A Systematic Literature Review of Advanced Machine Learning Techniques in Wireless Body Area Networks: Application, Challenges, and Future Directions

Received 27 October 2025, accepted 6 November 2025, date of publication 10 November 2025, date of current version 19 November 2025. Digital Object Identifier 10.1109/ACCESS.2025.3631230 A Systematic Literature Review of Advanced Machine Learning Techniques in Wireless Body Area Networks: Application, Challenges, and Future Directions ABDU IBRAHIM ADAMU 1, PRAVEEN KUMAR DONTA 2, (Senior Member, IEEE), DARMAWATY MOHD ALI 1, SOHAIL SARANG 3, (Senior Member, IEEE), GORAN M. STOJANOVIĆ 3, (Member, IEEE), AND SUZI SEROJA SARNIN1 1Wireless Communication Technology Group (WiCOT), Faculty of Electrical Engineering, Universiti Teknologi MARA (UiTM), Shah Alam, Selangor 40450, Malaysia 2Department of Computer and Systems Sciences, Faculty of Social Sciences, Stockholm University, 114 19 Stockholm, Sweden 3Faculty of Technical Sciences, University of Novi Sad, 21000 Novi Sad, Serbia Corresponding author: Darmawaty Mohd Ali ([email protected]) This work was supported in part by European Union’s Horizon Europe European Innovation Council (EIC) 2023 Pathfinder Challenge Program under Grant 101161032, in part by the Ministry of Higher Education Malaysia (MOHE) for the Fundamental Research Grant Scheme (FRGS) under Grant FRGS/1/2023/TK07/UITM/02/29, and in part by the Universiti Teknologi MARA (UiTM). ABSTRACT The development of machine learning (ML) in wireless body area networks (WBANs) has made great progress in healthcare monitoring. The sensors are made wearable to monitor the physiological data without any time lapse, continuously. Importantly, issues such as energy consumption, reliability of sensor data, patient privacy, and the need for transparent, interpretable models remain significant barriers to the application of this technology in clinical environments. This paper presents a systematic literature review (SLR) of peer-reviewed studies published between 2017 and 2025, conducted in accordance with Preferred Reporting Items for Systematic Review and Meta-analysis (PRISMA) guidelines to ensure rigor and reproducibility. From a total of 2,407 publications screened, 55 studies met the inclusion criteria. This SLR investigates how WBANs and ML have been applied to support anomaly detection, activity recognition, and lightweight data transmission to facilitate personalized healthcare. We compare the stated performance metrics of four different ML approaches, supervised learning, unsupervised learning, reinforcement learning (RL), and hybrid approaches, and map each to WBAN application contexts. To tackle these challenges, advanced ML techniques are studied, including generative artificial intelligence (GAI), federated learning (FL), lightweight models, representation learning, deep learning (DL), RL, and autoencoders. Moreover, low-latency deep neural networks (DNNs), edge computing, and eXplainable AI (XAI) techniques are recommended to increase interpretability and enable real-time decision-making. This synthesis highlights persistent gaps in energy efficiency, scalability, privacy preservation, and standardized evaluation. It also lays out a specific research agenda to direct future investigations. We conclude that four critical issues protecting patient data, extending battery life, processing data quickly, and ensuring accurate and reliable sensor readings must be addressed if ML-powered wearables are to genuinely transform the healthcare industry. INDEX TERMS Anomaly detection, energy efficiency, machine learning, personalized healthcare, real-time monitoring, wireless body area networks. The associate editor coordinating the review of this manuscript and approving it for publication was Ayman El-Baz . I. INTRODUCTION Through wireless body area networks (WBANs), wearable sensors are networked to monitor and transmit VOLUME 13, 2025 2025 The Authors. This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 194729 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques physiological data from the human body, making WBANs a rapidly evolving healthcare technology. These networks are proximity-based and utilize low-range wireless communication protocols, such as Bluetooth and Zigbee, for real-time health monitoring [1]. WBANs are primarily applied in various areas, including telemedicine, chronic disease management, and exercise monitoring, assisting healthcare providers in wirelessly taking patients’ vital signs and other clinical indicators [2]. With recent advancements in sensor technology, miniaturization, and energy harvesting, WBAN devices’ performance, reliability, and accessibility have significantly improved, resulting in their greater usage in everyday life [3]. The increasing prevalence of chronic health conditions has increased the need for ongoing healthcare monitoring, which has sped up the adoption of WBAN in sectors like sports, rehabilitation, and elder care [4]. Recent advances in artificial intelligence (AI) and machine learning (ML) have caused unprecedented growth in multiple domains due to improved algorithms, higher processing power, and the availability of giant datasets. The worldwide AI market is projected to reach $1.4 trillion by 2029, at a compound annual growth rate above 20% [5]. Advances in deep learning (DL) algorithms, such as transformer models in natural language processing and generative adversarial networks (GANs) in image synthesis, have further elevated AI’s potential. These technologies have the potential to completely transform healthcare, as evidenced by their growing use in predictive analytics, personalized medicine, and drug discovery [6]. When integrated with WBANs, AI/ML can enable physicians to process biosensor data in real time, create personalized care programs, and detect health issues early [7]. WBANs, enhanced by ML, can perform critical tasks such as activity recognition for personalized fitness plans and anomaly detection to identify unusual changes in vital signs, potentially preventing severe health complications. For instance, monitoring heart rate variability through ML could assist in early cardiac problem detection, demonstrating the transformative potential of these systems. Furthermore, AI-driven communication protocol optimization in WBANs improves data transfer efficiency and ensures timely access to critical health information. Nonetheless, there are still ongoing issues, especially in data privacy, which requires the protection of private health information, and energy efficiency, where battery life is a limiting factor. Ongoing research addresses these issues by developing privacy-preserving models that safeguard patient data without sacrificing performance, as well as energy-efficient algorithms that extend device lifespan. In addition, several surveys have reviewed WBANs; most focus on IoT integration, energy-efficient routing, or specific ML applications such as human activity recognition. Others emphasize energy optimization, classification techniques, WBAN design, security, or generative AI in niche healthcare tasks. However, no one offers a comprehensive, ML-focused synthesis that links algorithmic techniques directly to WBAN healthcare challenges. Along with improvements to algorithms, improvements to wireless communication hardware and sensing, like microwave-based sensors, substrate-integrated and leaky-wave antenna systems, will be very important for integrating ML into WBANs to make physiological monitoring more accurate and flexible [8],[9],[10],[11]. The current study fills these gaps by conducting an SLR on advanced ML/AI applications in WBANs using the Preferred Reporting Items for Systematic Review and Metaanalysis (PRISMA) guidelines, with an emphasis on solutions for energy efficiency, quality of service (QoS), security, anomaly detection, activity recognition, communication, and healthcare applications. Unlike previous surveys, this review follows a structured, reproducible methodology to consolidate fragmented knowledge, classify ML approaches, and identify persistent research gaps by answering the following research question: ‘‘How can ML algorithms be utilized in WBANs to solve problems, improve performance, and direct future advancements in healthcare applications?’’ This review’s main goals are to summarize the state-of-the-art in ML-driven WBANs for healthcare applications, point out the advantages and disadvantages of current strategies for tackling issues with energy, security, and real-time data analysis, and draw attention to unexplored areas while suggesting future lines of inquiry. The following are the primary contributions of this study: •The four primary application areas of ML-enabled WBANs are methodically identified in this systematic literature review (SLR). It introduces a new taxonomy that frames cutting-edge methods like generative AI and federated learning (FL) within the WBAN environment, while highlighting activity recognition, anomaly detection, communication optimization, and personalized care as the primary domains. •The SLR pinpoints various directions, including energy consumption, along with data-quality constraints and privacy risks, and opaque model behavior as the main obstacles for clinical adoption. The study proposes specific strategies, including lightweight model design and explainable AI (XAI) techniques, which enhance real-time system performance while reducing power usage and maintaining patient privacy. •To facilitate early disease detection, ongoing vital sign monitoring, and intelligent resource management, all of which contribute to a patient-centred approach to healthcare, our suggested architectural review integrates well-known WBAN hardware with deep reinforcement learning (DRL) elements and hybrid algorithms. •The comparative analysis demonstrates trade-offs between accuracy and latency against energy efficiency and interpretability while also identifying scalability 194730 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques limitations in large-node WBAN deployments. The assessment determines which optimization areas require immediate attention. •To highlight unresolved issues, such as privacy protection in high-volume sensor streams, WBAN interference management, edge-based ML processing, and neural interface integration, the analysis synthesizes findings from various disciplines. It then suggests state-of-the-art technical solutions with strong data security and privacy safeguards that adhere to ethical and legal standards. This will help to guide future research, build patient trust, and promote collaborative innovation. For the reader’s convenience, we have included a list of acronyms used in Table 1. The remainder of this paper is organized as follows: Section II presents the background of the study. Section IV Related Works. Section IV provides a systematic literature review. In Section V, advanced ML for WBANs is explained. Section VI presents a discussion and statistics of the findings. Section VII discusses the challenges and future research directions for integrating ML into WBANs, and conclusions and suggestions for additional research are presented in Section VIII. II. BACKGROUND OF THE STUDY According to a study by Abdu et al. in [12] WBANs are WSNs positioned as an emergent technological paradigm that intends to help healthcare systems function more effectively. WBANs are intentionally designed to support and allow real-time health monitoring and focus on timely intervention and treatment of people with life-threatening conditions [13]. Sensors in WBAN operate within or outside the human body to collect various physiological signals and transfer them to a central node called the ‘‘coordinator’’ node [14]. WBANs have attracted much attention because of their potential applications in healthcare, which enable them to routinely monitor important physiological signs, track patient mobility, and aggregate data related to various medical diseases [15]. This makes them ideal for remote patient monitoring, fall detection, and chronic disease management applications. A distinct trend in scholarly discourse is a strong emphasis on developing miniaturized, low-power sensors and wearable devices that will smoothly fit with the operational framework of WBANs [16]. Physiological data such as blood pressure, glucose levels, body temperature, heart rate, and muscle activity could be measured by these sensors [17]. Researchers have investigated a range of communication technologies, such as Bluetooth, Zigbee, UWB, and 5G. Every technique has pros and cons, and the decision is usually influenced by the requirements of the application [18]. WBAN devices must operate on a limited amount of battery power. Power consumption is an important consideration in increasing the lifespan of WBAN devices [19]. Many studies have concentrated on sensor designs, power management techniques, and energy-efficient communication protocols [20]. TABLE 1. Commonly used abbreviation in the SLR. The emergence of wearable platforms such as smartwatches and fitness trackers has boosted WBAN growth [21]. VOLUME 13, 2025 194731 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 1. (Continued.) Commonly used abbreviation in the SLR. These consumer-oriented devices have the potential to acquire useful health-related data and play a significant role in future healthcare applications. The WBAN communication architecture consists of three primary functional components, each of which is essential to enabling efficient data transfer and network connection [22],[23]. Nevertheless, Figure 1 shows the details. Tier-One: Since they serve as the basis for effective data interchange and network interaction, the communicative dynamics in the WBAN framework constitute a crucial focus of this paper. This tier focuses on the interaction of biomedical sensor nodes, which measure physiological parameters, including body temperature, heart rate, oxygen saturation, and blood pressure, and send the gathered data to a selected coordinator or sink node. WBANs rely on well-defined wireless standards to keep sensors talking while using as little power as possible. The most common choice is the IEEE 802.15.6 protocol, built specifically for body-worn devices that balance low energy use with reliable links and quality-of-service support [25]. Depending on the job, how far the signal must travel, how quickly data needs to move, and how much battery