Using Wearable Device Data for Step Measurement on Parkinson's Disease Population

Using Wearable Device Data for Step Measurement on Parkinson’s Disease Population Howon Ryu1 1University of California, San Diego, Herbert Wertheim School of Public Health and Human Longevity Science, San Diego, CA, USA Abstract Parkinson's disease (PD) is a progressive neurodegenerative disorder with various motor symptoms. Home detection and monitoring of such symptoms prove to be valuable, as it enables more constant monitoring at patient's convenience. Wearable devices equipped with inertial measurement unit (IMU) sensors are particularly essential in objective symptoms progression monitoring at home. Some literatures identify gait features, which characterize a person's walking or running movement, as important predictors for detecting PD symptoms. Such gait features can be derived from IMU signals. In this work, we propose a step measurement methodology using convolutional neural network architecture, which is an integral step in deriving important gait features. With the limited accessibility to such gait features, an open-source step-measurement model that translates raw IMU signals into gait features would be valuable to researchers in Parkinson's disease. We demonstrate the use of the proposed model through the WearGait-PD dataset. Key Words: digital health, wearable device, Parkinson’s disease, inertial measurement unit, time series, convolutional neural network 1. Introduction Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by a range of motor symptoms including tremors or change in gait patterns such as freezing of gait. For PD early detection, continuous and objective monitoring of such symptoms is critical. Home-based detection and monitoring approaches are particularly important, as they enable more frequent and consistent tracking without the need for patients to regularly visit clinical settings in person. In this regard, wearable devices equipped with inertial measurement unit (IMU) sensors have emerged as valuable tools, serving as objective means of assessing PD progression. IMU typically consists of accelerometers and gyroscopes, with an occasional inclusion of magnetometers; IMU tracks movement dynamics and spatial orientation in three-dimensional space. Specifically, accelerometers measure acceleration along the X, Y, and Z axes, while gyroscopes capture angular velocity across the same dimensions. Together, the IMU signals provide a description of how an object or a person moves over time. Figure 1: Example plot of IMU signal, showing accelerometer x, y, and z readings. IMU signals present several challenges in analysis. IMU signals are inherently time series and high-dimensional. De-noising is often necessary because of the noise in the input stemming from device placement, calibration, or sensor drift [1-2]. Additionally, IMU signals consist of multiple modalities of data, such as accelerometer and gyroscope, which adds to the complex correlational structure of the input data where the [x, y, z] readings of one modality are highly correlated while the readings across different modalities are correlated to a lesser degree. The complexity of the IMU signals as input data demands advanced modeling approaches. One effective way to summarize IMU data is through gait features, which are metrics that characterize walking or running patterns. Gait features include measures such as cadence (steps per minute), stride length (distance between the two footsteps of the same side), and step length (distance between the two footsteps of different sides). These metrics are particularly relevant in PD, where changes in walking patterns often reflect disease onset and progression. By condensing complex IMU signals into interpretable measures, gait features serve as a bridge between raw sensor data and clinically meaningful prediction models. Prior literature consistently demonstrates how using gait features contribute to effective and interpretable PD symptom classification and prediction models (Table 1). However, accessibility to gait features remains a significant challenge. Most gait metrics are derived using proprietary algorithms embedded within IMU devices which limits accessibility. Furthermore, device-specific algorithms also limit generalizability across datasets of different devices, as features may not be comparable. Given these challenges, there is a need for open-source models that derive gait features directly from raw IMU signals. Such models would enable reproducible and more universal gait derivation. One promising approach involves predicting foot contact information from the raw IMU signals, as foot contact provides the foundation for computing a wide range of gait features. By detecting foot contact information, we create a stepping stone for deriving clinically relevant gait features without relying on proprietary algorithms. In this respect, we propose a deep learning model designed to predict foot contact from raw IMU recordings. Specifically, we implemented a 1D convolutional neural network (CNN) with skip connections, trained on accelerometer and gyroscope signals. The CNN-based architecture effectively captures temporal dependencies within the time series data. The predicted foot contact events serve as a gateway to reconstructing critical gait features. Table 1: Summarized list of literature showcasing the use of gait features in Parkinson’s disease prediction. Paper Year Task Feature Identification of motor progression in Parkinson’s disease using wearable sensors and machine learning [3] 2023 Classification of PD severity (MDS-UPDRS-III rating scale) and identification of important features 122 gait features Accelerometry-Based Digital Gait Characteristics for Classification of Parkinson’s Disease: What Counts? [4] 2020 Classification of PD (non-PD, PD) and identification of important features 210 gait features Gait and tremor investigation using machine learning techniques for the diagnosis of Parkinson disease [5] 2018 Classification of PD (non-PD, PD) stance time, swing time, and stride time Biometric and mobile gait analysis for early diagnosis and therapy monitoring in Parkinson's disease [6] 2011 Classification of PD (non-PD, early PD, intermediate PD) 12 gait features 2. Method 2.1 Dataset The WearGait-PD [7] dataset was used for this analysis. This dataset is one of the first openaccess database that provides synchronized recordings of raw IMU signals, insole pressure sensor data, and pressure-sensing walkway measurements of PD patients. The dataset is composed of multiple sensor modalities including IMUs placed at 13 body locations, insole IMUs, and walkway foot contact information that provide