post-treatment phases. During the cleaning process, we had to exclude several participants who passed the inclusion criteria but had inaccurate sensor readings (probably due to a bad calibration at the beginning of the exercise, a wrong exercise tagging or a failure in the communication system of the i-Walker), remaining only 87 in total from the three pilots. Table 5.3 shows the inclusion criteria parameters of the I-DONT-FALL protocol and the mean and standard values of the studied population. 5.3.2 MAD Pilot This pilot took place in March 2015 during two weeks at Los Nogales (MAD) centre, in Madrid. People participating in this pilot were older adults living in a residential care centre, with reduced mobility but good cognitive conditions, or relatives. A total of sixty individuals participated in the study, with only one drop-out. Although the objective was to collect data only from healthy older adults with no falls, finally 15 of them presented a fall during that year (see Figure 5.6 for an overview representation of MAD population). The exercises included in the final study were the 10MWT performed with no helping parameters (λ= 0). The clinical team at Los Nogales provided the same assessment results to complete the i-Walker data. As it can be readily observed in Table 5.2, participants from the MAD pilot are apparently in a better cognitive condition and present higher scores in independent living despite being older than the IDF population. Moreover, although most of the MAD participants did not suffer any fall, they present a higher risk of falling in the Tinetti Scale than individuals from IDF. In the case of FSL. Although half of the population had a previous fall, results from physical and 71 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL Figure 5.6: Distribution of MAD participants represented demographically by age and gender, but also clinically by number of falls and risk of falling cognitive scales are much better than the other groups. This is probably related to age factor since only one participant from FSL was older than 80 years old. The reduction of autonomy and mobility that older adults usually face in residential centres could explain this distribution. 5.3.3 CV I Pilot This last pilot was collected in July 2017 during two weeks at the Centre de Vida Independent (CVI) in Barcelona. People in this pilot are quite heterogeneous regarding age, which goes from 22 up to 94 years old. This dataset is composed of 42 participants, 18 of them presenting cardiac problems or injuries caused by a fall. The clinical team that supported us assessed this group both physically and cognitively, but the rest of the volunteers were considered as healthy individuals with no risk of falling (all of them were autonomous non-retired adults, with no previous falls or neurological nor physical issues). This time, the exercise performed in this study was a short version of the 6mWT. The duration of the exercise reduced by recommendation of the clinical team at CVI to three minutes (3mWT). The objective was to collect more extended bags of strides of each participant since the 10MWT exercises where too short to obtain a good profile of each person. We executed the pilot in a long indoor corridor of more than 40 metres; once participants reached the end of the hallway, they were indicated to turn around and keep walking to the starting point until the time ends. The corridor had an area wide enough to allow round turns with a rollator. 72 5.4 Summary 5.4 Summary In this section, we have described the anthropometric characteristics of the participants in the different clinical trials. We aimed to create a baseline with similar criteria than in I-DONTFALL since it is essential to isolate anthropometric effects on gait analysis (Prakash et al. (2015)). In our case, the only differentiation among trials is the number of falls. Participants in both clinical tests used an i-Walker with the vertical position of handlers adapted to their height. A brief description of the pilot (start and end point, number of repetitions and led lights function) was given before the volunteers signed the protocol agreement. During the tests, we tried to let them navigate as independent as possible, although some individuals presenting visual impairments received some indications when turning and getting into a room. We also had some issues with individuals presenting hearing impairments, especially when we needed them to stop walking. In some occasions, these exercises where repeated. In the case of people stopping later than required, data was conserved but then processed to exclude unnecessary information (see §6). Although I already had participated in other clinical trials before this work, being involved at all stages, from the design of the protocol to the test execution and later processing has been challenging but has also provided me new expertise in the design and management of this kind of experimental research field. With every clinical test, new lessons were learnt from previous mistakes, which helped to tune the protocol. Also, when dealing with a robotic device, particular attention must be given to the precision of the execution of the exercises or the calibration of the sensor system. Many internal or external factors might affect the proper collection of data and further results and analysis. A total of 632 individuals have used the i-Walker, although only 168 of them will be included in the analysis proposed in this thesis. The size of the dataset is considerably good to validate the methodology presented here, but it would be suitable to increase it in the future, especially if the objective is to characterise gait patterns by pathologies. In that case, the protocol presented should be applied to specific groups of population, e.g., people diagnosed with Alzheimer Disease, Parkinson, post-stroke, etc. 73 5. CLINICAL TESTS: DESIGN AND IMPLEMENTATION OF A PILOT PROTOCOL 74 6.1 Gait Analysis based on Human-Rollator Interaction •wand pcorrespond to the width and prominences of the peaks and are returned as vectors as well. The prominence of a peak is the minimum vertical distance that the signal must descend on either side of the peak before either climbing back to a level higher than the peak or reaching an endpoint. On the other hand, the width of each peak is computed as the distance between the points to the left and right of the peak where the signal intercepts a reference line. •<Name,Value >options specify optional comma-separated pairs of Name, value arguments that will help adapting the local maxima identification with filtering options. Name is the argument name, i.e., the filter applied to the signal, and Value is the corresponding value for this filter. We will use these options to define peak filtering, which can be done by minimum prominence height or width and minimum peak distance. Other options include filtering by a threshold, by the maximum number of peaks or by peak sorting. •xand Fs are optional parameters that specify a location vector xor a sample rate Fs of the data. When using this argument, wand locs are given in terms of xor converted to time units for the second case. Another approach to gait detection is to use distance as a reference instead of time and force prominence. In this case, the xoptional argument is used, which will be a vector containing the incremental position of the i-Walker during the exercise. Figure 6.4 show a Fxdi f f signal filtered by minimum peak distance. Working with raw data implies that the signal is not smooth and thus, the search for local maxima peaks needs to be filtered to avoid false positives. In this case, we use the distance as reference or the findpeaks function. Depending on the physical condition or height of the person using the i-Walker, the distance between steps (and thus, the number of steps) varies significantly, needing several steps of filtering to include all types of behaviours in the model. After trying different approaches, the variable that best filtered the steps positions was the average speed, and the total walked distance to infer an estimated length of the step. As before mentioned, the detected peaks are steps performed with the right foot; between two peaks of right steps, there is always a negative peak, which indicates when the left foot step has taken place. Therefore, we can consider that the action between two detected positive peaks is a stride that begins with the right leg (a right stride from now on). 81 6. METHODOLOGY 0 1 2 3 4 5 6 7 8 9 10 11 12 12.3 Time (s) -6 -4 -2 0 2 4 6 8 10 12 14 lhfx (kg) -5 0 5 10 15 rhfx (Kg) Pushing Forces lhfx rhfx 0 1 2 3 4 5 6 7 8 9 10 11 12 12.3 Time (s) -15 -10 -5 0 5 10 15 rhfx - lhfx (Kg) Combined Pushing Forces Figure 6.3: Identification of steps in a 10 Meter Walk Test exercise. a) Individual longitudinal hand forces; b) Resulting pushing vector Fxdi f f . At this stage, a vocabulary of strides has been generated, where each instance is a time series containing the force applied at each sample time along the stride (i.e., every 100ms). Also, the spatiotemporal information of each right stride is also completed with identifiers for the user, exercise and step number within the exercise, the overall stride length and time, and the average pushing force exerted during the stride. This set of characteristics allows forming a dictionary (or vocabulary) of strides, aiming to emulate the dictionary of words used in the bag-of-words methodology. This technique is well-known in text and document classification, where the frequency of each word is used as a classifier feature. The bag-of-words has also been used in computer vision applied to image classification, where image features are treated as words. This approach is also known as bag-of-X due to its adaptability to other fields. In general, this technique relies on identifying relevant key-words and analysing their frequency of appearance. In this thesis, the stride vocabulary will be used to characterise types of exercises 82 6.1 Gait Analysis based on Human-Rollator Interaction Figure 6.4: Fxdi f f signal filtered by minimum peak distance. The incremental position vector has been used as location reference of the peaks. according to the occurrences of each type of stride. In this case, a bag-of-strides will be used to study the results of the clustering process. The concept of bag-of-steps was introduced in Pla et al. (2017), where authors used it to predict the rehabilitation length and discharge date of a patient using insole force sensors. Other than the final application, these two methodologies differ on the vocabulary generated (strides vs steps) and the tools used to collect them. The rest of the approach presented is developed with the aim to determine whether these bags-of-strides contain information that allows emerging new knowledge on walking patterns on strides or groups of individuals with similar walking patterns. 