life matters, designers also turn to Bluetooth Low Energy, Zigbee, or LoRa. That choice shapes the network’s day-to-day performance, influencing delays, battery drain, and resistance to interference. At the link layer, the system manages data aggregation, collision avoidance, and channel access so that every sensor can talk to the coordinator without hiccups. To push performance further, engineers add functions like cooperative relaying, dynamic channel selection, and adaptive modulation and coding. These upgrades help the network stay solid even when signals fade, patients move around, or outside devices cause noise. Running light data-processing tasks on the coordinator, an edge-computing step, also eases the burden on tiny sensor nodes and cuts transmission lag. Together, these measures keep the WBAN scalable, compatible with other gear, and quick enough for real-time monitoring. They lay the groundwork for higher-level tasks such as data analysis, anomaly detection, and automated medical decisions. Tier-Two: The inter-WBAN layer handles everything beyond the coordinator inside a single body network. It covers bodyto-body (B2B) links, overlapping WBANs, and connections to external access points [26]. Its job is to enable multiple WBANs to communicate with one another and with external systems, cloud servers, hospital databases, or remotemonitoring dashboards, allowing data to flow smoothly wherever it needs to go. Most links at this level utilize familiar wireless technologies, including Wi-Fi, cellular (4G/5G), or long-range, low-power options such as LoRaWAN. The choice depends on range, bandwidth, delay, and power use. LoRaWAN, for instance, is great when you need kilometres of reach but have tight battery limits, while 5G shines in real-time scenarios that demand millisecond latency and reliability [27]. Issues are common in crowded areas, busy hospital wards, and retail malls. Signal clashes occur when numerous WBANs use the same frequency spectrum. Smart interference-mitigation strategies, dynamic spectrum allocation, and frequency hopping are some of the techniques that keep the frequency clear. There are additional obstacles for B2B links: individuals move randomly, and bodies obstruct signals. Cooperative relays or adaptive routing protocols help sidestep those issues [28]. Interoperability matters just as much as connectivity. Standards such as IEEE 802.15.6, along with the latest 5G and IoT specs, lay out the rules for secure, efficient data exchange. Airtight security, which includes end-to-end encryption, blockchain-based authentication, and privacypreserving aggregation, is essential due to the sensitive health data involved. By tackling interference, following open standards, and leaning on modern wireless tech, this layer underpins a robust, scalable WBAN ecosystem [29]. In short, it is the bridge that links on-body networks to the wider healthcare world. Tier-Three: The Beyond-WBAN layer allows for the transmission of information to other destinations, like healthcare facilities, cloud storage, or telemonitoring systems. This layer establishes a communication model that transcends individual WBAN boundaries. In this layered architecture, the AP plays an essential role by allowing the passage of data packets from the WBAN to whatever destination via multiple communication systems. It covers all channels that ensure seamless connectivity and global reach: satellite communications, cellular networks, the Internet, and WANs. This tier should especially be considered by applications such as telemedicine, remote health monitoring, and big data analytics. For example, wearable sensors in a WBAN transmit 194732 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 1. WBAN architecture [24]. patient information to centralized healthcare systems that integrate the data with electronic health records (EHRs), store it for the long term, or analyze it in real time. By enabling effective data processing, storage, and decision-making at remote locations, the use of cloud and edge computing enhances this tier’s capabilities. The Beyond-WBAN layer addresses one of the most important challenges: ensuring reliable and secure data transmission through volatile or high-latency communication channels. Multi-path routing, error correction mechanisms, and QoS enhancement solutions play a vital role in maintaining data integrity while minimizing latency. Also, these techniques compress and aggregate the data to use less bandwidth and transmit more efficiently, since many WBANs generate massive amounts of information [30]. Because information sent outside the WBAN perimeter can readily travel through untrusted networks, increasing the risk of interception and unauthorized access, security and privacy are crucial issues. Advanced security measures such as blockchain-based authentication, secure tunnelling protocols like TLS/SSL, and end-to-end encryption can indeed mitigate these threats [31]. Moreover, while supporting collaborative analytics among different WBANs, privacy-preserving approaches such as FL and differential privacy ensure that confidential health information remains protected. Standardization at this level is also required because interoperability is what allows seamless data exchange between different healthcare systems and communication technologies. In Beyond-WBAN environments, standards such as IEEE 802.15.6, IoT-specific standards, and 5G-enabled healthcare frameworks provide safe and effective communication guidance while also ensuring better collaboration among these systems. Integrating ML and AI at this level further enables predictive analytics, anomaly detection, and personalized healthcare recommendations from the aggregated data of diverse sources. The Beyond-WBAN tier creates a crucial link for data dissemination outside the immediate WBAN area by tackling these issues and utilizing innovative communication and security technologies, as in [31]. This stratum enables remote diagnostics and real-time health monitoring and supports extensive healthcare programs like disease surveillance, population health management, and international health research. A. AREAS FOR WBAN APPLICATION WBAN has several present and potential uses as an extension of WSN [32]. Widespread WBAN use in people’s daily lives is made possible by the current increase in internet speed, the global number of internet users, the affordability of wearable technology, and advancements in AI and big data analytics. WBAN conforms to networked devices (wireless sensors) communicating across short distances within or outside the body. WBAN has several medical applications, as depicted in Figure 2. 1) MILITARY TRAINING AND SPORTS Several death incidents in sports have been reported, and medical results revealed that most casualties were caused by the failure of the body’s numerous important organs due to fatigue. Recently, wearable WBAN devices have been used to monitor the health conditions of athletes during physical training and military soldiers in various military operations [13]. VOLUME 13, 2025 194733 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 2. WBAN applications. 2) RESCUE OPERATIONS AND DISASTER MANAGEMENT Communication and other infrastructures often fail to perform their expected functions during a crisis. Rescue workers might be unable to complete a rescue operation if the dependable communication network malfunctions. WBAN coverage distance is limited, except for its use in information monitoring (health, etc.). As a result, the authors recommended using WBAN in addition to CR. Using the unlicensed frequency spectrum outside the ISM band, the CR transmits data based on opportunistic channel access, connecting data in the WBAN network to the targeted health units [33]. In smart homes, WBAN sensors (gyroscopes and accelerometers) are used to track the movements and shifting postures of elderly residents receiving assisted living. People can also easily connect to WBAN healthcare systems using remote controls to reduce their dependence [13]. When serious occurrences such as the development of a slump need to be addressed, monitored actions alert other family members and caregivers to intervene. 3) CONSUMER ELECTRONICS AND ENTERTAINMENT Millimeter-wave devices enable very short-range wireless links between gadgets. Wearables such as smartwatches and fitness bands can connect to smart TVs, power gaming controllers, or stream MP3s over Bluetooth, demonstrating how WBANs integrate into today’s entertainment and daily routines [34]. As technology develops, we witness increasingly customized wearables and smart gadgets. Wearable technology has evolved beyond health to become a fashion statement, thanks to 5G’s higher speeds. 4) EMERGENCY HEALTHCARE SERVICES, SURVEILLANCE, AND DEVICE MONITORING Surveillance devices, such as unmanned aerial vehicles and integrated satellites, use high-definition cameras and other sensors to track a range of social activities, such as position changes, interaction, gait movement, and the transfer of health data to data servers via UWB transceivers. In addition to surveillance, unmanned aerial vehicles (UAVs) linked to WBANs can be used to collect visual health data from several sources. Depending on the type of data, controllers, doctors, and engineers must react to current circumstances based on the health status of people or equipment [35]. In remote or isolated places, WBAN paired with satellites can provide health emergency assistance with minimal delays in physiological signals. In remote devices, embedded sensors (e.g., WBAN) measure various attributes during communication and coordination. Remote devices can interact with the control station, save data in the logger (or servers), and independently control their operations, depending on the embedded intelligence [36]. 5) HEALTH MONITORING WBAN is a crucial component of health monitoring. WBAN sensors have been integrated into the human body as wearable devices, surface contacts, or implants to collect and transmit vital patient health data to remote medical facilities. Linked biosensors measure numerous parameters, including blood sugar, heart rate, and body temperature. Acquired health data are utilized to predict the course of a disease or to recommend remedial actions using AI algorithms [37]. Moreover, WBAN has numerous uses in stroke-related illnesses, diabetes management, early cancer cell detection, rehabilitation progress tracking, and critical care unit decision assistance [34],[38]. 6) TELEMEDICINE WBAN offers network infrastructure for telemedicine services, enabling patients to receive virtual health care rapidly, in addition to monitoring astronauts’ health during space exploration [29]. Health status evaluations and decisions are made by coordinating biological sensors, data processing units, and miniature actuators. The actuators then perform the decision as feedback. For instance, if it senses elevated blood sugar, the processor instructs the actuator to regulate insulin in the blood arteries. B. MACHINE LEARNING Doctors now have a new assistant in AI that can help them with diagnosis and even prognosis. A thorough understanding of a patient’s problems may also be facilitated by the capacity to learn from experience. WBANs that employ AI algorithms enable the creation of computer programs that learn and develop from experience, instead of being specifically designed to make predictions or recommendations. Over the 194734 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques last few decades, advances in computer power have enabled the development of resource-intensive AI techniques such as ML solutions. This study also illustrates how ML programs statistically create a predictive model using data samples or information for training. This training is used to classify objects or make predictions in applications necessary for wise decisions. Health data collected from sensors is processed by smart healthcare apps employing ML logic before being sent to the cloud for processing by ML algorithms. The results were sufficiently replicated, and the collected data were referred to as the testing data. The outcomes will also be utilized in the training phase of the subsequent testing data after they have been replicated. As a result, the data collected by the sensors is considered testing data and, after processing, training data for additional medical evaluations. Robust implementation, detection accuracy, and computational cost are crucial when selecting ML algorithms for WBAN applications [39]. DL, supervised learning, and unsupervised learning are the main categories in which ML systems are incorporated with WBAN applications. Figure 3illustrates how ML programs statistically create a predictive model using data samples or information for training. FIGURE 3. Machine learning process [39]. 1) SUPERVISED LEARNING Supervised learning is the process of training algorithms to classify or predict data or outcomes correctly using labelled datasets. The objective is to teach the algorithm a general rule that links inputs to outputs by providing instances of inputs and the expected outputs that they should generate. WBANs use supervised learning algorithms, especially where data can be appropriately categorized and a wealth of prior knowledge about the human body is available. One prominent example is the estimation of a mobile node’s location using an algorithm trained on signal propagation characteristics (inputs) and designated locations (outputs). The application of supervised learning has effectively handled several obstacles within the field of WBANs, including the MAC [40], routing [41], disease prediction in WBAN [42], detection of human body movements in WBANs [43], human activity recognition [44], and security [45],[46]. 