binary labels (1 for contact and 0 for no contact) of left or right foot contact. IMU data were collected from the subjects performing eight tasks: self-paced walking, hurried pace, self-paced walking on a mat, hurried pace on a mat, self-paced turning on a mat, tandem gait, the Timed-Up-and-Go (TUG) test, and balance. For the analysis, of the 123 subjects, 20% (25 subjects) was reserved as an external test set. The data from the remaining 98 subjects were used for training, of which 14 were set aside as an internal validation data. 2.2 Model Architecture Figure 2: Model Architecture where “WN” refers to weight normalization, “Conv” refers to convolutional layer, and “ReLU” refers to rectified linear unit activation. The 1D CNN-based neural network model was used to predict the output (binary prediction for left and right foot contact per datapoint) from the input (time series input of 6 or 12 channels). The input passes through seven layers of 1-D CNN layers with the classifier layer attached at the end for two binary predictions (left and right foot contact). Skip connection was used to give the middle convolutional layers more context of the original input [8-9]. Detailed model architecture is presented in Figure 2. 2.2 Performance Measurement Point-wise accuracy for binary classification was calculated for both performance evaluation measurement, and training loss. The accuracy was calculated using binary cross entropy where the loss L is defined: 𝐿 =−1 𝑁&[ ! "#$ 𝑦"log(𝑝")+(1−𝑦")log(1−𝑝")0], where N is the total number of samples, 𝒑𝒊 is the predicted probability of contact for the ith sample, and 𝒚𝒊 is the true (ground truth) label for the ith sample with i being the index for data points. 3. Results 3.1 Step Measurement Accuracy Our results demonstrate that ankle sensors yield highly accurate predictions of foot contact events, achieving 97% point-wise accuracy across all test samples. However, prediction performance is less reliable when using sensors placed at other body locations, highlighting the importance of sensor placement in gait-based models. Table 2: Point-wise accuracy per sensor inputs (column) and prediction target (rows) on the test set Ankle Lower back Wrist Right contact 97.23 92.97 91.79 Left contact 97.17 92.57 91.63 Overall 97.20 92.77 91.71 Figure 3: Example plots showing accelerometer [x, y, z] inputs (blue, yellow, and green respectively), left foot contact prediction (red), left foot contact ground truth (black), and different types of movement denoted as background colors. The plots are of ankle (top), wrist (middle), and lower back (bottom) sensors. 4. Discussion Conclusion In this work, we proposed a CNN-based step measurement model which can be used to produce clinically meaningful gait features for PD predictions. Our results demonstrates that the IMU sensors placed on ankle produce the best result when it comes to step measurement with 97% point-wise accuracy, followed by lower back and wrist placements. Caveats Despite promising preliminary results, more investigation is further required in model fine-tuning and sensitivity analysis. Model performance can be further improved through several refinements with the following approaches. First, smoothing can be applied over physiologically implausible “spikes” in predicted foot contact when the predicted step duration is shorter than a pre-defined threshold. Second, sensitivity analysis can be performed regarding the probability threshold for declaring a contact event (currently set at 0.5). Finally, incorporating additional reference signals such as foot pressure measurements could provide complementary information for model improvement. Challenges There remains challenge in adopting the walkway contact data as the ground truth in training. Cases such as partial footfalls or edge cases, where a subject’s foot contacted the walkway pad only partially or the step is made entirely out of the pad, are often not counted as contact in the walkway data, whereas the model may correctly detect them. This discrepancy complicates the evaluation process and necessitates additional assumptions regarding model performance when computing accuracy metrics. A more flexible definition of ground truth may be required to account for such cases. Future Direction Future work involves validating the model on external datasets that lack explicit contact labels but have higher-level gait information such as step counts or durations. Given such datasets, an evaluation can be made between the original gait features and the features that are derived based on our step measurement model. This type of feature-level validation would demonstrate our model’s generalization capabilities extending to external datasets and further proves our model’s ability as an effective open-source framework in creating clinically meaningful features. References 1. Nirmal, K., and A. G. Sreejith. "Noise modeling and analysis of an IMU-based attitude sensor: Improvement of performance by filtering and sensor fusion." Advances in optical and mechanical technologies for telescopes and instrumentation II. Vol. 9912. SPIE, 2016. 2. Tedaldi, David, Alberto Pretto, and Emanuele Menegatti. "A robust and easy to implement method for IMU calibration without external equipments." 2014 IEEE international conference on robotics and automation (ICRA). IEEE, 2014. 3. Sotirakis, Charalampos, et al. "Identification of motor progression in Parkinson’s disease using wearable sensors and machine learning." npj Parkinson's Disease 9.1 (2023): 142. 4. Rehman, Rana Zia Ur, et al. "Accelerometry-based digital gait characteristics for classification of Parkinson's disease: what counts?" IEEE Open Journal of Engineering in Medicine and Biology 1 (2020): 65-73. 5. Abdulhay, Enas, et al. "Gait and tremor investigation using machine learning techniques for the diagnosis of Parkinson disease." Future Generation Computer Systems 83 (2018): 366-373. 6. Barth, Jens, et al. "Biometric and mobile gait analysis for early diagnosis and therapy monitoring in Parkinson's disease." 2011 annual international conference of the IEEE engineering in medicine and biology society. IEEE, 2011. 7. Anderson, Anthony J., et al. "WearGait-PD: An Open-Access Wearables Dataset for Gait in Parkinson’s Disease and Age-Matched Controls." medRxiv (2024): 2024-09. 8. Zhou, Zongwei, et al. "Unet++: Redesigning skip connections to exploit multiscale features in image segmentation." IEEE transactions on medical imaging 39.6 (2019): 1856-1867. 9. Drozdzal, Michal, et al. "The importance of skip connections in biomedical image segmentation." International workshop on deep learning in medical image analysis. Cham: Springer International Publishing, 2016.