6.1.3 Clustering Time Series A time series is a sequence of observations collected at a given period in chronological order. A time series dataset Dis thus a collection of time series TS with the same time base, although, the number of observations can vary among the instances of the dataset and is represented as D ={TS1,TS2, ..., T Sn}. Traditional time series analysis focus on smoothing, decomposition and forecasting, but they are later used also to solve clustering and classification problems. Time series clustering is becoming popular in recent studies since it allows exploring large amounts 83 6. METHODOLOGY of data in more complex data mining algorithms such as rule discovery, indexing, classification and anomaly detection. Time series clustering has been applied to different fields, such as finances or medical research (the most commonly used example would be in ECG analysis). The increasing use of sensors in many sorts of studies has also enhanced the interest in temporal analysis. Time series can be analysed on frequency or time base, depending on the expected outcomes and the type of analysis applied. Usually, time series clustering has two steps: the first step consists of working out an appropriate distance/similarity measure and then, at the second step, to apply an existing unsupervised partitioning technique, such as k-means, hierarchical clustering, density-based clustering or subspace clustering. As a result, the time series dataset Dwill be divided into C={C1, C2, ..., Ck}clusters, in such a way that homogeneous time series are grouped together based on a specific similarity measure (Aghabozorgi et al. (2015)). The primary challenge on time series clustering relies on the high dimensionality of datasets and the selection of the similarity measure. Also, time series are naturally noisy and might contain outliers, hence it is required to apply some signal filtering to obtain a useful similarity matrix. Time series clustering can be applied (i) as a whole, where all the individual time series are clustered concerning their similarity; (ii) as a subsequence of time series obtained with a sliding window; (iii) as a time point clustering, where time series are segmented. In the case of the study presented here, a time series clustering is applied to both baseline and validation datasets. The size is not excessively large due to the short time of exercise, but the i-Walker can work for several hours and generate large amounts of data. Moreover, since data is broken into strides, each instance of the dataset will be presumably short and of different size. This might present a computational challenge when comparing the similarity among them, but solutions can be found in the literature to tackle this. In traditional clustering, the distance between objects of a dataset is calculated based on exact match, but this is not possible in time series clustering due to the nature of time series objects, where sample intervals and lengths can be irregular. As mentioned, there are different distance measures applied to time series, such as Hausdorff distance, Dynamic Time Warping, Euclidean distance or Longest Common Sub-Sequence among others. The decision on which one is more appropriate relies on the clustering objectives, i.e., the type of similarity among objects that is being studied. Three primary goals have been identified: similarity in time, in shape or change. Finding similarities in time means to look for the similarity at each step time and is usually based on Euclidean distances. However, 84 6.1 Gait Analysis based on Human-Rollator Interaction due to the dimensionality of time series datasets, it is recommended to apply first a transformation to reduce the computational cost. Similarities in change look for similar structural differences between time series, i.e., time series with identical correlation structures. This approach is not recommended for short time series like the ones used in this thesis. In the case of this thesis, it aims to find similar strides by the shape of the pushing forces along the stride. A popular method to compare and identify patterns of time series is the Dynamic Time Warping (DTW from now on, Deza and Deza (2013)). DTW finds optimal alignment between two-time series, regardless of the time points that each one contains. In fact, DTW has already been used in the context of gait analysis. For example, in Barth et al. (2015) authors propose a methodology for automatic single stride segmentation from walking exercises. Sequences were extracted from inertial movement sensors located at the participants’ footwear. In Boulgouris et al. (2004) authors use video sequences for gait recognition and use DTW to compare test and reference gait cycles. Derawi et al. (2010) presents another methodology using video sequences and wearable sensors to recognise gait cycles and DTW is again the similarity measure to compare them. In this thesis, a DTW distance is calculated for each pair of strides from the previously generated vocabulary of strides, generating a M×Mdistance matrix, where Mis the number of objects in the time series dataset. This was performed in RStudio using the dtw and dist libraries; each DTW distance matrix took several hours to be generated. The resulting distance matrix is then used as input of the clustering, instead of the original time series, which will save a massive amount of time when computing the time series clustering analysis. Since the learning algorithm used is unsupervised, the clustering process will require several repetitions until finding the right partition; once the distance matrix is generated, the computational time of trying different values of k(partitions) is reduced to few seconds. In this case, the partition around medoids algorithm (Kaufman and Rousseeuw (1990)), also known as k-medoids, was applied in Rstudio using cluster and proxy libraries. The term medoid refers to an observation within a cluster for which the sum of the distances between it and all the other members of the cluster is a minimum. The main challenge of working with an unsupervised learning technique is to determine the number of clusters (k) that better represents the analysed data. Unfortunately, there are no definitive criteria to determine that value, but it is somewhat subjective and depends on the method used for measuring similarities and the parameters applied to the clustering algorithm. The partition around medoids algorithm can be evaluated with some well-known techniques 85 6. METHODOLOGY such as the within-cluster sums of squares or the average Silhouette among others (Arbelaitz et al. (2013)). The latter is an output of the pam algorithm available in RStudio when performing k-medoids. It returns an average value for each object of the dataset, representing how well does it lay within its cluster. Values can be either negative or positive, but the higher the values are, the better is for the clustering quality. Thus, the optimal number of clusters kis the one that maximises the average silhouette over a range of possible values for k. However, in this case, the Silhouette method might not be entirely suitable for evaluating DTW distances since it is based on Euclidean similarity. Therefore, clustering will be executed with different values of kand assessed in the second part of the gait analysis. At this stage, the time series clustering returns a clustering vector that assigns each stride of the time series dataset to a group of strides, regardless of the user, they belong to. This approach applies to any exercise performed with the i-Walker within a straight line and will be applied to both the 10MWT exercises from IDF and MAD pilots and to the straight lines extracted from the 3mWT collected at the CVI pilot. 6.1.4 Exercises as bags-of-strides Once the optimal k(types of strides) is defined, the outputs of the partitioning can be grouped by clusters to study the distribution of strides from the same exercise in the different obtained kgroups. As a result, a new dataset is created, with as many objects as exercises had the original datasets (IDFMAD and CVI), where each object represents a pair {user, exercise} and the representation of the strides extracted from that exercise in terms of clusters i.e., the proportion of strides of an exercise that fall in each strides cluster. In other terms, we obtain a histogram for each pair {user, exercise}representing the frequency of appearance of each stride cluster by exercise. A set of histograms H={h1,h2, ..., hn}is generated, where each hk is composed by a number of bins which sum up to one. Clustering histograms have become popular thanks to the bag-of-words categorisation method (Nielsen et al. (2014)). This provides a first visual overview of the distribution of each exercise, but there are still too many objects to analyse. Therefore, a second k-medoids partitioning is applied to this new dataset, using two different similarity metrics (Deza and Deza (2013)): •Euclidean distance is the similarity measure used by default in partitional clusterings. •Kullback-Leibler (KL) Divergence: also called relative entropy, measures how one probability distribution diverges from a second expected probability distribution. 86 6.2 Spatio-temporal Analysis The symmetrised KL Divergence is used to obtain a distance matrix of the dataset. Previous research has shown that it has better performance than the Euclidean distance. This second partition will be executed with both distances, although only KL Divergence will be considered for the final analysis. Again this clustering process will be performed with different values of kto determine the most suitable partition. The result of this clustering will be a set of bag-ofstrides that represent exercises characterised by a given vocabulary of strides. In Chapter §7.2 a graphical representation of these bag-of-strides is given as well as the analysis of the content of the clustering result. 