2) UNSUPERVISED LEARNING Unsupervised learning refers to algorithms that find patterns in datasets, including unlabelled and uncategorized data points. The learning algorithm acts as a feature extraction tool, uncovering hidden patterns in the data without labelling. Problems, such as automatically clustering wireless sensor nodes according to their current observed data values (without knowing the group membership of each node beforehand), can be solved with unsupervised learning. Anomaly detection, such as clustering or outlier detection, can be used to identify unexpected patterns or abnormalities in physiological data acquired by WBAN sensors [47],[48]. This is particularly useful in healthcare applications, where deviations from normal patterns may indicate health issues or sensor malfunctions. Data compression and dimensionality reduction can be achieved using unsupervised learning methods, such as PCA. WBANs benefit from this solution because it enables the transmission of vital information while minimizing data transfer. Patient profiling through clustering algorithms enables the grouping of patients based on similar health characteristics drawn from WBAN sensor information [49]. This can help create detailed patient profiles, making it possible to offer personalized healthcare treatments and approaches [50],[51]. It also supports network self-organization, such as enabling WBAN nodes to group themselves based on shared characteristics and form clusters [52]. This self-organization makes network communication more efficient overall by allowing for better resource allocation and coordination, as well as dynamic channel allocation, which means that channels can be analyzed for conditions and interference patterns without supervision. This information is then utilized to dynamically assign channels, decrease collisions, and improve the dependability of data transmission within a WBAN [53]. 3) REINFORCEMENT LEARNING RL is a prominent advanced subfield of ML [54]. Here, an agent is assigned the responsibility of making a sequence of decisions while interacting dynamically with its environment. The basic objective of RL lies in improving the decision-making policies to accumulate a reward signal, which is attained through iterative exploration based on trial and error. This process is formalized using the framework of a Markov Decision Process (MDP), which offers a structured mathematical model for problems involving sequential decision-making under uncertainty. RL, along with supervised and unsupervised learning, is one of the three core ML paradigms. RL algorithms can sense and understand their surroundings, act and learn from failures, and choose the best action for intelligent agents to maximize the cumulative reward in each environment. The learning algorithm interacts with a dynamic environment as it moves through its problem area and receives feedback that it may compare to rewards, which it aims to optimize [39] VOLUME 13, 2025 194735 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques The research that depicts one type of RL technique, which is the Q-learning procedure, is used for slot distribution in a WBAN setting. This illustrates how Q-learning can be applied to intelligent and adaptive resource allocation in WBANs by highlighting the interactions among the agent, environment, state transitions, and incentives. The framework’s multiple components operate in a reinforcement learning loop to maximize resource allocation [55]. An agent representing one of the WBAN nodes begins the process by interacting with its surroundings. The agent tracks the WBAN nodes’ current state (St), which encompasses several network states and metrics. Based on this state, the agent chooses an action (at), like how to assign communication slots, to keep everything running smoothly. In the WBAN environment, nodes communicate with each other to carry out an action and see what happens. If the action works well, the environment moves to a new state (St+1), and the agent gets a reward (Rt). This reward shows how well the chosen slot allocation helps improve system performance, like reducing delays or saving energy. The Q-learning algorithm helps the agent improve its decision-making based on the rewards it receives and the new situations it encounters. Through repeated learning, the agent gradually figures out the best way to allocate time slots. It does this by balancing the need to try new actions (exploration) using what it already knows works well (exploitation), allowing it to adjust to changes in the network. This adaptability helps improve the performance of the WBAN system. Figure 4shows how Q-learning enables smart and flexible resource management in WBANs by illustrating the interactions between the agent, the environment, state changes, and reward signals. FIGURE 4. Reinforcement learning. 4) DEEP LEARNING DL is a subset of supervised learning techniques where output is generated by processing input through some nonlinear transformations. Reference [56] claims that DL makes it possible for computational models composed of multiple processing layers to learn data representations at different levels of abstraction. The capacity of DL to automatically extract high-level features from complex data is a key advantage over traditional ML systems. Handcrafting previous features is much simplified by the fact that the learning process does not have to be constructed by a human [56]. However, the interpretability of the model suffers due to the performance of DL techniques, such as deep neural networks (DNNs). Since DNNs make unique decisions, they are sometimes perceived as ‘‘black boxes’’ with little understanding. Additionally, DNNs often have numerous hyperparameter tuning issues, making it difficult and time-consuming to determine the best configuration. Training DL networks can also be computationally expensive, requiring robust parallel computing capabilities like graphics processing units (GPUs). Therefore, it is crucial to consider the energy and computing limitations of embedded or mobile devices when deploying DL models. Furthermore, DL can learn to perform categorization tasks directly from text, audio, or images. Synthetic neural networks (NNs) support these functions. Sensors collect vital sign data, allowing AI to identify patterns instantly. AI-powered triage systems expedite emergency response by prioritizing cases based on their urgency. Large volumes of data generated by biosensors are readily and swiftly processed with the help of AI approaches. This quick processing reduces latency by facilitating timely diagnosis. AI-enabled remote monitoring devices, automated patient screening, real-time alerts, and predictive algorithms also help physicians make informed decisions and significantly reduce reaction times. NN architectures with multiple layers and substantial amounts of sample data are used to develop DL models to reach state-of-the-art accuracy [57]. Two examples of analytical DL models that provide computational intelligence solutions by studying vast datasets are CNN and deep belief networks, which are used when shallow learning cannot assess the necessary meaningfulness of trends. Precision medicine often uses these learning models to diagnose illnesses and generate new treatments [58]. C. MACHINE LEARNING ALGORITHMS This section takes a close look at the ML approaches commonly used in WBAN research, with a critical review of their strengths and limitations. 1) SUPPORT VECTOR MACHINE SVM is a method used to solve classification problems. It works by transforming the data into a higher-dimensional space, where a hyperplane can separate the different classes clearly. Figure 5illustrates the four-layered architecture, showing how WBANs can be integrated with ML algorithms to enhance healthcare services. 194736 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 5. Four-layered architecture illustrating how WBANs can be combined with ML algorithms to improve healthcare services [59]. The BAN layer uses biosensors throughout the body to track critical indications, including temperature, oxygen saturation, and heart rate. These sensors continuously collect physiological data. However, the user interface layer allows cellular networks or a wireless router to gather data wirelessly and send it to consumer devices, including smartphones and tablets (such as the iPhone). This layer makes it easier to retrieve data initially and connect to the internet for additional processing. In addition, the information that is sent to a server that applies supervised ML techniques, including random forest (RF) and decision tree (DT), and SVM stages, including data collecting, filtering (to eliminate noise or unnecessary information), and analysis, makes up this layer. After analysing the filtered data, the algorithms produce useful insights for healthcare decision-making. Lastly, doctors and other medical professionals can utilize the ML analysis’s conclusions for monitoring, diagnosis, or emergencies. Doctors can evaluate patients’ health and administer treatments from a distance. Emergency services are notified to attend urgent medical demands in critical circumstances. 2) RANDOM FOREST One type of supervised ML method that works well for both regression and classification problems is RF. It falls under ensemble learning strategies, especially the bagging (bootstrap aggregating) technique. A bagged decision tree-based ensemble learning technique is called RF. Bootstrap aggregating, or bagging, is training several classifiers on various subsets of the training data and averaging their results to generate the final prediction. By successfully lowering the ensemble model’s variance, this method improves the model’s stability and prediction precision. DTs, sensitive to input data changes and prone to high variance and overfitting, are especially well-suited for bagging. RF is an effective tool for classification and regression tasks because it combines the outputs of several decision trees to produce strong and dependable predictions. Figure 5represents a four-layered architecture illustrating how WBANs, combined with SVM, RF, and DT algorithms, improve QoS. 3) K-MEANS To arrange data into mutually exclusive groups (or clusters), K-means aims to make observations from the same cluster as similar as possible while keeping observations from different clusters as diverse as possible. In K-means clustering, each cluster’s center, or centroid, corresponds to the means of the observation values assigned to the cluster. Assigning several points to the number of groups or clusters is intended to produce high intra-cluster and low inter-cluster similarities [60]. However, the data that is produced in WBAN is diverse. Additionally, the distribution of biosensors is not consistent. The study demonstrates that the effectiveness and dependability of data collection and communication in WBANs are significantly impacted by data clustering using an unsupervised ML technique like K-Means [61]. The clustering VOLUME 13, 2025 194737 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 2. (Continued.) Comparison of recent and related surveys on the WBAN application. as well as wireless technologies and architectural frameworks like LPWAN, 5G, and 6G, was the focus of early works [13], [97]. These contributions provided thorough mappings of the technologies and system architectures that were available, but they frequently lacked viewpoints on ML and AI, which limited their ability to meet the increasing demand for data-driven and adaptive WBAN solutions. By 2022, surveys had become more specialized, with several studies concentrating on data classification and energy-efficient routing [98],[99],[100],[101]. These studies acknowledged the potential of ML to enhance the QoS and optimize routing. Nevertheless, they provided little advice on how to implement ML in WBANs, even though they acknowledged it as a promising tool. Given the resource-constrained nature of WBAN devices, where the direct application of traditional ML models is limited by computational complexity, energy constraints, and hardware integration issues, such advice is especially crucial. Applications for HAR and healthcare were investigated in parallel [102],[103], where ML showed distinct advantages in classification and adaptive learning. However, problems like interpretability, concept drift, and integration with real-time communication channels were still unsolved. Considering the growing significance of WBANs in delicate healthcare settings, more recent surveys (2023–2024) placed a greater emphasis on security and privacy. Threats to data integrity and confidentiality were studied in studies like [6],[104],[105], and [106]. Some even introduced generative AI models to improve WBAN security [79]. Domain-specific applications such as medical imaging for the diagnosis of lung cancer also demonstrated the expanding role of deep learning in WBAN-enabled healthcare [107]. Despite these developments, most surveys remain primarily conceptual in nature and fail to provide practical validation. Furthermore, even though ML/AI viewpoints are being introduced gradually, interpretability, computational viability, and privacy-preserving strategies like lightweight cryptographic models or FL are frequently overlooked in discussions. Considering this, our survey’s contribution is its thorough integration of ML/AI viewpoints from all thematic areas of WBAN research, such as architecture, routing, QoS, energy efficiency, privacy, and healthcare applications. This survey sheds light on the specific obstacles preventing ML adoption in WBANs, in contrast to previous works that either ignore AI considerations entirely or concentrate only on a limited number of technologies [4]. These include the interpretability of predictive models, privacy issues, data quality and variability, and energy and resource limitations. Our survey aids in the creation of intelligent, safe, and effective healthcare applications by bridging communication-centric and AI-driven viewpoints. It also offers a comprehensive roadmap for furthering WBAN research. Table 2summarizes the details. IV. SYSTEMATIC LITERATURE REVIEW METHODOLOGY FOR ML IN WBANs This study’s methodology follows the PRISMA and MetaAnalyses framework, which ensures that the systematic review can be conducted in an open and repeatable manner [108]. Identification, screening, eligibility, and inclusion are the four primary steps in the PRISMA framework. The identification step involved a comprehensive search using pre-set keywords related to advanced ML techniques in WBANs in reputable databases such as Web of Science, Scopus, and IEEE Xplore. In addition to terms like ‘‘Machine Learning,’’ ‘‘Deep Learning,’’ and ‘‘Reinforcement Learning,’’ the search query included ‘‘Wireless Body Area Networks’’ or ‘‘WBAN.’’ A focus on ML applications, relevance to WBANs, and publication in peer-reviewed journals were among the inclusion and exclusion criteria used to filter studies during the screening phase. We excluded studies that didn’t fit these requirements, such as not being in English or being shorter than four pages. During the eligibility process, we used established evaluation measures to ensure the validity and reliability of the studies we selected. Data from qualifying studies were finally combined during the inclusion phase to determine trends, obstacles, and potential avenues for future research regarding the application of ML techniques to WBANs. The PRISMA framework is demonstrated in the paper ‘‘Energy Efficient and Reliable Routing in Wireless Body Area Networks Based on Reinforcement Learning and Fuzzy Logic’’ by Guo et al. in [109]. Because of its substantial contribution to improving WBANs’ energy 194744 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 9. PRISMA Framework [110]. TABLE 3. List of research questions. TABLE 4. Search strings used in the selected databases. efficiency and dependability, this work was found during the initial search, validated for relevance, evaluated for quality, and finally included in the review. The PRISMA framework, which ensures a rigorous, methodical, and objective review process, is used in this study following best practices for systematic literature reviews. All four of the SLR workflow’s primary steps are depicted in Figure 9. A. IDENTIFICATION 1) FORMULATION OF RESEARCH QUESTIONS (RQS) In developing the research topic, two sources were used: first, concepts from prior studies [103],[111]. All the publications focused on how ML algorithms can be best used within WBANs to address difficulties, improve performance, and lead to future improvements in healthcare applications. Second, the mnemonic PICo, which stands for ‘P’ (population or problem), ‘I’ (interest), and ‘Co’ (context), based on these notions, contained three noteworthy features as part of the review [112]. The population comprises academics and research communities focused on WBANs and ML. The aim is to completely review and understand the use of ML algorithms in the context of WBANs and investigate various ML techniques and their impact on WBAN performance in healthcare. The interest lies in comprehensively reviewing and understanding the utilization of ML algorithms in the context of WBANs and exploring various ML techniques and their impact on the performance of WBANs in healthcare. The setting is WBAN-related, with a focus on ML integration and applications. This enabled the authors to formulate the three research questions of this study as listed in Table 3. 2) KEYWORDS FORMULATION Six primary keywords were found based on the developed study questions: ML techniques, AI, DL, RL, and the WBAN approach. To supplement these keywords, the author used an internet thesaurus such as thesaurus.com, consulted previous studies’ keywords, consulted Scopus’ keywords, and sought expert advice. Several terms were searched using this process, including ML techniques, AI, DL, RL, WBAN, and wearable sensors. VOLUME 13, 2025 194745 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques 3) SEARCHING CRITERIA USING DATABASES Using field code functions, phrase searches, and Boolean operators, the keyword combinations were looked up in three databases: Web of Science, Scopus, and IEEE Xplore, as shown in Table 4. 4) RECORD EXTRACTION Whittemore et al. [113] state that using qualitative or mixed method approaches that enable the researcher to compare primary data sources iteratively is the best way to synthesize or analyse integrative data. In this study, an integrative review was used. Several papers intended for quantitative, qualitative, and combination reviews were included thanks to this process. The researcher carefully examined the abstract, results, and discussion portions of the articles. The data abstraction was based on the research questions, meaning that all information from the examined studies that could help answer the research questions was extracted and included in Table 5. Subsequently, the researcher employed a thematic analysis to discern themes and sub-themes by examining the abstracted data for patterns and topics, clustering, numbering, and identifying parallels and correlations [114]. For synthesizing a mixed study design (integrative), thematic analysis is thought to be the most suitable method [115]. It is characterized as a descriptive approach that connects with other data analysis approaches and flexibly analyzes data [116]. Creating themes is the first stage in a thematic analysis. We looked for trends in the abstracted data from every reviewed paper throughout the process. Three major categories were formed by pooling relevant or comparable abstracted data into a group. After that, the authors re-examined the three data groups and discovered the remaining thirty-two subgroups. A review of these themes’ correctness was the next step. The writers re-examined each major theme and sub-theme produced during this process to guarantee their applicability and correct depictions of the data. The writers then moved on to the following phase: identifying the topics for every group and its subgroup. Before naming the topics for the subgroup, the authors first identified the themes for the leading group. This method was used to establish themes in a group of co-authors and corresponding authors who shared the findings’ theme. Until the point of agreement on modifying the produced themes and sub-themes, the researcher talked about any contradictions, concepts, conundrums, or thoughts that might be connected to the interpretation of the data. Two panels of experts, each with backgrounds in community development studies and qualitative methodology, were shown the produced themes and sub-themes. The experts were asked to assess 32 sub-themes such as: Performance evaluation of the model, EnergyEfficient, Reliable, and Low-Latency, Reduction in Energy Consumption and Latency, Computational Time, Precision, Overall Accuracy, Sensitivity and Specificity, Throughput, Model Performance Improvement, Hyperparameter Tuning, Model Outperformance, Addressing Challenges, Reduction in Transmission Cost, Efficiency and Reliability Enhancement, Improved System Energy Efficiency, Increased Utility of WBAN, Increased Effectiveness of the Sensor, Significant Increase in System Lifetime, Power Optimization and Interference Reduction. Favourable anti-interference performance and frequency resource utilization, Duty Cycle Optimization, Decreased data traffic, Successful Data Delivery, Reduced Attack Possibility of Jammer, Improved Block Error Rate, QoS Satisfaction, Contributions to Healthcare Decision-Making and Outcomes, Adaptive Performance in a Fast-changing WBAN Topology, High Prediction Rate for Heart Diseases and Model Efficacy in WBAN, Development of an ML-Based mHealth System, Effectiveness for Image Encryption/Compression, Contribution to Reducing Misdiagnosis and False Positives, and 3 main themes such as (1) the most common ML models utilized in WBAN research, (2) performance evaluation metrics and the insights they offer, and (3) future research directions and challenges. subjectively. Both concurred that the themes and sub-themes were appropriate and pertinent to the review’s findings. B. SCREENING The second process was screening, in which articles were either added to or removed from the study depending on a predetermined set of criteria (either manually by the author or with the help of the database). This review restricted the screening method to include publications published between 2017 and 2025 to adhere to the concept of research field maturity, which is highlighted by [117]. This chronology was selected because there was enough published research to conduct a representative review. Because empirical research papers include primary data, the writers chose to review them. To prevent misunderstandings, only English-language texts were considered. As a result, 2,407 articles were left for review in the next round. Table 5depicts the details. C. ELIGIBILITY AND INCLUDED This section combined the two phases together, making it an eligibility and inclusion phase. In this section, a total, 633 documents were obtained from the Web of Science, 1,123 from Scopus, and 651 from the IEEE Xplore. Data retrieved from search engines included authors, names, titles, digital object identifiers (DOIs), abstracts, and keywords. The three lists were then integrated into Mendeley, removing redundant articles. Duplicate papers were found and eliminated using the Mendeley reference manager to preserve the integrity of the records gathered during the SLR. The data was exported in research information system (RIS) and comma-separated values (CSV) formats from academic databases, including Web of Science, Scopus, and IEEE Xplore, and then imported into Mendeley for additional processing. Mendeley’s built-in Check for duplicates function was used to compare metadata fields, such as titles, authors, publication years, and DOIs, to find the duplicated papers systematically. 194746 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 5. Inclusion and exclusion criteria. The deduplication procedure began with a review of the duplicates that were found. Mendeley automatically grouped articles with metadata that matched or almost matched into sets for verification. Articles with similar names or identical DOIs were thoroughly inspected to confirm duplication. Based on title similarities, Mendeley’s matching system identified potential duplication for entries lacking DOIs. This was brought on by metadata issues that led to slight variations in titles or author order, which were fixed by manual inspection. After this procedure, the records were reduced from 2,407 to 1,744 unique entries after 663 duplicate records were detected and removed. Unique articles were kept for review once the cleaned dataset was exported for further analysis. By avoiding the inclusion of duplicate articles in the literature review, our Mendeley approach expedited the deduplication process without sacrificing accuracy. Figure 10 depicts the Mendeley deduplication procedure. However, a total of 1,501 out-of-scope records were removed from the 1,744 records obtained in Mendeley, and all abstracts were reviewed to exclude extraneous items from the study’s objectives. Nevertheless, the details of the data extraction have been identified. In this stage, the data extracted for inclusion and exclusion criteria is 243 articles. Out of the number of selected articles, 81 were excluded based on title, 162 were selected based on their title and relevance to the study, and 14 reports were not retrieved based on the abstract. Furthermore, 148 Reports were assessed for eligibility, and 52 articles were excluded because they did not have an ML application. 21 articles were removed for not being original contributions, 9 for having fewer than four pages, 10 for not being published in peer review, 3 for not being published in English, and 3 for being review studies. Finally, after all the assessments based on the exclusion and inclusion criteria, 55 papers were selected for the study [108]. Figure 11 presents the PRISMA flowchart designed for this study. D. BACKGROUND OF THE SELECTED STUDIES Moreover, concerning years of publication, 2 articles were published in 2017–2018, 9 articles were published in 2019–2020, 20 articles were published in 2021–2022, and 24 articles were published in 2023–2025. Figure 12 shows the frequency of articles published between 2017 and 2025. This indicates a notable increase in the use of ML techniques in WBANs. The most recent publication years are 2023–2025, with 24 articles. E. SUMMARY OF THE SECTION An extensive review of ML methods used in WBANs is discussed in this section. From traditional supervised and unsupervised techniques to advanced DL models, it thoroughly examines a variety of ML algorithms, emphasizing their applications in activity recognition, anomaly detection, communication optimization, routing protocols, QoS, personalized healthcare, and security improvements. The popularity and effectiveness of ML models in tackling WBAN issues, including energy efficiency, real-time monitoring, and data reliability, are demonstrated by examining 55 chosen publications. RL maximizes energy usage and adaptability, while DL models, such as CNNs and LSTMs, perform better when handling complicated data. Nevertheless, there are still interpretability, scalability, and processing cost issues. The taxonomy of ML approaches in WBAN highlights the significance of choosing the right models for certain applications and the need for more study to get beyond current obstacles and increase their usefulness in healthcare systems. V. BASIC AND ADVANCED ML-BASED TECHNIQUES FOR WBANs The importance of identifying frequently used ML algorithms and understanding their functions in handling the complex and dynamic data generated by WBANs was underlined by this section, in which a variety of approaches have been studied, including more complex DL models as part of supervised learning strategies [12],[111],[118],[119],[120],[121], [122], to the more sophisticated DL models [63],[120], [121],[123],[124],[125],[126],[127],[128],[129],[130], [131],[132],[133],[134],[135],[136] and conventional unsupervised techniques [137],[138]. The advantages of each kind of algorithm increase the precision, effectiveness, and adaptability of WBAN applications. Because they are simple to use and comprehend, basic ML techniques like DT, SVMs, and KNN are well-liked in resource-constrained environments. Conversely, sophisticated machine learning methods like CNNs, RNNs, LSTM networks, and RL excel at handling high-dimensional data, identifying intricate features, and streamlining dynamic processes like slot allocation and energy management. The 55 selected studies demonstrate that a wide variety of ML models are being applied and are effective in WBANs [12],[63],[109],[111],[118],[119],[120], [121],[122],[123],[124],[125],[126],[127],[128],[129], VOLUME 13, 2025 194747 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 10. Mendeley deduplication process. [130],[131],[132],[133],[134],[135],[136],[137],[138], [139],[140],[141],[143],[145],[146],[147],[148],[149], [150],[151],[152],[153],[154],[155],[156],[159],[163], [164],[166],[167],[173],[174],[175],[176],[181],[182], [183],[184]. Most studies clearly identify the algorithms they used. These findings demonstrate how crucial ML is to encourage creativity and improve WBAN performance. A. ML-BASED TECHNIQUES FOR WBANs This section provides a thorough examination of the ML-based methods used in WBANs. To address critical issues in WBANs, including activity recognition, anomaly detection, communication optimization, routing protocols, QoS, personalized healthcare, and security enhancements, it systematically categorizes and evaluates a variety of ML techniques. This review lines up a wide range of ML approaches, from staple supervised and unsupervised methods to newer DL, RL models, and shows how they stack up in everyday WBAN use. For each group, we include a table that spells out how the algorithms learn, how complex they are, how accurate they can be, and what they do well or poorly. We make it clear that the ‘‘best’’ model depends on what the WBAN needs: can it scale, save battery power, and keep up with real-time demands? We also call out the headaches that come with using ML in these networks: heavy computation, lopsided datasets, hard-to-explain decisions, and ethical concerns. We finish by flagging the open questions and pointing to where future work could make ML an even better fit for healthcare and other WBAN applications. In short, the review offers a straightforward roadmap for anyone looking to see how machine learning is pushing WBAN technology forward. The 55 chosen studies show that an enormous range of ML models are being used and are successful in WBAN applications. Figure 13 is the taxonomy of the ML-based techniques used in WBANs regarding this SLR. 1) ML-BASED ACTIVITY RECOGNITION TECHNIQUES FOR WBANs The study compares different ML methods for human activity recognition (HAR) and related applications in healthcare WBANs. Key findings and developments in the field are highlighted as the analysis examines their learning paradigms, model complexity, performance, and limitations. Because they can handle complex data and extract high-dimensional characteristics, DL techniques like CNNs and LSTM networks are widely used. On the other hand, supervised learning, which provides intermediate complexity but typically falls short of the accuracy attained by DL models, is the foundation of conventional techniques like SVMs. DL models continuously perform well in terms of accuracy and model complexity. For instance, El-Adawi et al, and Boga et al. [133] and [139] both use CNNs to attain remarkable accuracy, highlighting the potency of these architectures in feature extraction. In a similar vein, 194748 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 11. PRISMA flow diagram. FIGURE 12. Publication years of selected studies. Kedjar et al. [135] combine CNN and LSTM to achieve prominent position identification and classification accuracy in both line-of-sight (LoS) and non-line-of-sight (NLoS) situations, but in contrast to DL models [119]. The SVM-based VOLUME 13, 2025 194749 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 13. Taxonomy of the ML-based techniques for WBANs. method offers intermediate complexity and lower accuracy, but it still shows gains in metrics like loss, delay, and throughput. The models have significant limitations despite their excellent accuracy. El-Adawi et al. [139] point out that interpretability and sensor-specific difficulties are not considered, although they are essential considerations in delicate healthcare sectors. The performance in the study by Kedjar et al. [135] is hampered by data imbalance, especially regarding NLoS location prediction. Treatment delays result from older methods like SVMs’ incapacity to accurately and promptly identify patient problems, as noted by Kathuria et al. in [119]. In similar research by Boga et al. in [133] highlights difficulties in modelling active learning paradigms and evaluating decision-making when wearable sensors are included. Notwithstanding these drawbacks, the models make noteworthy contributions. The model by El-adawi et al. in [139] demonstrate its effectiveness by achieving accuracy, Fmeasure, and Matthews Correlation Coefficient (MCC). The study by Kedjar et al. in [135] demonstrate its reliability in location identification tasks by exhibiting good classification accuracy in LoS-NLoS scenarios and a decreased root mean square error (RMSE) in channel prediction. Despite its reduced accuracy, Kathuria et al. [119] shows minor benefits in loss, delay, and throughput measures. Lastly, the study by Boga et al. in [133] emphasizes how well feature selection works to enhance HAR recognition performance. Table 6 depicts details of ML-based activity recognition techniques for WBANs. 2) ML-BASED ANOMALY DETECTION TECHNIQUES FOR WBANs A comparison of ML approaches used for goals in healthcare WBANs is presented in this section. It emphasizes the ML techniques, their intricacy, learning models, precision, constraints, and salient features. The emphasis is on tackling issues like harmful data patterns in WBAN systems and anomaly detection in physiological parameters. The goal of Boga et al. [133] was creating a technique for identifying abnormalities in physiological data captured by WBAN sensors. They used supervised learning with an ANN. The model’s durability in anomaly detection is demonstrated by its medium complexity and good accuracy, precision, recall, and F1 score. Nevertheless, the study points out drawbacks in assessing decision-making in HAR tasks, especially when combining wearable sensors with active learning paradigms. Nonetheless, the technique works well to provide dependable performance for detecting anomalies in WBANs. Finding dangerous data patterns brought on by sensor malfunctions, inaccurate readings, and possible hostile activity was the main goal of [129]. They employed a DL strategy that delivers high complexity and performance by combining 194750 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 6. ML-Based activity recognition techniques for WBANs. CNN and LSTM networks. The model’s effectiveness in managing various data is demonstrated by its high accuracy measure across various subjects in the datasets. However, the study highlights new inconsistencies in WBANs and difficulties with huge data streams. More complex solutions that can handle dynamic and expansive WBAN environments are required, as current statistical and ML techniques frequently fall short in identifying such patterns. Table 7shows the details of ML-based anomaly detection techniques for WBANs. 3) ML-BASED COMMUNICATION TECHNIQUES FOR WBANs This section provides an academic comparison of ML approaches used in WBANs, along with information on their advantages and disadvantages, and important performance indicators such as throughput, energy efficiency, latency, reliability, and packet delay ratio (PDR). Fernandes et al. [124] reduced data packet loss by 10% by using an ANN in supervised learning to increase communication reliability in WBANs. However, this method led to higher latency because of the dependability of the best-attained action. CNN was used by Liu et al. in [126] for DL-based performance gains in energy efficiency and reliability; however, the result’s generalizability is limited by the absence of comprehensive real-world testing. Chen et al. [140] used Q-learning intending to reduce energy use; however, the evaluation was based only on simulations, which might not accurately reflect realworld circumstances. Although the energy needed to run DRL models may offset some of the advantages in practice, Gupta et al. [141] found that DRL models improve energy efficiency in real-time applications. While maximum energy consumption was still a problem, Ahmad et al. [127] discovered that using CNN in routing protocols improved packet transmission and decisionmaking, decreasing path loss ratios and increasing energy efficiency. High-performance transceivers for WBANs were designed by Ali et al. in [142] using DL; nevertheless, their strategy might be constrained by data rates of up to 1.312 Mbps, which would not be suitable for high-throughput applications. Although other crucial aspects like latency and reliability were not thoroughly investigated, He et al. [143] developed a deep Q-learning-based power controller in RL that outperformed baseline controllers in terms of energy efficiency. While a study by Roy et al. in [144] emphasized ML’s revolutionary potential in automating healthcare procedures, their findings overlooked crucial factors, including ethical considerations and legal constraints that could hamper ML adoption in healthcare. Similarly, He et al. in [143] pointed out that although their method effectively increased network lifetime and reduced packet loss rates, scaling up networks may be significantly hampered by the computing cost and time required for RL. Although other important aspects like latency and reliability were not thoroughly examined, RL by He et al. in [143] proposed a deep Q-learning-based power controller that outperformed baseline controllers in terms of energy efficiency. It highlighted the revolutionary potential of ML in automating healthcare procedures but failed to consider key factors such as moral dilemmas and regulatory restrictions that could prevent ML from being widely adopted in the medical field. Roy et al. [144] emphasized ML’s transformative role in automating healthcare processes. However, their findings overlooked critical issues like ethical concerns and legal regulations, which might hinder ML adoption in healthcare. Similarly, a study by He et al. in [143] noted that while their model successfully extended network longevity and reduced packet loss rates, the computational cost and time required for RL could pose a significant barrier when scaling up networks. VOLUME 13, 2025 194751 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 7. ML-based anomaly detection techniques for WBANs. DRL was used for intelligent reflecting surface (IRS) aided methods by Xiao et al. [145] who pointed out that even while there are advantages in terms of energy efficiency and lower eavesdropping rates, the computing demands of the learning process may result in significant energy overhead. Although important aspects like security and scalability were not fully addressed, Rao et al. [146] demonstrated how RL could dynamically modify channel allocation in WBANs, boosting dependability in medical data transfer. The QL-based approach by Mohammadi et al. in [147] increased network connection, energy efficiency, and has drawbacks such as sleep/wake problems and emergency packet loss. In another research by Kim et al. in [148] DRL enhanced WBAN performance, emphasizing increasing throughput while lowering energy usage. They did point out that it is still unclear how well the suggested DRL strategy will work with larger-scale WBAN installations. The complexity of scheduling and power control may rise with the number of devices in the network, which could impact the system’s overall performance. To evaluate the suggested approaches’ robustness and practical applicability, the authors underlined the necessity of additional validation through testing in various environments and real-world implementations. To ensure QoS in healthcare systems, Abdu et al. [12] used ANN through supervised learning to create energy-conscious routing for WBANs. Their study did note one drawback, though: the frequent requirement to replace sensors because of resource constraints. This could pose a fundamental problem for sustaining system performance over the long run. Despite this, their method showed significant energy management and routing efficiency gains, making it a viable option for Internet of Things-based medical applications. DNNs were used by Cwalina et al. in [128] to increase communication efficiency by addressing the problems of LoS and NLoS categorization. Nonetheless, when working with learning datasets, the instability of the SVM and Threshold Method (THM) limited the investigation, which would have affected the classification’s resilience and dependability. Despite these obstacles, their DL method successfully increased the classification effectiveness in WBANs, indicating its potential to boost system performance. Mehrani et al. [131] reduced data communication and increased energy efficiency in WBANs by combining fuzzy logic with LSTM networks. They acknowledged a drawback in not investigating other potentially effective ML techniques that may have further optimized the performance, even though their method produced a stunning reduction of communicated data by 81% and energy expenditure by 73%. Even though their strategy effectively lowered energy usage, including more algorithms to handle other aspects of WBAN optimization would be advantageous. Regarding FL, Consul et al. [149] noted that while federated RL enhanced latency, energy consumption, and throughput in WBANs, it may also result in additional computing overhead. Although they overlooked the energy cost of edge servers handling offloaded tasks, it is also worth noting that FL may increase communication and energy consumption because the sensor nodes need to periodically exchange model updates or gradients with a central server. However, most of the reviewed studies did not provide key information, such as the size of the shared model, the degree of compression used, or how often updates are transmitted, making it difficult to estimate this overhead or determine its impact on battery life under IEEE 802.15.6 power constraints. Future work should report these parameters more clearly so that the performance and practicality of FL-based WBANs can be evaluated more accurately. Yin et al. [150] demonstrated how DQN in RL may optimize the QoS in WBANs. Table 8provides an ML-based communication approach for WBANs. 4) ML-BASED ROUTING PROTOCOL TECHNIQUES FOR WBANs With an emphasis on their use cases, QoS enhancements, drawbacks, and significant contributions, this part studies ML approaches used in routing protocols for healthcare WBANs. 194752 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 8. ML-based communication techniques for WBANs. VOLUME 13, 2025 194753 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques In practice, these metrics, throughput, latency, energy use, error rates, and more, guide both researchers and engineers. They reveal each design’s strengths and weak spots and serve as yardsticks for improving future WBAN hardware and protocols. A well-rounded set of benchmarks, spread across multiple sub-themes (scalability, reliability, energy efficiency, data quality, and security), lets us judge whether a system is truly ready for clinical use. By viewing WBAN performance through this multi-lens approach, we can keep pushing the technology forward and deliver more dependable, patient-friendly health-monitoring solutions. 