6.1.5 Cluster stability In this second clustering, it was noticed that each execution of the pam algorithm returned a different clustering vector. Therefore, each <pam,k0>, where k0={2,3,4,5,6}, was executed 20 times in order to determine the optimal partitions. Then the normalised mutual information (NMI, Wagner and Wagner (2007)) index was applied to measure the mutual dependence between each pair of clustering vectors obtained. As a result, for each <pam,k0>, a 20 ×20 matrix was generated with the NMI results. The partition is considered stable when all the results from the 20 ×20 NMI matrix returned 1 (i.e. all the clustering vectors are equal). In case any combination returns 1 for all the repetitions, then the most stable should be chosen, considering the number of 1 appearing in the 20 ×20 matrix and the value of the other cases (the higher, the better). This process was applied using both the Euclidean and the KL divergence as a distance measure for the k-medoids. The pairs <pam,k0>where NMI was equal to 1 were the partitions selected for the final analysis. Table 6.1 shows the combinations that will be obtained and presented in Chapter §7. This table only includes the clusters of exercises with the KL divergence, as it is more used in the literature when clustering histograms. In addition, the combinations obtained with the Euclidean distance were part of the KL results as well. 6.2 Spatio-temporal Analysis In this section, we will describe in detail the different parameters that were proposed in Prakash et al. (2015) and relate it to the corresponding i-Walker outcomes. Although parameters are here presented separately, we will later combine them to obtain a better understanding of our studied population and the differences they present in relation to the i-Walker data. 87 6. METHODOLOGY Pilot Stride Clustering Exercise Clustering Name IDFMAD k = 4 k’ = 5 idfmad4k5k’ CVI k = 3 k’ = 3 cvi3k3k’ k = 5 k’ = {3,5}cvi5k3k’ cvi5k5k’ k = 6 k’ = {3,4}cvi6k3k’ cvi6k4k’ Table 6.1: Resulting combinations of strides per exercises to be analysed in different scenarios Spatio-temporal parameters provide the simplest form of objective gait evaluation in terms of time and distance. This section describes the metrics that are commonly used in the literature to assess individuals’ walking performances. Relation of these metrics to the extracted variables from the i-Walker raw data is also provided, although the previous section already showed how these were obtained. To review the definitions of step, stride and gait cycle, see Chapter §2. 6.2.1 Descriptive Gait Parameters Most of the human gait research focus on the same set of metrics to assess the gait quality of the user. In normal walking patterns, a gait cycle is the continuous repetition of strides or, as defined in Pappas et al. (2001), a succession of 4 phases for a given foot: stance, heel-off, swing and heel-strike. Considering a step as the movement of one foot in front of the other, and a stride (or gait cycle) as a step from one foot followed by a step from the other foot, we can define the following quantitative metrics: •Step length: distance (in meters or centimetres) between the heel-off and heel-strike of one foot. •Step time: duration (in seconds) between the heel-off and heel-strike of one foot. •Number of steps: total number of steps during an exercise. •Number of stops: total number of stops during an exercise •Stride length: distance (in meters or centimetres) of a gait cycle. 88 6.2 Spatio-temporal Analysis •Stride time: duration (in seconds) of a gait cycle (or the time between steps of the same foot) •Cadence: Number of steps per minute The metrics included in this thesis are the stride length and time, the number of steps and cadence. In this work, the number of strides was collected instead, and the rhythm was modified to represent the number of strides per minute. Although the presented methodology allows also extracting the information at step level, these were not considered in the analysis. Stride-tostride fluctuations have been traditionally used in literature to study gait variability, which is a complementary way to evaluate human locomotion and its change with age or disease. It has also been closely related to gait disorders, leading the results to categorise individuals participating in these studies in terms of frailty or risk of falling (Hausdorff (2005a)). As mentioned before, steps (and thus, strides) are obtained by local maxima of the Fxdi f f pushing force. We use the walking distance as the reference. Therefore, it is easy to extract the distance between peaks of local maxima. In addition, we know that data is collected at a periodical and chronological time, so the time between peaks can be obtained by counting the number of observations and converting it to seconds. We provide the number of strides at the end of the dictionary of strides generation, and the cadence is a simple conversion from the number of strides along the exercise (usually few seconds) to minutes. 6.2.2 Gait Velocity There are several approaches that use gait velocity (e.g.,Montero-Odasso et al. (2012)) as an assessment metric for cognitive decline and risk of falling, although one of the most commonly used is the 10 Meter Walking Test. Traditionally, this test is measured by direct observation of the performance and a simple time measurement through a chronometer. More recent studies include the use of wearable sensors (usually accelerometers placed at wrist or ankles) or walking platforms for foot tracking (see §2). Less frequent are the studies where assistive devices are involved. However, authors like Wang et al. (2014) and Ballesteros et al. (2015) have proposed different methodologies using different versions of robotised rollators. This sort of studies ought to be very careful with the target population selected since the rollator is not adequate to specific pathologies, e.g., people with hemiparesis who have no lateral balance generally due to the consequences of a stroke. As mentioned in Chapter §4.3.1, the i-Walker can provide a helping force that would give the required balance to walk safely. However, this 89 6. METHODOLOGY was not considered in this thesis due to the difficulty to pass the protocol in each pilot and to get access to people with this pathology. According to Montero-Odasso et al. (2005), older adults can be divided in three groups based on gait velocity (GV): high GV (>1 m/s), medium GV (0.7-1 m/s) and low GV (<0.7 m/s). It is considered that people not belonging to the high GV group, and especially those belonging to the low GV group, have more probabilities to suffer from adverse events due to physical or cognitive decline. We divided participants according to this methodology, and it will be used to observe the trends in different periods of time for the I-DONT-FALL population. We will perform an analysis on Gait Velocity for both the IDF as an independent dataset, the baseline datasets (IDFMAD) and the validation dataset (CVI) to study the distribution of our population. Participants will be classified according to the approach proposed by MonteroOdasso et al. (2005), and we will determine whether our population follows the clinical hypothesis mentioned during this work: (i) poor performances of gait velocity are related to age, and cognitive status; (ii) physical and/or cognitive training helps to improve gait velocity and, thus, reducing gait variability; (iii) training will also help to reduce the risk of falling and fear of falling. For the IDF data, we will compare the performance before and after the three-months training (T0 and T1 respectively), and we will study the evolution for each of the training groups (motor, cognitive, mixed and placebo). 6.3 User Driving Skills There are several methods of evaluating the human-robot interaction while driving. Urdiales (2012) provides methods to (i) analyse user driving skills and obtain user profiles according to their performance, and (ii) interpret user disagreement in a collaborative control driving strategy. In our case the i-Walker works with a purely reactive control for now, as no control strategies have been added to assist in the effectiveness of the driving skills of the user. Hence, it will act as a sensing platform that collects data from different sensors during the exercise at a rate of 10Hz. In this thesis, we will follow the same philosophy than in Urdiales (2012) to assess human driving skills, which has been fully described in §3.2.1, but adapting it to the characteristics of the i-Walker architecture. The main difference is that the i-Walker requires human body motion to move, i.e., users must use angle joints such as legs to create motion and arms to exert a force on the handlers of the rollator. By analysing how does an elder person use the 90 6.5 Modelling Fall Risk Assessment 6.5 Modelling Fall Risk Assessment One of the objectives of the I-DONT-FALL project was to reduce the risk of falling in elderly population through different types of training. As mentioned before, data obtained from clinical scales at baseline was compared to the post-training results to evaluate the effectiveness of the system. We used the i-Walker data generated during the 10MWT exercises before and after the training and applied a machine learning technique to the prediction of the risk of falling for an individual. We selected several variables for the 10MWT exercises from the study sample using the i-Walker. These variables combined raw data with other calculated variables. They correspond to the average values from the force sensors (X,Yand Zdirections) of both hands and the average speed. The dataset has a total of 20 variables. The average was computed for all the 10 meters of the exercise and also for the values discarding the first and last two meters (the central part of the exercise). The reason for dropping values from the beginning and the end of the exercises was to reduce noise from acceleration and deceleration phases and focusing on the stable regime of the exercise. The exercises were divided into two datasets. The first corresponds to the exercises performed before treatment (T0) and the second to the exercises performed after treatment (T1). The data before treatment was used as a training set to obtain a predictive model for the risk of falling. The second dataset is used as test to detect if the population has changed their status as an effect of the treatment. To obtain the model for fall risk, a logistic regression was performed using L1regularisation, first with all the variables, and then selecting only the relevant ones, by discarding all variables that were assigned zero weight by the regularised logistic regression. From all the variables only 10 were used by the model that included variables for all the exercise (average speed, mean force on the Xdirection, right hand force on the Xdirection and left hand forces on the three directions) and only for the central part (left hand forces on the Xand Zdirections and right hand forces on the Yand Zdirections). A Support Vector Machine (SVM) model with linear kernel was also computed, obtaining identical results. It is clear that the information collected from the force exerted by the user to the i-Walker is essential to characterise its walking behaviour. This model was not applied to the rest of datasets since they did not follow the same protocol. The rest of pilots executed the test in one day, and hence no temporal comparison of their 97 6. METHODOLOGY evolution is possible. 6.6 Summary The i-Walker offers several sources of information that can be related to the human walking behaviour. The main idea is to translate data sensor readings into readable and understandable reports, aiming to complement the observations of clinicians while the user is performing an exercise. Fusing the expert knowledge on human body motion with the data obtained from older adults using i-Walker would allow the design of messages to alert or inform the user about a situation (see a first approach to the implementation of this service in Appendix B). Moreover, the i-Walker would be able to take decisions in dangerous situations and assist the user in his/her mobility. One of this PhD work aims is to find those known gait parameters represented in data obtained from the i-Walker for the selected target population. For this, we have first focused on a descriptive analysis of the studied parameters. We have also proposed a model to predict the risk of falling for the I-DONT-FALL population. Finally, previous results on applying clustering techniques to identify individuals with falls are promising (see Chapter §4.3.3), although they consider exercises as a whole. Thus, it cannot process information on user’s interaction at each moment. We expect that by clustering users’ strides, we will be able to identify gait disturbances associated with non-healthy older adults. As a result of the methodology presented here, a bag-of-strides and a bag-of-exercises are created for each dataset (IDFMAD and CVI), containing the spatiotemporal metrics abovementioned and the exerted forces by the human to the i-Walker at stride level while performing the exercise. We are confident as we have robust sets of data coming from I-DONT-FALL and other campaigns. 98 Chapter 7 Results In this chapter, we present the results obtained from the proposed methodology and the already introduced datasets. To describe these results we will discuss the relevance of each of the gait characteristics extracted from the interaction between the participants and the i-Walker. For this analysis, we have selected some evaluation metrics that can be found in the literature of human gait research (see Chapter §6) that will be evaluated by means of the data obtained from the i-Walker sensors. First, the methodology will be applied to the baseline population (both fallers and healthy older adults) regarding gait velocity while performing the 10 Meter Walk Test. The relevance of this measure in the early detection of the decline in older adults has already been explained in this document. The objective is to observe how are the participants distributed regarding gait velocity and whether significant differences are observable between groups. Second, the results of the spatiotemporal analysis are shown, focusing on the characteristics of an elder individual’s gait (length, time, velocity). Following, a study users’ driving skills concerning directionality and laterality is given. Then, the results from the clustering approach to gait analysis are presented for both the baseline and validation datasets (i.e., observations for 10MWT and 3mWT). Finally, the results of a regression model of fall risk prediction applied to the IDF dataset are introduced in this chapter. 7.1 SpatioTemporal Analysis In this section, we will show preliminary results on the time and distance parameters of human gait through its interaction with the i-Walker. 99 7. RESULTS 7.1.1 Descriptive Gait Parameters We have seen that steps can be represented by frequency or by distance, but both require the use of filters to discard noise from human oscillations. Once steps are detected and located in the data signal, it is possible to represent it in different understandable ways as shown in Figure 7.1. Here the signal has been filtered by distance, i.e., by a minimum of walked distance between strides based on the average walking speed. According to our interpretation, exercises start when the first right stride is detected. With a similar process, we can extract the strides performed with the left leg. Furthermore, this figure adds a new layer of information, introducing the concepts of prominence and width defined in §6.1.2. In this case, the width represents the time in seconds between peaks, i.e., the average time the user has spent pushing between steps. This concept can also be interpreted in terms of distance travelled between peaks. However, for the rest of the analysis, the width was represented in terms of distance, as already explained in §6.1. As above said, the positive parts of the signal corresponding to those moments where the user was exerting more force with its right hand. It is expected that in an exercise like the 10MWT, where the trajectory is assumed to be a straight line, the signal will be balanced. A person drifting to one side while driving will present less zero-time moments since the increase in the opposite side will not be high enough to compensate the movement. The comparison of the force prominence for both arms aims to illustrate how does the individual interact with the i-Walker and associate it with its physical or cognitive dysfunction. Figure 7.2 represents the evolution of the step length for each foot while performing the 10MWT. Usually, 10MWT assessments disregard the beginning and the end of the exercise (2 meters on each side) since they correspond to the acceleration and deceleration phases and the gait pace is not regular. In Figure 7.2.a, we can observe this behaviour for both feet. Moreover, it strengthens the hypothesis that the individual has started the exercise with the right foot and finishes it with the left one. Figure 7.2.b represents the increments of each foot’s strides in terms of distance (meters). As we observe, in this case, the individual does bigger strides with the right foot. The same analysis and representation can be given in terms of stride time. After applying the bag-of-strides approach described in the methodology, a database is generated where, for each stride, the following information is being stored: •user ID, gender, age range, fall risk (Tinetti), faller (yes/no) 100 7.1 SpatioTemporal Analysis 0 1 2 3 4 5 6 7 8 9 10 11 12 Time (s) -15 -10 -5 0 5 10 15 FxDiff (Kg) signal peak prominence width (half-prominence) Figure 7.1: Right leg strides, pushing force increments and average pushing time within peaks •timestamp and number of exercise (being T0-T1 for the IDF dataset, or round number in the case of MAD and CV I) •stride ID (incremental within the same exercise) •total number of strides in exercise •distance travelled, duration of the exercise, average speed and cadence (in metres, seconds, metres/seconds and strides/minute respectively) •stride length and time (in metres and seconds respectively) •average stride length and time in exercise(in metres and seconds respectively) •average pushing force along the stride in terms of Fxdi f f This will allow depicting the distribution of our studied population from different perspectives that will be presented in the following sections. Table 7.1 shows the main spatio temporal characteristics for the IDFMAD pilot population. 7.1.2 Gait Velocity The Gait Velocity has been calculated for both the IDF and MAD pilots for the 10 Meter Walk Test. It was also applied to the CVI pilot, where each straight line was considered as an 101 7. RESULTS Left Right Feet 0 1 2 3 4 5 6 7 8 9 10 Distance (m) 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 Steps -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Stride Length (m) Right Left Figure 7.2: Evolution of an individual’s stride length during the 10MWT: (a) Estimated feet position; (b) Stride length for each foot. independent exercise. In the case of MAD and CVI, gait velocity was obtained by calculating the average walking distance and walking time of all the exercise performed by the same participant; this technique is known in the literature as the test-retest. For the final analysis, some participants were excluded due to sensor failure or drop-out. At a first stage, participants from MAD with falls were excluded, but we finally decided to treat IDF and MAD as a single baseline dataset. The distribution of each pilot is depicted in separated tables and commented below. The Spatio Temporal Characteristics Biological Characteristics N Cadence Speed sLength sLength CV sTime sTime CV FxDiff Age M 85 49,88 0,80 0,93 42,45 1,04 54,31 1,72 Gender F 49 49,14 0,80 0,94 42,63 1,05 55,37 1,36 Gender M 36 50,89 0,79 0,91 42,21 1,02 52,87 2,19 Age O 200 45,83 0,61 0,76 42,06 1,28 55,98 1,66 Gender F 140 45,56 0,57 0,71 42,12 1,35 55,91 1,60 Gender M 60 46,48 0,71 0,87 41,91 1,12 56,16 1,81 Total general 285 47,04 0,67 0,81 42,18 1,21 55,49 1,68 Table 7.1: Spatio-temporal characteristics of the IDFMAD population. 102 7.1 SpatioTemporal Analysis Characteristics Low GV (N= 52) N(%) or Mean ±SD Median GV (N= 27) N(%) or Mean ±SD High GV (N= 6) N(%) or Mean ±SD Age <80 ≥80 Total 14 (26.9%) 38 (73.1%) 85.3 (±6.8) 15 (55.5%) 12 (44.5%) 79.2 (±9.4) 6 (100%) 0 (0%) 75.3 (±5.2) Gender Female Male 36 (69.2%) 16 18 (66.7%) 9 4 (66.7%) 2 Treatment Mix Motor Cognitive Placebo 15 8 12 17 4 6 7 10 4 1 1 0 Risk Low Medium High 1 11 40 3 17 7 4 2 0 Table 7.2: Distribution of the IDF participants before 3-months training (IDF-T0) performing the 10MWT. Data is represented by Gait Velocity groups. These results are used to evaluate the effectiveness of the IDF solution after the training phase. distribution has been observed in terms of age, gender, treatment and original risk of falling. Table 7.2 and Table 7.3 provide the same structure for the IDF population. Tables 7.4 and Table 7.5 show the same information for the MAD and CVI pilots respectively, except for the treatment information which only regards the IDF protocol. In these two latter pilots, participants repeated several times the same exercise. Thus the walking speed has been obtained from the average walking speed of each exercise performed by the same user, as before-mentioned. Even though the MAD population is supposed to be healthier than the one from IDF, they are also older (the proportion of people aged 80+ is significantly higher than in the IDF pilot, which corresponds to the typical age of people living in care centres). Hence, only three individuals (5%) reached the High GV group and, surprisingly, they present a medium or high risk of falling. Unfortunately, the Low risk of falling is poorly represented, and thus no definite conclusions can be extracted from this result. The rest of groups are coherent with the litera103 7. RESULTS Characteristics Low GV (N= 50) N(%) or Mean ±SD Median GV (N= 22) N(%) or Mean ±SD High GV (N=12) N(%) or Mean ±SD Age <80 ≥80 Total 15 (30%) 35 (70%) 84.8 (±6.6) 10 (45.5%) 12 (55.5%) 82.7 (±9.3) 11 (91.7%) 1 (8.3%) 73.3 (±6.9) Gender Female Male 37 (74%) 13 13 (59.1%) 9 7 (58.3%) 5 Treatment Mix Motor Cognitive Placebo 13 11 10 16 4 3 8 7 6 1 2 3 Risk Low Medium High 0 24 26 4 17 1 7 4 1 Table 7.3: Distribution of the IDF participants after 3-months training (IDF-T1) performing the 10MWT. Data is represented by Gait Velocity groups and is used to assess the usability of the fall risk prevention system. ture: most of the people with higher risk of falling are classified in the low GV group. After contacting Los Nogales clinical team, it was possible to retrieve updated assessments one year after the pilot took place of those participants that were still living in the care centre. It was observed a general improvement in terms of Tinetti scale results, especially for those having a better cognitive status (measured with the MMSE scale). Thus, it is presumed that people with a particular cognitive condition and ability (or willingness) to keep an active living were able to improve their mobility and thus reduce