1) PERFORMANCE EVALUATION OF THE MODEL This part assesses the performance of models used in WBANs, including predicted accuracy capabilities. Evaluating this model’s performance is critical for making trustworthy and effective decisions in healthcare applications. However, when evaluating the algorithm’s overall performance in identifying DDoS attacks in cloud-assisted WBANs, the examined metrics of classification accuracy, tree size, time, and memory are essential [122]. In addition, the study, which integrates priority-based adaptive scheduling with DRL for power control, demonstrates superior performance in terms of PDR and throughput [159]. Compared to traditional methods, it also decreases latency and lowers power usage. Another research by Li et al. in [160] claims that the DRL optimization (DRLOPT) method efficiently adjusts to environmental changes, continuously achieving excellent performance with minimal computational cost. a: QUANTITATIVE PERFORMANCE COMPARISONS OF THE MODEL ON WBAN In this SLR, we investigate the use of sophisticated ML techniques in WBANs, focusing on important performance measures such as accuracy, latency, and power consumption. These factors are essential for determining how well ML models work in real-life health monitoring situations that need quick responses and smart use of energy. DL models such as CNNs and LSTM networks are notable for their high accuracy, with a 97% success rate in predicting user behaviour. Not only are these models accurate, but they also minimize data transfer by 81%, hence lowering latency. Furthermore, they increase energy efficiency by 15% as compared to previous scheduling approaches utilized in WBANs [126],[131],[132]. SVMs, on the other hand, produce reliable findings at a 92.1% accuracy level. While they maintain minimal latency (typically less than 20 milliseconds), they are less capable of handling more complex input than DL models. Nonetheless, SVMs are very power-efficient, making them an excellent choice for battery-powered devices. They consume less energy than deep learning models, which usually demand greater processing power [161]. RF and DT also perform well, particularly with high-dimensional sensor data, with an accuracy of 88.6%. These models strike a fair mix between accuracy and the necessity for real-time performance, with low latency (around 20 milliseconds). They also use 30% less power than deep learning models, making them appropriate for wearable devices that must conserve energy [120],[162]. Table 14 summarizes the performance comparisons of various ML techniques used in WBANs, focusing on accuracy, latency, and power consumption. b: ENERGY-EFFICIENCY, RELIABILITY, AND LATENCY Due to power limits and real-time data transmission needs, WBANs require energy-efficient, reliable, and low-latency wearable devices. Metrics in this sub-theme evaluate the energy consumption, reliability, and latency of data transmission protocols and algorithms to optimize resource utilization and improve system performance. Nevertheless, Table 15 indicates the details of this sub-theme. c: REDUCTION IN ENERGY CONSUMPTION AND LATENCY This section focuses on lowering latency and energy consumption in WBANs. WBANs can increase responsiveness and extend device battery life by reducing energy consumption and communication delays, ultimately improving user experience and system efficiency. The outcomes of the sub-themes are shown in Table 16. d: COMPUTATIONAL TIME The time needed for data processing, analysis, and decision-making in WBANs is measured by computational time. Since quick responses are essential for patient monitoring and intervention, real-time applications require a minimum processing time. The suggested unsupervised coloring algorithm (UCA) demonstrated convergence outcomes with both correct and inaccurate positioning in the study by Ma et al. [137]. It performed better than current techniques in terms of robustness against topological changes, faster algorithm convergence, reduced interference strength, and less time complexity. Likewise, in another study by Li et al. [160] the DRL optimization (DRL-OPT) algorithm effectively follows environmental variations and maintains good performance with low computational complexity. e: PRECISION, ACCURACY, SENSITIVITY, AND SPECIFICITY This combines 4 relevant sub-themes: precision, overall accuracy, sensitivity, and specificity. The result is shown in Table 17. These metrics evaluate the precision, accuracy, sensitivity, and specificity of diagnostic and predictive models in WBANs. Reliable illness identification, treatment planning, and healthcare decision-making all depend on elevated levels of precision and accuracy. f: THROUGHPUT The rate of successful data transmission in WBANs is quantified by throughput measurements, which show how well the network can handle data traffic. The throughput for a WBAN 194760 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 14. Summarizing the performance comparisons of various ML techniques used in WBANs, focusing on accuracy, latency, and power consumption. TABLE 15. Summary of the energy-efficient, reliable, and low-latency sub-theme. TABLE 16. Summary of Reduction in Energy Consumption and Latency Sub-theme. TABLE 17. Summary of precision, accuracy, sensitivity, and specificity sub-theme. network is denoted as TH [165] as follows: TH =RchennelTdata Tsimulation (1) where Tchennal Is the licensed channel’s data transmission rate Tdata Is the total transmission time for all successful transmissions, and the Tsimulation Is the system simulation time? Throughput optimization is critical for ensuring that sensor data is delivered to healthcare decision-support systems on time. Table 18 compares the throughput between some models and the traditional method. TABLE 18. Comparison in terms of throughput. g: MODEL PERFORMANCE IMPROVEMENT This section evaluated the performance improvements made by model optimization and refinement strategies.Researchers have increased the prediction models’ efficacy and accuracy in WBANs by iteratively improving model performance. Thus, the author [127] proposed an improved quality of routing protocol (IM-QRP) that shows notable improvements, including a 10% increase in residual energy, a 30% reduction in path loss ratio, a 10% improvement in packet transmission reliability, and a 7% improvement in SNR, in contrast to the VOLUME 13, 2025 194761 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques current cooperative link energy-efficient balanced algorithm (CO-LEEBA) and QoS path routing discovery (QPRD) routing protocols. With 98.8% accuracy, 99.9% recall, 97.9% F-score, and 99.8% precision, the proposed hybrid artificial NN and grasshopper optimization algorithm (ANN-GOA) for anomaly identification outperforms other conventional techniques in a distinct study [118]. Furthermore, the experimental findings showed a remarkably low false alarm rate and an accuracy of over 96% for problem detection. The overall efficacy of the technique is demonstrated by the fault management prototype, which uses the suggested framework to classify faults, automate sensing node profiling, and make it easier to train and validate new models [120]. h: HYPERPARAMETER TUNING Measures of hyperparameter tuning assess how well model hyperparameters can be optimized for improved performance. Within WBAN applications, adjusting model parameters can lead to notable improvements in anticipated accuracy and resilience. However, a high prediction rate for cardiac disorders is achieved by the RNN flexible architecture and parameter tuning, which are made possible by the tunicate swarm-sail fish optimization (TS-SFO). The suggested model successfully raises prediction accuracy in the WBAN setting [125]. i: MODEL OUTPERFORMANCE Measures of model outperformance to identify the most effective methods for certain tasks or applications. This sub-theme evaluates the performance of multiple models or algorithms within WBANs. Researchers can find the most effective approaches and solutions for WBAN adoption by contrasting model performance with industry norms or rival approaches. Nevertheless, Table 19 summarizes the results under the sub-theme model outperformance. j: SUMMARY OF THE SECTION WBANs are transforming thanks to ML, which has improved their functionality and addressed significant issues. From communication within WBANs to interactions between WBANs and beyond, this part examined the integration of ML techniques across several WBAN layers. It has been demonstrated that advanced ML models, including DL, RL, Gen AI, and FL, enhance real-time processing, energy efficiency, data quality, and security in WBANs. Robust anomaly detection, individualized healthcare monitoring, and effective resource management are made possible by these methods. However, privacy protection, scalability, and model interpretability remain crucial topics for further study. This section demonstrates how ML can transform WBANs and open the door for scalable, secure, and patient-centred healthcare systems by analysing existing research and pointing out unresolved issues. In addition to addressing technical constraints, ML integration into WBANs meets the growing need for intelligent, flexible healthcare solutions. VI. DISCUSSION AND STATISTICS The study found that supervised learning was the most used ML model across the analysed studies, with 41% of the total methods used. Supervised ML techniques, such as, ANN, SVM, DT, SVR, KNN, LRE, RF show how crucial labelled data and supervised learning algorithms are in developing prediction models for classification and regression tasks in WBANs like effective in classifying faults, automating sensing node profiling, detecting hidden patterns in healthrelated data, and convergence time, and accuracy for disease prediction due to their exceptional performance in various applications [111],[118],[119],[120],[121],[122],[156], [167]. DL techniques are part of this learning technique, notable for their ability to automatically learn and improve data representations [111],[118],[119],[120],[121],[122], [156],[167]. Convolutional networks (CNN, DCNN, R-CNN) and recurrent architectures (RNN, LSTM, ConvLSTM, and BiLSTM) are two examples of DL models that have significantly impacted WBAN applications. DL has been shown to improve energy efficiency, extend network lifetime, lower transmission power, and lessen interference at the system level [113],[114],[115]. It also improves communication reliability by allowing for fault classification, automatic node profiling, and efficient picture compression and encryption models [117],[118],[119],[120],[121]. By reducing false positives in disease detection, uncovering hidden patterns in physiological data, and achieving high prediction rates in applications like heart disease classification and early-stage lung cancer diagnosis models, arrhythmia detection, stresslevel prediction, and fall detection, DL supports accurate health monitoring at the clinical level [12],[122],[123], [124],[125],[126],[127]. These results highlight the dual role of DL in enhancing healthcare outcomes and network performance in WBAN systems. DL models as a subset of supervised category make the proportion of supervised learning approaches become higher, confirming that supervised techniques remain dominant in WBAN-related ML applications. RL, which accounted for 35% of the data, was the second most common ML model behind supervised learning. RL techniques have demonstrated encouraging outcomes in optimizing resource allocation, routing protocols, and adaptive control mechanisms in WBANs [168],[169],[170], [171]. These methods enable agents to experiment with their surroundings and discover the best decision-making policies. This technique includes DRL, which is a hybrid technique that uses DL and RL concepts. This combination of methods is perfect for applications like adaptive resource allocation and individualized healthcare monitoring in WBANs because it can resolve complex optimization problems and adjust to shifting environmental variables [168],[172]. In addition to increasing overall usefulness to WBAN security, improving 194762 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 19. Summary of the results of the sub-theme model outperformance. WBANs’ effectiveness and dependability for medical applications, and recording a variety of network interactions [141], [145],[148],[155],[164],[173],[174],[175],[176]. The study demonstrated that, in addition to these DRL methods, specialized algorithms such as the FL and FRL are also a part of an RL framework that is used to provide collaborative ML that reduces time delay, enhances throughput, lowers energy consumption, and protects privacy when combining ML and WBAN data [149]. Unsupervised ML approaches, which have shown promise in anomaly detection and clustering analysis in WBANs, were applied in 24% of the records. Without the need for labelled instances, these methods make it possible to find underlying patterns and structures in data, and they increase the network lifespan, enhance adaptability to changing network topologies, reduce interference strength, and accelerate algorithm convergence [137],[138],[152]. Figure 14 depicts the details of the identified ML in WBANs from 2017–2025. Figure 15 shows the issues in WBANs that ML addressed between 2017 and 2025. The most important problem that ML addresses in WBANs is communication, which receives a significant amount of attention (40%) and is crucial to dependable data transfer and system performance. Personalized healthcare receives significant attention (25%) as evidence of the expanding trend toward patient-centered solutions. Another crucial area is security (11%), which emphasizes how crucial it is to protect private health information and preserve system integrity. Focusing on moderate areas, like Activity Recognition (8%) and QoS (8%), enhances the efficiency and adaptability of the system. In contrast, more specialized areas such as Routing Protocols (4%) and Anomaly Detection (4%) receive less attention in WBANs research, despite their significance. Body mobility, limited energy resources, and the need for ultra-low latency constrain routing in WBANs, making the direct application of ML techniques challenging. Similarly, anomaly detection requires large, diverse, and labelled datasets to effectively identify abnormal physiological patterns; however, these datasets are often difficult to obtain due to privacy FIGURE 14. Distribution of machine learning approaches applied in WBAN research (2017–2025), showing that Supervised Learning (41%) dominates, followed by Reinforcement Learning (35%) and Unsupervised Learning (24%). concerns and limited access to patient data. Despite their real-world importance, these challenges explain the reduced focus of machine learning in these fields. We propose oneclass/self-supervised and federated methods, enhanced with privacy-preserving data synthesis and uncertainty estimation, for anomaly detection, as well as budgeted/safe or model-based RL and lightweight, graph-aware policies for routing to address these issues. To make results clinically relevant, we also recommend evaluating cross-subject generalization and reporting deployment-focused metrics (energy per decision, latency, and false alarms). This distribution demonstrates a strategy that balances attention to other complementary areas while giving priority to communication, personalized healthcare, and security. The way we monitor health may change if ML techniques are incorporated into WBANs. Tiny wearable sensors, which feed real-time data to machine learning algorithms, enable physicians to identify early symptoms of disease, personalize treatment plans, and monitor daily activities. It is difficult to translate this vision into a working system, according VOLUME 13, 2025 194763 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 15. Issues in WBANs addressed by ML (2017–2025). Communication (40%) is the largest focus, followed by Personalized Healthcare (25%) and Security (11%). Activity Recognition (8%) and Quality of Service (QoS, 8%) receive moderate attention, while Routing Protocols (4%) and Anomaly Detection (4%) are comparatively underexplored. Overall, the literature prioritizes reliable connectivity and patient-centric outcomes, with a meaningful—but smaller—emphasis on security and specialized tasks.’’ to several studies. For example, well-known problems with WBAN deployments include sensor unreliability and data noise brought on by electrode displacement, battery drain, or environmental interference [161],[177]. Similarly, energy limitations continue to be crucial: the majority of WBAN devices run on tiny batteries, and sophisticated machine learning models use more power, which reduces their long-term usefulness [178]. Clinicians are frequently hesitant to trust complex model predictions without clear, transparent explanations, so the lack of model interpretability has emerged as a major obstacle to clinical adoption, surpassing hardware limitations [171]. Additionally, since performance is reported using a variety of metrics (accuracy, latency, energy, sensitivity, etc.), the absence of standardized evaluation frameworks makes it more difficult to compare studies, as noted by [136] and [179]. Finally, the existing WBAN literature does not adequately address the issues of scalability and cross-population generalization [180]. It will take developments in explainable ML layers, standardized testbeds, ultra-efficient algorithms, and data preprocessing to overcome these obstacles. In WBAN settings, robust security and privacy features like FL and lightweight encryption are also essential for safeguarding private health information [157],[158]. For the creation of scalable, interpretable, safe, and energy-efficient ML models that are adapted to WBAN constraints, multidisciplinary cooperation between engineers, computer scientists, and healthcare professionals is therefore crucial. A. DEVELOPED THEMES Three primary themes emerged from the thematic analysis: (1) the most common machine learning models used in WBAN research, (2) performance evaluation metrics and the insights they provide, and (3) upcoming research directions and challenges, along with their subthemes such as accuracy, efficiency, and computational effectiveness. These were among the key performance metrics employed to evaluate WBAN models in the first theme, Performance Evaluation. Model performance assessment, energy-efficient and latency-reducing designs, reduction in energy consumption, computation time, accuracy, sensitivity, specificity, throughput, performance improvements, hyperparameter tuning, and model outperforming others, along with reducing transmission costs, enhancing sensor functionality, increasing the usefulness of WBANs, and making the 55 selected studies demonstrate that a wide variety of ML models are being applied in WBANs are effective and efficient. Other critical areas included system lifespan, power efficiency, interference mitigation, frequency utilization, duty cycle optimization, traffic reduction, and reliable data delivery. Another significant issue was security, with sub-themes focused on methods to prevent jamming attacks, decrease block error rates, and ensure QoS. Additionally, there was progress in developing ML-based mHealth systems, improving image encryption and compression, reducing the risk of misdiagnosis and false positives, enabling adaptive performance in shifting WBAN topologies, and enhancing prediction accuracy for heart diseases. By methodically examining ML approaches, trends, and preferences, this theme provides us with information about the most effective ways to apply ML in various WBAN contexts. Additionally, it clarifies the distinction between simple and sophisticated processes. These three domains offer a thorough summary of the current state of WBANs, highlighting research challenges, performance requirements, and the groundbreaking potential of ML-based solutions. 1) THEME 1: MACHINE LEARNING MODELS USED The first theme examined how ML algorithms are transforming the current state of WBAN systems. The importance of identifying frequently used machine learning algorithms and understanding their functions in handling the complex and dynamic data generated by WBANs was underlined by this theme. A variety of approaches have been studied, including more complex DL models as part of supervised learning strategies [111],[118],[119],[120],[121],[122], to the more sophisticated DL models [12],[63],[120], [121],[123],[124],[125],[126],[127],[128],[129],[130], [131],[132],[133],[134],[135],[136] and conventional unsupervised techniques [137],[138]. The advantages of each kind of algorithm increase the precision, effectiveness, and adaptability of WBAN applications. Because they are simple to use and comprehend, basic ML techniques like DT, SVMs, and KNN are well-liked in resource-constrained 194764 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques environments. Conversely, sophisticated ML methods like CNN, RNN, LSTM networks, and RL excel at handling high-dimensional data, identifying intricate features, and streamlining dynamic processes like slot allocation and energy management. The chosen studies show that an enormous range of ML models are being used and are successful in WBANs. These results show how important ML is for fostering creativity and enhancing WBAN performance. In addition, this theme emphasizes how supervised learning is most utilized to maximize energy efficiency and improve WBAN communication reliability, accounting for 41% of the reviewed studies. Although less common, RL has potential applications in dynamic resource allocation and adaptive routing. A small but increasing percentage of studies use emerging techniques like FL, especially in healthcare settings where privacy is a concern. According to this distribution, there is growing experimentation with decentralized and adaptive models, even though traditional supervised methods continue to serve as the foundation. Furthermore, Table 20 provides an overview of the ML models found in WBAN research between 2017 and 2025, emphasizing the most common methodologies (e.g., supervised, unsupervised, and RL) and their main uses, including anomaly detection, communication optimization, and personalized healthcare. In addition to the use-case-specific discussions previously provided, this offers a comprehensive summary of the modelling techniques used. 2) THEME 2: PERFORMANCE EVALUATION Accuracy, efficiency, and computational effectiveness were among the key performance metrics used to evaluate WBAN models in Theme 2 ‘‘Performance Evaluation’’. Model performance evaluation, energy-efficient and latency-reducing design, energy consumption reduction, computation time, accuracy, sensitivity, specificity, throughput, performance improvement, hyperparameter adjustment, and model outperformance were some of the sub-themes. The standards by which WBAN systems and ML-based models are evaluated, enhanced, and validated are comprised of these sub-themes. However, the evaluation of performance in WBAN-ML studies is covered under this theme. The most used metric is energy efficiency, emphasizing the importance of battery life for wearable technology. While healthcare applications focus on accuracy, sensitivity, and specificity, communicationrelated studies prioritize throughput and latency. Additionally, some research examines model outperformance and hyperparameter tuning as key evaluation aspects. This case study highlights the diversity of performance issues in the field by showing that evaluation metrics are not uniformly applied but depend on the specific WBAN application. 3) THEME 3: CHALLENGES AND EMERGING PATHWAYS This theme examined clinical and technical issues that are impeding the advancement of WBAN applications. This category consists of 21 sub-themes, including reducing transmission costs, improving the functionality of sensors, increasing the usefulness of WBANs, and making systems more dependable and efficient. The system’s lifespan, power efficiency, interference reduction, frequency utilization, duty cycle optimization, traffic reduction, and reliable data delivery were other critical areas. Another major issue was security, with sub-themes concentrating on methods to prevent jamming attacks, reduce block error rates, and ensure QoS. The development of ML-based mHealth systems, improved image encryption and compression, reduced risk of misdiagnosis and false positives, adaptive performance in shifting WBAN topologies, and improved prediction accuracy for heart diseases were also included in this theme. Hence, the theme suggests that applying ML techniques to WBANs presents a complex environment with numerous opportunities and challenges. Identifying and resolving these issues and future research directions are crucial for maximizing the effectiveness of ML techniques in WBAN applications. Within this discourse, we delineate pivotal challenges and propose prospective trajectories for future research. VII. OPEN CHALLENGES AND FUTURE RESEARCH DIRECTIONS FROM THE INTEGRATION OF ML IN WBANs Although promising, incorporating ML into WBANs offers a distinct collection of difficulties and opens new research directions. These difficulties include handling the varied and dynamic nature of WBAN environments, coping with the diversity in human physiology, guaranteeing data privacy and security, and processing data in real-time with constrained computer resources [111]. ML models still struggle to perform reliably across different patient groups and clinical scenarios. A good first step is simply acknowledging these gaps and setting an agenda for deeper study. Promising directions include stronger data-encryption methods, ultralightweight learning algorithms, and adaptive models that improve each time new data arrives. Pointing out these shortfalls and laying out a roadmap for future work is critical if we want WBAN-based ML systems to deliver on their promise of better health monitoring and decision-making [79]. Researchers must tackle several practical hurdles: scarce training data, the need to merge records from many sources, and the day-to-day challenges of changing established clinical routines [185],[186]. Addressing these issues should raise the efficiency, accuracy, and usability of ML-driven monitoring tools [181]. If we overcome them, ML-enabled WBANs could transform healthcare by supporting continuous, personalized treatment that improves patient outcomes [187]. Integrating ML into WBANs also produces a fastmoving, complex landscape. Solving today’s problems and identifying tomorrow’s research questions will make these techniques far more useful in practice. One persistent issue is noisy, highly variable sensor data. Human movement, device faults, and changes in the surrounding environment can all VOLUME 13, 2025 194765 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 20. Summary of Machine learning models in WBANs Research (2017-2025). Highlights the predominant machine learning approaches and their primary applications, including communication optimization, anomaly detection, and personalized healthcare. distort readings, reducing model accuracy. Two advanced approaches, autoencoders for noise reduction and RL for adaptive sensing, show real promise. Autoencoders clean the raw stream and pull out the most informative features, while RL adjusts sampling parameters on the fly to suit current conditions [177]. Robust anomaly-detection pipelines and good data pre-processing are still essential to feed the model reliable input. Finally, since most wearables run on small batteries, energy efficiency is a constant concern. Demanding ML workloads 194766 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques can quickly drain power and shorten service life. Future research must consider power budgets, reconciling model complexity with the practical constraints of wearable devices. Lightweight ML models like MobileNet and model compression strategies like pruning have been created to lessen this. These methods preserve model performance while lowering the computational overhead [178]. In addition, FL minimizes the energy centralized processing uses by enabling dispersed computation across several devices. By utilizing these strategies, WBANs can extend battery life without sacrificing functionality. Because WBANs handle highly sensitive health information, keeping that data secure and private is essential. A breach during transmission or storage could have serious consequences for patients. FL helps reduce this risk by keeping raw data on each person’s device; only the updates from the