their risk and/or fear of falling. This data will not be shared for ethical issues, as it was not included in the consented inform signed by the participants at the time of the pilot testing. In the case of the CV I pilot, the age range is expanded, including now the Young category (≤65 years old). In this test, 13 participants had some cardiac pathology or physical injuries resulting from a fall and were living in the care centre. The rest of people are healthy people, most of them are still employed or retired in the last year, and were living independently. The distribution of the GV groups by age follows the same trend than in the rest of pilots. It also 104 7.1 SpatioTemporal Analysis Characteristics Low GV (N= 34) N(%) or Mean ±SD Median GV (N= 21) N(%) or Mean ±SD High GV (N= 3) N(%) or Mean ±SD Age <80 ≥80 Mean 3 (8.8%) 31 (91.2%) 88.3 (±5.8) 4 (19%) 17 (77.78%) 86.9 (±6.2) 0 (0%) 3 (100%) 87.6 (±5.8) Gender Female Male 27 (79.4%) 7 9 (42.8%) 12 1 (33.3%) 2 Risk Low Medium High 1 8 21 1 10 10 0 2 1 Table 7.4: Distribution of the MAD participants after a 10MWT test-retest represented by Gait Velocity groups. shows that 8 of the 13 challenged participants fall in the Low GV group; these people where even the eldest from this challenged group, which was the expected. The other five users are classified in the Median GV group correspond to middle-aged people with premature cardiac pathologies. It has also been noticed that young participants change their walking behaviour when using an assistive device, reducing their walking speed, which might affect the results of the clustering presented in §7.2. It was also expected that participants (especially challenged or elder) would show some trend in the walking speed among rounds (i.e., elder people performing each straight line slower than the previous one), but this is merely observable. People tend to walk at a similar speed between each round, especially the youngest and eldest participants; people in middle age are the ones that show more variations, but never more than 0.2 m/s, and usually always within the same GV class. Results on Table 7.2 and Table 7.3 show that the group of participants are distributed as expected in terms of age and gait velocity: the younger they are, the faster they walk. Moreover, people under 80 years old tend to improve their performance in terms of velocity after the 3-months training. On the other hand, men walk faster than women on average, taking into account the proportion of gender representation, although this could easily be related to 105 7. RESULTS Characteristics Low GV (N= 8) N(%) or Mean ±SD Median GV (N= 18) N(%) or Mean ±SD High GV (N= 16) N(%) or Mean ±SD Age <65 65-80 ≥80 Mean 0 (0%) 0 (0%) 8 (100%) 87 (±4.2) 8 (44.4%) 3 (16.7%) 7 (38.9%) 63.2 (±24.4) 9 (56.3) 1 (6.2) 6 (37.5%) 52.5 (±19.1) Gender Female Male 7 (87.5%) 1 14 (77.8%) 4 13 (81.2%) 3 Risk Low Medium High 0 1 7 11 6 1 14 2 0 Table 7.5: Distribution of the CV I participants performing the 3mWT. Data is obtained by extracting the straight lines of the exercise, which go from 25 to 35 metres long, and applying the test-retest method to obtain the average speed. Data is represented by Gait Velocity (GV) groups. anthropometric parameters, such as height. Finally, as it was expected in the outcomes of the I-DONT-FALL project (see I-DONT-FALL), people having been assigned to the Mixed Training group have obtained better performances than the rest of the groups. However, the Motor Training group did not present the expected improvements in terms of velocity, although the sample is small to extract a reliable conclusion with this single metric. In general terms, results are coherent with those existing in the reviewed literature: (i) gait velocity is related to some biological characteristics of the individual; (ii) it is essential to promote active, healthy ageing both in physical and cognitive tasks; (iii) gait velocity can be used as an early indicator of decline in older adults. Results were also represented graphically to observe the distribution among the different pilots. Figure 7.3 shows the difference in gait velocity between T0 and T1. The red area represents the MAD pilot, which is concentrated around the zero value (i.e., no differences between T0 and T1). The three studied pilots from IDF (HGG in green, FSL in blue and SERMAS in yellow) show different distributions in gait velocity. Areas on the right side of the 106 7.2 Clustering Results of the Gait Analysis Bag-of-strides in Exercise Cluster 2 1 2 3 4 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 2 1 2 3 4 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 3 1 2 3 4 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in exercise Cluster 4 1 2 3 4 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 5 1 2 3 4 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Figure 7.5: Scenario 1: idfmad4k5k’. Cluster representation of bags-of-strides with four types of strides, grouped in five sorts of exercises. we aimed to focus on a common pathology (in this case, cardiologic dysfunction) for the elder group of participants. However, it was not possible to reach the minimum number of volunteers. Thus we had to increase the number of young participants, obtaining an unbalanced pilot in terms of biological data. Results are presented in the following section. 7.2.2 CV I pilot The CVI dataset provided different combinations of stable clustering results, which are presented in Table 6.1. Each combination will be studied in this section and justifications of the selected option will be given at the end. In principle, the more clusters we obtain, the more accurate will their information be. However, we must observe the type of groups that have been formed in order to determine their intra similarity in terms of gait shape and anthropometric characteristics of the people belonging to a same bag-of-strides. 113 7. RESULTS Figure 7.6: Right deviation during a 10MWT 7.2.2.1 First Scenario The first scenario studies the results of categorising strides in three clusters. The bags-ofstirdes of each exercise are then distributed three types of exercises. Figure 7.8 depicts the distribution of the bag-of-strides for each kind of exercise and Table 7.7 contains some details about the anthropometric and spatio-temporal characteristics of the participants belonging to each cluster, along with the general distribution. Users’ information is classified by age range (Middle, Old, Young) and then by gender (Female, Male). The risk of falling has not been included since it is highly correlated to the age group in this dataset. Figure 7.8 depicts that each group of exercises has a gait shape which prevails over the rest, and in each case, it is a different stride type. In the first type of exercise, users performed mainly strides of type 1, with over a 70% of representation. In the second group of exercises, again over 70% of strides belong to the second cluster of strides. Finally, more than 80% of the strides performed in the third kind of exercise belong to the third cluster of strides. As it can be observed in Figure 7.7, the shape of these strides share some common characteristics at first glance. The two first clusters have similar time length, while strides in the third group are 114 7.2 Clustering Results of the Gait Analysis Cluster 1 2 3 Age N Speed Cadence Stride Length N Speed Cadence Stride Length N Speed Cadence Stride Length Age M 6 1,17 61,58 1,18 6 1,09 59,81 1,14 26 1,05 55,20 1,18 F 6 1,17 61,58 1,18 6 1,09 59,81 1,14 16 0,99 53,24 1,14 M 10 1,14 58,34 1,25 Age O 7 0,95 54,16 1,11 24 0,75 53,03 0,82 16 0,56 52,82 0,65 F 2 0,70 56,22 0,77 19 0,73 51,12 0,80 16 0,56 52,82 0,65 M 5 1,05 53,33 1,24 5 0,86 60,31 0,86 Age Y 37 1,05 56,62 1,13 7 1,20 59,48 1,28 24 1,05 55,87 1,18 F 37 1,05 56,62 1,13 7 1,20 59,48 1,28 16 1,07 56,91 1,18 M 8 1,00 53,80 1,18 Total 50 1,05 56,87 1,14 37 0,89 55,35 0,96 66 0,93 54,87 1,05 Table 7.7: Anthropometric and spatiotemporal gait characteristics of Scenario 1 slightly longer. The other difference relies on the amount of pushing force applied. The second group differs from the other having the biggest difference in the resulting pushing force: users in this cluster uses a larger pushing force on the part of the body that is making the step. Most of the exercises from this cluster are performed by participants aged 85+ years old with several comorbidities; it is presumable that these individuals require more support on the i-Walker to perform the exercise and thus present this gait pattern that clearly shows the foot transition in the swing phase. The other two groups remain in positive values even when the left step is taking place, which means that the compensation between both sides of the body while walking is more unbalanced. This situation is more remarkable in the third cluster which is the most heterogeneous group in age and gender representation. Finally, Exercise cluster 1 has mainly classified female participants, especially young women. 7.2.2.2 Second Scenario The second scenario considers five types of strides. The correspondent bags-of-strides are then grouped in either three or five sorts of exercises, as represented in Figure 7.10 and Figure 7.11 respectively. Gait shapes of this clustering are depicted in Figure 7.9. In the first case, it can be observed that each group of exercises has one or two dominant types of strides, and in each case, these differ. In the first group, the second stride type prevails with over 65% of the strides performed in the exercises clustered in this group. This group of strides is also the one with more observations, including one-third of the total amount of them. People in this group are 115 7. RESULTS 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi3k-75-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi3k-3862-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi3k-1230-medoid and neighbours in cluster Figure 7.7: Gait shapes in CVI dataset with three clusters of strides Bag-of-strides in Exercise Cluster 1 123 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 2 123 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 3 123 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Figure 7.8: Scenario 1: cvi3k3k’. Cluster representation of bags-of-strides with three types of strides, grouped in three sorts of exercises. 116 7.2 Clustering Results of the Gait Analysis those presenting a minimal force variation. The gait shape indicates that a significant righthand force was exerted during the phases regarding the right step, and then the forces applied from the two parts of the body get balanced to perform the left step. This cluster is again the most heterogeneous in age and gender and coincides with the one obtained in the previous scenario. In the second group of exercises, almost 80% of the strides are represented by two types, strides 1 and 3, which are also the shorter ones. As depicted in Table 7.8, this group is mainly formed by exercises performed by women, especially young ones, and one single man from the older age group. The stride length is strongly related to the individual’s height, which is generally lower in women. In this case it appears to be the main discriminator, along with the amount of force employed, to categorise most women together. Finally, the third cluster of exercises is mostly represented by the two last types of strides from Figure 7.9. The shape of the fourth stride reveals that the force compensation between both parts of the body during the left step are more balanced, since they are close to zero. The shape of the fifth stride is the one with more variation between right and left hand force, but both start and end the gait cycle (i.e., the right step phases), using a considerable amount of force with the right hand. As previously observed, this is a pattern in exercises performed by the older volunteers, which were diagnosed with several comorbidities (mostly related to cardiological problems). Most of the people executing this type of force along the gait cycle, and especially those falling in this last Exercise Cluster, are those presenting more comorbidities, and thus have presumably more disturbances while walking. Table 7.8 shows that this cluster is composed by half of the exercises performed by the older adults, but has also some representation from the young and middle-age groups. During the execution of this pilot, it was observed that some people from the control group (i.e., not needing a mobility support device) modify they walk when using the i-Walker1, which could explain the walking behaviour of those young people appearing in this Exercise cluster. If we take a look at the second case of this scenario (Figure 7.11), where bags-of-strides are categorised into five sorts of exercises, it can be observed that in general, a pattern is emerging in two clusters. In this case, the first Exercise cluster corresponds to the previously explored clusters that were mainly formed by young women (in this case 28 exercises performed by 1This remark is given under a personal, empirical evidence. It was observed that some young and middle age participants could cause some false lead during the analysis: some changed their walking speed during the exercise, or others were very focused on the interaction they were having with the handlers. 117 7. RESULTS 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi5k-29-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi5k-1230-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi5k-318-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi5k-1297-medoid and neighbours in cluster 5 10 15 20 25 30 35 40 45 50 Stride Time (#instances in timeseries) -6 -4 -2 0 2 4 6 FxDiff Force (in N) cvi5k-3797-medoid and neighbours in cluster Figure 7.9: Gait shapes in CVI dataset with five clusters of strides Bag-of-strides in Exercise Cluster 1 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 2 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 3 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Figure 7.10: Scenario 2.1: cvi5k3k’. Cluster representation of bags-of-strides with five types of strides, grouped in three sorts of exercises. 118 7.2 Clustering Results of the Gait Analysis Cluster 1 2 3 Age N Speed Cadence Stride Length N Speed Cadence Stride Length N Speed Cadence Stride Length Age M 26 1,05 59,52 1,18 5 1,18 67,46 1,18 7 1,09 63,84 1,14 F 16 0,99 57,32 1,14 5 1,18 67,46 1,18 7 1,09 63,84 1,14 M 10 1,14 63,04 1,25 Age O 13 0,50 54,13 0,58 11 0,89 57,80 1,02 23 0,76 55,71 0,82 F 13 0,50 54,13 0,58 7 0,74 58,25 0,80 17 0,74 53,15 0,82 M 4 1,17 57,01 1,41 6 0,81 62,96 0,82 Age Y 20 1,01 59,26 1,15 40 1,05 61,37 1,14 8 1,27 66,42 1,32 F 12 1,01 59,84 1,14 40 1,05 61,37 1,14 8 1,27 66,42 1,32 M 8 1,00 58,39 1,18 Total 59 0,91 58,25 1,04 56 1,03 61,21 1,12 38 0,93 59,46 0,99 Table 7.8: Anthropometric and spatiotemporal gait characteristics of Scenario 2.1 young female and two from a middle-age). The second Exercise cluster characterises the mixed group, where now all the exercises executed by male volunteers appear together. Surprisingly, the exercises coming from women have the worst average performances of all age groups and clusters (i.e., women in this cluster performed at the lowest average gait velocity in comparison to women in same age category but different clusters). The second Exercise cluster is again the one containing more instances, being most of them of middle-aged people (26 out of 58). Since most of the strides are of type two, it is likely to say that it is the more general gait shape of all the strides collected in this dataset. This cluster of strides falls in the middle of stride length, being some of them very short (around 2.5 seconds). The force compensation in the swing phase of the gait cycle is quite balanced, with a general trend of applying more right forces, especially when reaching the end of the period. The other three clusters could be understood as a splitting of the group formed mainly by old ladies in scenario 2.1. As it can be observed in Figure 7.11, the step type has lost relevance in the three first Exercise clusters to become more significant in the last two classes of exercises. It is especially important in the fifth Exercise cluster since people in this group combine the two steps with major force variation during the swing phase, which should be analysed by an expert eye to determine if this is a possible identifier of gait disturbance or fall risk. 7.2.2.3 Third Scenario The third scenario presented with the CVI dataset is the one with six types of strides. Once the bags-of-strides are generated, the clustering of the histograms returns two possible solutions 119 7. RESULTS Cluster 1 2 3 Age N Speed Cadence Stride Length N Speed Cadence Stride Length N Speed Cadence Stride Length Age M 2 1,16 64,61 1,16 26 1,05 59,52 1,18 2 1,03 65,22 1,04 F 2 1,16 64,61 1,16 16 0,99 57,32 1,14 2 1,03 65,22 1,04 M 10 1,14 63,04 1,25 Age O 13 0,50 54,13 0,58 11 0,88 61,33 0,93 F 13 0,50 54,13 0,58 8 0,85 60,00 0,91 M 3 0,97 64,88 0,97 Age Y 28 1,05 61,38 1,14 19 1,02 59,25 1,17 5 1,18 65,22 1,23 F 28 1,05 61,38 1,14 11 1,02 59,87 1,16 5 1,18 65,22 1,23 M 8 1,00 58,39 1,18 Total 30 1,06 61,59 1,14 58 0,91 58,22 1,04 18 0,98 62,84 1,02 Cluster 4 5 Age N Speed Cadence Stride Length N Speed Cadence Stride Length Age M 6 1,16 65,50 1,20 2 1,12 65,78 1,14 F 6 1,16 65,50 1,20 2 1,12 65,78 1,14 M Age O 12 0,84 57,46 0,97 11 0,68 50,26 0,75 F 5 0,70 55,68 0,80 11 0,68 50,26 0,75 M 7 0,95 58,74 1,09 Age Y 7 1,23 64,84 1,31 9 1,02 60,80 1,10 F 7 1,23 64,84 1,31 9 1,02 60,80 1,10 M Total 25 1,03 61,46 1,12 22 0,86 55,98 0,93 Table 7.9: Anthropometric and spatio-temporal gait characteristics of Scenario 2.2 with three and four sorts of exercises, which are represented in Figures 7.14 and 7.15 respectively. As a general observation from these two figures, and in comparison with the other two scenarios, it is clear that the more types of strides we have, the more difficult it will be to obtain a type of stride which strongly represents an Exercise cluster. In this third scenario,the more prevalent types of strides reach up to half of the strides appearing in that cluster, while in the first scenario this could go up to 80%. Also, the more types of strides we have, the harder it is to find significant differences among them. Observing Figures 7.12 and 7.13, we can see that the first strides cluster represent the shorter strides in time and cluster 5 has the longer ones; the rest of clusters have a mix of short to long strides, going from 3 to 5 seconds per stride. In the first case (Figure 7.14), from the three Exercise clusters, only the second has a prevalent type of stride. This cluster is again mainly composed by a mix of young and middleaged people but also has a significant representation of the more elderly population: 25% of 120 7.2 Clustering Results of the Gait Analysis Bag-of-strides in Exercise Cluster 1 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 2 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 3 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 4 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Bag-of-strides in Exercise Cluster 5 12345 Clusters of strides 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Strides distribution Figure 7.11: Scenario 2.2: cvi5k5k’. Cluster representation of bags-of-strides with five types of strides, grouped in five sorts of exercises. the exercises performed by people aged 80+ fall in this cluster, despite being those with the lowest gait velocity (0.5 m/s) within the older people. As it can be observed, people in this group apply more right-hand force since values of Fxdi f f are always positive, although they remain balanced during the swing phase (there are no sharp changes in the gait shape during the gait cycle). The third Exercise cluster is mainly represented by strides type 1 and 4 (which are barely present in the other two exercise clusters). These strides are the shortest in time. The stride shape also indicates that the force applied to the i-Walker varies at each phase of the cycle: when the right step is taking place, the pushing force is mainly given by the right force and changes to the left force at the left step. The amount of force applied in each case is quite similar; thus in that sense, they are compensated. In this case, 38 out of 46 exercises falling in this cluster were performed by young people, and in general, the gait velocity was high, even for the three exercises performed by old adults. It is coherent that strides that are short in time belong to those people who walk faster or to those with shorter legs, as stated before. The first Exercise cluster is the most homogeneous in stride distribution, except for the 121 7. RESULTS Cluster 1 2 3 Age N Speed Cadence Stride Length N Speed Cadence Stride Length N Speed Cadence Stride Length Age M 7 1,09 66,04 1,14 26 1,05 61,68 1,18 5 1,18 69,97 1,18 F 7 1,09 66,04 1,14 16 0,99 59,36 1,14 5 1,18 69,97 1,18 M 10 1,14 65,39 1,25 Age O 32 0,78 57,16 0,87 12 0,52 56,50 0,59 3 0,82 58,56 0,96 F 23 0,72 54,71 0,81 12 0,52 56,50 0,59 2 0,68 60,92 0,73 M 9 0,94 63,42 1,01 1 1,11 53,83 1,41 Age Y 10 1,27 69,51 1,33 20 1,01 61,46 1,15 38 1,04 63,34 1,13 F 10 1,27 69,51 1,33 12 1,01 61,97 1,14 38 1,04 63,34 1,13 M 8 1,00 60,69 1,18 Total 49 0,93 60,95 1,00 58 0,92 60,53 1,05 46 1,04 63,75 1,12 Table 7.10: Anthropometric and spatiotemporal gait characteristics of Scenario 3.1 stride types three and four which have almost no representation: these two types are the two more relevant in the other two clusters of exercises. Strides can be grouped in two categories, according to the amount of force required during the right step phase: in the first two strides, people employ around 2N of pushing force in Fxdi f f , which is a reasonable quantity. The two last types of strides exert around 6N of force in the same phases, which means that users performing these steps rely more on the support offered by the rollator. In addition, these strides can also be differently grouped by shape: the second and fourth stride represent gait cycles where body forces were compensated during the left step part of the cycle, with an increasing amount of right force applied during the right step phases. The first and last gait shapes present a higher oscillation of forces during the swing phase, i.e., participants performing these strides require higher support for both parts of the body while walking. In Table 7.10, we can see that this group is mainly composed of older adults, with some representation of the younger and middle-age groups. The strides more relevant from the k-medoids result belong to challenged older adults with vascular dementia or hypertension diagnosed, among others. The second case, with four types of exercises, is quite similar to the previous one: clusters 1 and 4 have a prevalent stride each one, which is the same than Scenario 3.1, but also those found in the second scenario. Cluster 1 is mainly composed of the younger female participants (although ten have moved to other clusters) but still contains few exercises performed by women from different age ranges. Cluster 4 is once more the heterogeneous one, with a similar representation of all groups of ages but slightly higher in middle-age individuals. This cluster also has grouped together all the exercises performed by men aged 80years old. As 122 7.3 CVI Cluster Explanation from the Spatio-Temporal Gait Characteristics y = Ec1 y = Ec2 y = Ec3 Weigth Feature Weigth Feature Weigth Feature 1.49 rhfx 0.83 rhfx 2.18 rhfx 0.41 rhfz 0.66 rhfz 0.51 lhfy 0.31 sLengthCV 0.59 lhfy 0.48 lhfx Table 7.14: Most relevant features of SVM in Scenario 1 of representation. In the case of the SVM, Table 7.14 shows the most relevant features by Exercise cluster(which correspond to each ycolumn). The weight is an absolute value and it mainly represents the importance of a feature over the others: the larger are the weight values, the more significance has their associated variable. In this first scenario, Exercise cluster 1 contains mainly instances of young female that, as previously described based on Table 7.7 and Figure 7.7, have the shortest stride shape in length. The pushing force applied in these exercises is quite balanced. This is coherent with the relevant features obtained in Table 7.14. Moreover, this is the only class containing a spatio-temporal gait characteristic (the coefficient of variance of the stride length) among its most relevant ones. This support the observations made in the previous section related to the common pattern of the stride shape of this group. The second Exercise cluster represents those strides with higher force variation during the gait cycle. In this case, the SVM considered almost as important the pushing and leaning forces, but it also gives relevance to the lateral force. This class contains most exercises performed by older adults with cardiologic pathologies. This again backs up the sorts of strides that represent this Exercise cluster: people requiring of walking assistance show a higher variance in the pushing force, a pattern that is probably present in the rest of applied forces. The third class considers rh f x as the most relevant feature, far from the scores obtained by the rest of variables. 7.3.2 Second Scenario The second scenario has five sorts of strides and presents two possible options for the number of Exercise clusters, separated in scenarios 2.1 and 2.2. In this case, the RF obtains its worst accuracy (0.73) while SVM has its best performance (tied with the first scenario result, see Table 7.12). Comparing Tables 7.14 and 7.15, we observe that the accuracy results are quite similar. On the one hand, one of the Exercise clusters have 129 7. RESULTS precision recall f1-score support Exercise Clus 1 0.98 0.87 0.92 67 Exercise Clus 2 0.89 0.77 0.83 65 Exercise Clus 3 0.37 0.67 0.47 21 avg / total 0.86 0.80 0.82 153 Table 7.15: Accuracy results of the SVM in Scenario 2.1 y = Ec1 y = Ec2 y = Ec3 Weigth Feature Weigth Feature Weigth Feature 1.85 rhfx 1.71 rhfx 0.74 rhfz 0.52 speed 0.36 speed 0.70 lhfz 0.50 lhfx 0.26 rhfz 0.55 lhfy Table 7.16: Most relevant features of SVM in Scenario 2.1 very low precision scores (0.43 in the first scenario, 0.37 in this case). On the other hand, another class has obtained a 0.98 of precision in both scenarios. The exercises falling in each case are very similar. The cluster with better performances corresponds to the previously named as the mixed group, with higher gender and age balance (it is also the cluster with more observations). On the other hand, the third cluster is mainly composed by older participants and a few young and middle-aged women. Almost half of the original exercises of this cluster are classified in the second cluster (which includes most of the exercises performed by young women). It is possible that people performing the fourth stride type depicted in Figure 7.9, which has a similar shape than the second stride type, corresponds to the younger individuals found in the third Exercise cluster. The most relevant features for the RF are again rh f x,lh f z and rh f z with weight values of 0.38, 0.16 and 0.14 respectively. In the case of the SVM, Table 7.16 depicts how the pushing force along with the speed are equally represented in classes 1 and 2, while the third class gives more importance to the leaning forces. If we observe the fifth stride type in Figure 7.9, we can conclude that individuals in this cluster rely more on the i-Walker as a support device while walking. The amount of exerted pushing force can only be compensated by a significant leaning force in order to compensate the body balance during the gait cycle. Moreover, the left lateral force is also relevant in this class. It seems that the combination of the applied forces is different than in the other age ranges, which could be an indicator of a pathological gait. 130 7.3 CVI Cluster Explanation from the Spatio-Temporal Gait Characteristics Scenario 2.2, with five types of strides and of exercises, differs from the previous one in the accuracy deviation for the SVM algorithm. This case is the first time that the RF Cross Validation accuracy excels the SVM. In here, the results in accuracy are better in average compared to the previous scenarios, as depicted in Table 7.17, being the worst score 0.52 for the fourth Exercise cluster. The bags-of-strides belonging to this class are the most heterogeneouly distributed as depicted in Figure 7.11, with almost 40% of the strides belonging to the first type and the rest equally distributed (except for the second stride which has almost no relevance in this group of exercises). It is pressumable that the walking pattern of these people show characteristics of gait variability. In addition, based on the spatio-temporal data depicted in Table 7.9, people in this group have the longest stride lengths, independently of their age range. Table 7.18 shows the three more relevant coefficients per class. In the case of the fourth cluster, previously commented, the SVM gives more relevance to spatio-temporal characteristics such as the walking speed or the stride length. The instances from this cluster that have fallen in the fifth class belong to and elder male adult with high risk of falling, with a stride length and walking speed significantly below the performances of other male adults in this class. In addition, these characteristics are more similar to those found in the fifth Exercise cluster. However, due to the lack of male representation in this dataset, we cannot be sure of the accuracy of this affirmation. The third class, which contains exercises that have a similar age distribution and stride length mean values, also relies in spatio temporal features for the classification (walking speed and the coefficient of variance of the stride time). However, these two classes have obtained the worst precision and recall scores; they are also those paying less attention to the hand forces respect to the others. In general, we can observe that the more stride and exercise types we have, the more relevant become the spatio-temporal features, although the pushing or leaning forces are still present at each class. In fact, RF keeps considering them the most relevant features, giving 0.38 of importance to rhfx, 0.14 to rhfz and 0.11 to lhfz. We can say that the interaction of the human with the i-Walker is very important to classify correctly the exercises, although results are better when the walking speed is also considered. 7.3.3 Third scenario The third scenario considers two possible numbers of exercises for bags-of-strides composed by six types of strides. 131 7. RESULTS precision recall f1-score support Exercise Clus 1 0.80 0.73 0.76 33 Exercise Clus 2 0.97 0.92 0.94 61 Exercise Clus 3 0.67 0.75 0.71 16 Exercise Clus 4 0.52 0.65 0.58 20 Exercise Clus 5 0.82 0.78 0.80 23 avg / total 0.82 0.80 0.81 153 Table 7.17: Accuracy results of the SVM in Scenario 2.2 y = Ec1 y = Ec2 y = Ec3 y = Ec4 y = Ec5 Weigth Feature Weigth Feature Weigth Feature Weigth Feature Weigth Feature 2.79 rhfx 5.53 rhfx 4.471 sTimeCV 2.21 sLength 6.96 rhfx 2.63 sLengthCV 4.41 speed 2.93 speed 1.90 speed 5.19 lhfx 1.58 lhfx 3.83 rhfy 2.36 lhfz 1.28 rhfx 2.98 cadence Table 7.18: Most relevant features of SVM in Scenario 2.2 The first case represents three clusters of exercises. The accuracy obtained with Cross Validation is 0.79 for RF and 0.75 for SVM. The most important features in RF are again rh f x,rh f z and lh f z with 0.41, 0.19 and 0.14 of the total reprentation. In this case, the left leaning force is the one discriminating the observations of the second Exercise cluster, while rh f z is more relevant for the other two clusters. If we go a step further in the decision tree, we find that the cadence and the stride time are strongly correlated to the classification of the Exercise clusters 1 and 3; more specifically, small values of these variables are distinctive of the exercises classified in cluster 3. This is coherent with the distribution of bags-of-strides in Exercise cluster one (see Figure 7.14) and with the prevalent strides of this histogram (strides 1 and 4, depicted in Figures 7.12 and 7.13 respectively). This is also represented in Table 7.20, where the stride time appears as one of the most valuable coefficients of the third class. The first Exercise cluster has obtained the worst precision score, with more than half of the exercises distributed in the other two clusters, as shown in Table 7.19. The pattern is similar to the ones described in the previous scenarios. The most relevant variables of this cluster according to the SVM explanation in Table 7.20 are rh f z and lh f z, which is coherent with the stride shapes represented in this Exercise cluster (see Figures 7.12 and 7.14). It is pressumable that the SVM has classified people with a more latent force interaction with the i-Walker together, taking the younger participants to other clusters. That means that the individuals showing this 132 7.3 CVI Cluster Explanation from the Spatio-Temporal Gait Characteristics precision recall f1-score support Exercise Clus 1 0.47 0.74 0.58 31 Exercise Clus 2 0.98 0.85 0.91 67 Exercise Clus 3 0.85 0.71 0.77 55 avg / total 0.83 0.78 0.79 153 Table 7.19: Accuracy results of the SVM in Scenario 3.1 y = Ec1 y = Ec2 y = Ec3 Weigth Feature Weigth Feature Weigth Feature 0.94 rhfz 1.86 rhfx 1.42 rhfx 0.68 lhfz 0.50 lhfx 0.55 rhfz 0.59 lhfz 0.48 lhfy 0.38 sTime Table 7.20: Most relevant features of SVM in Scenario 3.1 interaction with the i-Walker are those with higher dependence on the supporting device and, thus, those with higher risk (or fear) of falling. Finally, the second case of this third scenario consists of six types of strides and four clusters of exercises. The RF has a mean accuracy score of 0.8 while the SVM returns the worst mean accuracy among all (0.70). This time, the three most important features of RF are rh f x (0.33), rh f z (0.23) and lh f z (0.09, which is closely followed by the stride time coefficient of variance), i.e., almost 60% of the data can be explained with these variables. Surprinsingly, the right leaning force does not appear among the most relevant coefficients of SVM, as shown in Table 7.22. These variables are specially relevant for Exercise cluster 4, which is the one with best precision performance and is also the one with more classified objects, which corresponds to already mentioned the mixed group (see Table 7.11). According to Table 7.22, its most important variables are the walking speed, the right pushing force rh f x and also the stride length. This cluster contains most of the exercises performed by male participants, which pressumably are stronger and taller than women. Thus it combines spatio-temporal characteristics with the force interaction between the individual and the i-Walker. On the other hand, the first and thrid Exercise clusters are those with worst results on precision and recall, although they have significantly improved compared to the results obtained in the first case of this scenario. Its most relevant features are rh f x and lh f z and rh f y; however, it has been previously mentioned that this cluster is mainly formed by young women, which main gait characteristic is the short stride length. It is likely to think that the first case of this 133 7. RESULTS precision recall f1-score support Exercise Clus 1 0.69 0.76 0.72 29 Exercise Clus 2 0.85 0.74 0.80 47 Exercise Clus 3 0.61 0.78 0.68 18 Exercise Clus 4 0.95 0.92 0.93 59 avg / total 0.83 0.82 0.82 153 Table 7.21: Accuracy results of the SVM in Scenario 3.2 y = Ec1 y = Ec2 y = Ec3 y = Ec4 Weigth Feature Weigth Feature Weigth Feature Weigth Feature 2.56 rhfx 6.47 sLength 14.75 vel 9.18 vel 1.94 lhfz 6.35 rhfx 12.56 sLength 6.11 rhfx 1.44 rhfy 6.17 vel 4.53 sLengthCV 5.92 sLength Table 7.22: Most relevant features of SVM in Scenario 3.2 scenario is more accurate in this sense, since it also considers gait characteristics to obtain a more general description of a gait cycle. 7.4 Modeling Fall Risk As already explained in Chapter §3.3, the main objective of the I-DONT-FALL project was to design a physical and cognitive treatment focused on older adults at high risk of falling or having suffered a fall along the last year. This treatment aimed to reduce, not only the number of falls, but also the risk and fear of falling. The IDF dataset used, introduced in §3.3.3 and Table 5.1, was also used to develop a model able to predict the fall risk of those individuals after the treatment period. The methodology has been introduced in Chapter §6.5 and the following results were published in Cort´ es et al. (2016). The best accuracy obtained from the model measured using 10-fold cross validation was 0.88 with standard deviation 0.1. The accuracy of the model for all the data was 0.95. Table 7.23 presents the results from the model for each class. The resulting model was applied to the T1 dataset, dividing the data between the individuals that were submitted to treatment and the individuals that performed no treatment (placebo). The goal is to test if their fall risk status had changed or not. 134 7.4 Modeling Fall Risk precision recall f1-score support Low Fall Risk 0.91 0.77 0.83 13 High Fall Risk 0.96 0.99 0.97 72 avg / total 0.95 0.95 0.95 85 Table 7.23: Accuracy results of the logistic regression model for the training set. The placebo group is formed by 27 individuals, 24 of them had high fall risk, and 3 had low fall risk. The model obtained with the T0 data fits well to this data, predicting the correct label for all the subjects except for two high fall risk individuals that are classified as low fall risk. The treatment group is formed by 68 individuals, 48 of them had high fall risk, and 10 had low fall risk. The model predicts the same label for 43 individuals (36 high fall risk and seven low fall risk). The remaining 20 individuals have changed their status. This means that this part of the dataset has deviated from the initial model and the treatment has had some effect on the fall risk status of the individuals. Specifically, 12 high-risk fall individuals are now considered low risk. Still, it seems that many individuals have not had enough benefit from the treatment in terms of fall risk, and just a minority have improved their condition. Also, some individuals have worsened their status, probably because of the natural course of their medical conditions. Most of the observations in this dataset belong to people at high risk of falling. Moreover, it is hard to observe evident improvements in people at this age, since they continue to develop comorbidities that in the end affect their locomotion and other abilities. It would be interesting to extend the dataset with a broader age range and risk fall. It could also be instrumental to enlarge the treatment period and the post-treatment evaluation to study the effects of such rehabilitation at long-term perspective. Possible future work could combine the predicted risk of falling with the categorisation of the walking performance to complement the strategies of control to provide assistive and safe navigation. 135 7. RESULTS 136 Chapter 8 Conclusions Older adults are becoming the predominant demographic group in most developed societies. The process of this ageing society is already affecting, or will soon affect, both developed and developing countries. Seniors usually suffer from one or more diseases and disabilities related to age, causing a loss of residual skills and, hence, autonomy. This often poses a barrier for the elderly, limiting their access to relatives, friends and social activities, which may lead to isolation, depression and severely impacts their QoL. Ageing also challenges the ability to perform the activities of daily living, such as dressing, bathing or self-feeding, but also to the functional mobility. As such, solutions both efficacious and cost-effective need to be sought with two main objectives: (i) improve, or at least maintain, the QoL of seniors and their relatives, and (ii) rethink public health systems, taking leverage of technological solutions that allow a remote and ubiquitous health management (see §3.2,§Bor Barru´ e et al. (2017)). The Internet of Medical Things is a new concept emerging recently that takes into account this current situation. It aims to enable the machine to machine interaction and real-time intervention solutions to create a network of connected devices continuously collecting and/or processing data. Concepts already explained in this document, such as smart wearable or assistive devices, home-use medical devices or mobile healthcare applications, will be the principal components that will allow communicating with medical experts remotely. This system could not only be used for monitoring, but also for prevention, healthy lifestyle promotion or remote intervention in emergency situations. This is also empowered in Europe by the H2020 funding programmes, which aim to include the new generation of mHealth solutions that will help users to manage their health and communicate with related stakeholders of the care process (e.g., doctors, informal caregivers, relatives, etc.). The i-Walker is a definite candidate to play 137 8. CONCLUSIONS a role in this health system architecture as it meets the requirements needed to contribute to its achievement. Through this work, the strong relationship between cognition and mobility has been shown, as well as the importance of being active while ageing to maintain, not only our skills but also our autonomy in community-dwelling. In the case of mobility recovery and assistance, we have reviewed several solutions developed during the last two decades that involve different assistive technologies and sensors and interact at various levels with end-users and health professionals. Within this field, smart walkers offer an exciting opportunity to the senior population as a support tool for gait and balance. However, they also present some limitations for specific users with physical or cognitive impairments depending on the number of legs and wheels, such as post-stroke users (see §4.3.1 and Giuliani et al. (2012), Morone et al. (2016)). The i-Walker has been presented in this document as a smart walker with a system of sensors and actuators that is already able to assist people in different types of environments with a reactive control, providing safety and self-confidence to the user. We have also reviewed previous research works in which the i-Walker has an essential role as a rehabilitation tool (Giuliani et al. (2012)) and as an intelligent service integrated into a medical social network (Barru´ e et al. (2015)). One of the main challenges when working with a robotic tool with sensors is the interpretation of raw data. In the previous works developed with the i-Walker (see Chapter §4.3 and Annex §B). In this PhD, we have been working in turning these data to readable and understandable information for clinicians in different formats such as activity reports, graphics showing the evolution of a sensor reading through an exercise or messages via social network communication channels. The results obtained in the works as mentioned above were studied from a clinical perspective, where scales and performances are compared between periods. In §4.3.3, we have also presented two approaches for data analysis using unsupervised learning techniques to categorise exercises by type (straights, turns) or by participants’ age. However, it is also interesting to analyse the human-robot interaction from a biomechanical point of view, i.e., by interpreting the data extracted from the onboard sensors and relate it to the study of gait characteristics of old-age individuals with high risk of falling. The i-Walker offers a unique opportunity to learn the effect of forces exerted directly by a person at every moment while walking and to validate our hypothesis that with this information it is possible to determine patterns of user interaction with the i-Walker as well as assessing the risk of falling in a near-future. Through this document, we have mentioned different techniques of gait analysis 138