trained model are shared, so the original information never leaves the patient’s control. Furthermore, GANs and GenAI can generate privacy-preserving datasets for training ML models [157],[158]. As shown in Table 13, most existing ML-based security approaches in WBANs concentrate on countering spoofing, tampering, and information leakage at the communication layer. In contrast, very little attention has been given to safeguarding the learning models themselves. None of the reviewed studies directly addresses the risk of model poisoning or Byzantine behavior in federated or distributed WBAN settings. This gap suggests a need for future work on robust aggregation strategies, trust-aware federated learning, and secure validation of model updates, to maintain the integrity and reliability of ML models used in continuous healthcare monitoring applications. Blockchain-based authentication systems, decentralized data processing, and homomorphic encryption are all viable options for secure computations. These solutions enhance the security and privacy of WBAN systems while boosting stakeholder and user confidence. Clear and understandable ML models for WBANs are necessary for healthcare practitioners to trust and utilize these solutions in clinical. The frequent lack of transparency in black-box models’ explanations of forecasts may make them harder to accept. To address this issue, XAI methodologies such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) have been formulated. By showing how models make decisions, these methods make them clearer [188]. Furthermore, attention-based models make it easier to understand by showing which parts of the decision-making process are most important. These advancements are crucial for the integration of ML into medical practices. For prompt healthcare actions, real-time processing is crucial, especially in important situations like identifying irregularities or anticipating crises. However, in dynamic contexts, typical ML models could have trouble with latency and adaptation. Low-latency DNNs or RL in conjunction with edge computing have been proposed to remedy this issue. Through data processing locally on the device, edge computing eliminates the delays that come with sending data to distant servers [189]. These methods allow for real-time analysis and flexibility, guaranteeing that WBANs can react quickly to changing situations. When used in WBANs, ML models must function reliably across patient demographics and medical situations. Poor generalization may result from variations in data distribution and environmental factors. Transfer learning and meta-learning have been presented as solutions to this problem. Meta-learning teaches a model of how to learn, so it can pick up new activities or settings with very little extra training. Transfer learning lets a model that was trained in one job adapt quickly to a different one [180]. Teaming up several models in an ensemble makes predictions sturdier and more accurate, an advantage when you scale up tiny bio-nano sensors or wireless brain transceivers, where reliability matters most. Progress is slower than it should be, though, because researchers still don’t have a common yardstick for judging ML tools in WBANs. We need a shared framework that weighs accuracy, response time, battery use, and how easy the model is to interpret, all at once. A multi-metric scorecard like that would spotlight trade-offs, steer design choices, and make it possible to compare results across studies [179]. With standardized benchmarks in place, the field could innovate and improve much faster. However, there are advantages and disadvantages of integrating ML into WBANs. To fully exploit the potential of ML-driven WBANs, specific solutions are needed for data quality, energy efficiency, security, interpretability, real-time processing, scalability, and evaluation metrics. Advanced ML techniques such as XAI, FL autoencoders, and edge computing provide promising approaches to tackling these issues. By implementing these concepts, researchers and practitioners may improve WBANs’ usability, efficiency, and dependability, improving patient care. Table 21 lists the main ML options now used to tackle core WBAN challenges. Beyond algorithmic improvements, the integration of ML in WBANs will heavily rely on innovations in sensing and wireless communication hardware to support more accurate and adaptive physiological monitoring. A. EMERGING MICROWAVE SENSING AND COMMUNICATION TECHNOLOGIES FOR INTELLIGENT WBANs The incorporation of ML in WBANs will progressively rely on enhancements in sensing and communication hardware capable of delivering more comprehensive and dependable data. Future WBAN systems are anticipated to utilize microwave-based sensors that function on the principles of dielectric perturbation, wherein variations in biological tissue permittivity result in quantifiable alterations in resonant frequency and quality factor. These sensors, which use dualfrequency, meandered microstrip, and substrate-integrated VOLUME 13, 2025 194767 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques TABLE 21. Summary of the ML options to address various problems in WBANs. waveguide architectures, are better for continuous physiological monitoring because they are more accurate, smaller, and more flexible [8],[9],[190],[191],[192],[193]. These advancements create novel research opportunities for machine learning models that can handle high-frequency biomedical data, execute feature extraction, and adjust to real-time physiological fluctuations. Meanwhile, in WBAN environments, next-generation wireless communication technologies, like substrateintegrated and leaky-wave antenna systems, will enable low-latency, interference-resistant, and energy-efficient data transmission [10],[11],[194],[195]. Future studies should concentrate on cross-layer integration, in which machine learning algorithms dynamically engage with the communication and sensing layers to accomplish intelligent energy management, adaptive slot allocation, and optimal data flow. Therefore, new developments in wireless and microwave technology will be crucial to the development of intelligent, self-optimizing WBAN architectures driven by ML. B. TOWARD STANDARDIZED BENCHMARKS FOR ML EVALUATION IN WBANs A key challenge found across the reviewed studies is the absence of standardized datasets and evaluation protocols for testing ML models in WBANs. Most studies depend on private or inconsistent data gathered under varying experimental conditions, which makes it hard to reproduce results or compare models accurately. This inconsistency affects performance reporting for key parameters, including accuracy, latency, and energy consumption. Recent studies in wearable computing and HAR have shown that open and unified benchmarking frameworks improve research comparability and transparency [196], [197],[198]. Publicly available datasets such as Mobile Health (MHEALTH), Wearable Stress and Affect Detection (WESAD), Physical Activity Monitoring (PAMAP2), and Massachusetts Institute of Technology–Beth Israel Hospital (MIT-BIH) Arrhythmia already provide diverse physiological and motion data that are relevant for WBAN applications. However, these datasets are often used inconsistently across studies. To overcome this limitation, future work should focus on developing an open-source WBAN-ML benchmark framework that includes: 1) Dataset integration: A curated collection of representative public WBAN datasets (e.g., MHEALTH, WESAD, PAMAP2, MIT-BIH). 2) Standard preprocessing pipeline: Common data cleaning, segmentation, and normalization steps to ensure consistent input across models. 3) Unified evaluation metrics: A fixed set of indicators such as F1-score, latency (ms), energy per inference (mJ), model size (kB), and interpretability score (XAI-score) to enable objective performance comparison. 4) Community leaderboard: A shared platform for researchers to upload results and compare models under the same experimental settings. Creating such a standardized benchmark would enhance reproducibility, fairness, and collaboration in WBAN-ML research. It would also accelerate the development of reliable, energy-efficient, and explainable ML models for real-time healthcare monitoring. 194768 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques FIGURE 16. Explainability–Accuracy Pareto frontier of ML models integrated into WBANs for healthcare applications. The plot visualizes the trade-off between model explainability (transparency) and predictive accuracy across different ML paradigms. Larger bubble sizes indicate a higher frequency of occurrence in the reviewed studies (2017–2025). Models such as DL and DRL achieve high accuracy but low explainability, whereas fuzzy logic (FL), unsupervised, and supervised ML approaches provide better interpretability with moderate accuracy. C. EXPLAINABILITY–ACCURACY TRADE-OFF IN ML MODELS FOR WBAN HEALTHCARE A key challenge in applying machine learning within WBAN-based healthcare systems is finding an appropriate balance between predictive accuracy and interpretability. Deep learning models—including CNNs, RNNs, and newer hybrid deep reinforcement learning approaches— often achieve extremely high performance, with reported F1-scores frequently above 95%. However, these models operate as ‘‘black boxes,’’ making it difficult for clinicians to understand or verify how decisions are being made. On the other hand, more transparent models such as decision trees, random forests, and fuzzy logic systems allow the reasoning process to be examined and validated, though they may offer slightly lower predictive accuracy. Figure 16 illustrates this trade-off by showing the accuracy–explainability Pareto frontier derived from the 55 studies reviewed. The figure indicates that deep learning and deep RL approaches cluster in the region of highest accuracy but lowest interpretability, whereas methods like decision trees, random forests, and fuzzy logic occupy a middle ground where both interpretability and performance are balanced. This observation is consistent with recent literature stressing that explainability is essential for clinical trust and acceptance in medical IoT and WBAN settings [121],[139],[182]. Looking forward, research should move toward multiobjective optimization strategies that explicitly account for both accuracy and interpretability. This may involve evaluating models using interpretability metrics such as SHAP or LIME alongside conventional performance measures like F1-score and precision. Establishing such dual evaluation standards will support the development of machine learning systems that are not only accurate but also transparent and reliable for real-time healthcare monitoring in WBAN environments. D. CLINICAL VALIDATION PATHWAYS, PATIENT COMPLIANCE, AND REGULATORY CONSIDERATION 1) CLINICAL VALIDATION PATHWAYS Clinical validation tests WBANs’ safety and performance in their intended environment. The process includes preclinical VOLUME 13, 2025 194769 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques [155] G. Chen, X. Liu, M. Shorfuzzaman, A. Karime, Y. Wang, and Y. Qi, ‘‘MEC-based jamming-aided anti-eavesdropping with deep reinforcement learning for WBANs,’’ ACM Trans. Internet Technol., vol. 22, no. 3, pp. 1–17, Aug. 2022, doi: 10.1145/3453186. [156] M. M. Islam and M. Shamshuzzoha, ‘‘Securing wireless body area networks data transmission with machine learning: A cross-tier framework for anomaly detection and intrusion prevention,’’ Comput. Structural Biotechnol. Rep., vol. 2, Jun. 2025, Art. no. 100031, doi: 10.1016/j.csbr.2025.100031. [157] G. Aceto, F. Giampaolo, C. Guida, S. Izzo, A. Pescapè, F. Piccialli, and E. Prezioso, ‘‘Synthetic and privacy-preserving traffic trace generation using generative AI models for training network intrusion detection systems,’’ J. Netw. Comput. Appl., vol. 229, Sep. 2024, Art. no. 103926, doi: 10.1016/j.jnca.2024.103926. [158] C. Uddagiri and B. V. Isunuri, ‘‘Ethical and privacy challenges of generative AI,’’ in Studies in Computational Intelligence. Singapore: Springer, 2024, pp. 219–244, doi: 10.1007/978-981-97-8460-8_11. [159] J. Kim, B. Kim, C. You, and H. Park, ‘‘Priority based adaptive slotted ALOHA method using Q-learning,’’ J. Korean Inst. Commun. Inf. Sci., vol. 48, no. 3, pp. 350–358, Mar. 2023, doi: 10.7840/kics.2023.48.3.350. [160] S. Li, H.-C. Yang, F. Xu, H. Hu, and F. Hu, ‘‘Energy-efficient relay transmission for WBAN: Energy consumption minimizing design with hybrid supervised/reinforcement learning,’’ IEEE Internet Things J., vol. 11, no. 10, pp. 17770–17779, May 2024, doi: 10.1109/JIOT.2024.3361772. [161] A. P. Kunal and B. T. Sagar, ‘‘Critical healthcare assessment using WBAN and SVM,’’ Int. J. Innov. Technol. Exploring Eng., vol. 10, no. 9, pp. 84–86, Jul. 2021, doi: 10.35940/ijitee.i9362.0710921. [162] N. Bilandi, H. K. Verma, and R. Dhir, ‘‘An intelligent and energy-efficient wireless body area network to control coronavirus outbreak,’’ Arabian J. Sci. Eng., vol. 46, no. 9, pp. 8203–8222, Sep. 2021, doi: 10.1007/s13369021-05411-2. [163] R. Mohammadi and Z. Shirmohammadi, ‘‘DRDC: Deep reinforcement learning based duty cycle for energy harvesting body sensor node,’’ Energy Rep., vol. 9, pp. 1707–1719, Dec. 2023, doi: 10.1016/j.egyr.2022.12.138. [164] Y. Chen, S. Han, G. Chen, J. Yin, K. N. Wang, and J. Cao, ‘‘A deep reinforcement learning-based wireless body area network offloading optimization strategy for healthcare services,’’ Health Inf. Sci. Syst., vol. 11, no. 1, p. 8, Jan. 2023, doi: 10.1007/s13755-023-00212-3. [165] Z. Sadreddini, Ö. Makul, T. Çavdar, and F. B. Günay, ‘‘Performance analysis of licensed shared access based secondary users activity on cognitive radio networks,’’ in Proc. Electric Electron., Comput. Sci., Biomed. Eng. Meeting (EBBT), Apr. 2018, pp. 1–4, doi: 10.1109/EBBT.2018.8391442. [166] N. Sharma, H. Chadha, K. Singh, B. M. Singh, and N. Pathak, ‘‘A novel hybrid clustering based transmission protocol for wireless body area networks,’’ Comput., Mater. Continua, vol. 69, no. 2, pp. 2459–2473, 2021, doi: 10.32604/cmc.2021.014305. [167] P. Khoshvaght, J. Tanveer, A. M. Rahmani, M. Mohammadi, A. Mehranzadeh, J. Lansky, and M. Hosseinzadeh, ‘‘H-TERF: A hybrid approach combining fuzzy multi-criteria decision-making techniques and enhanced random forest to improve WBAN-IoT,’’ Internet Things, vol. 32, Jul. 2025, Art. no. 101613, doi: 10.1016/j.iot.2025.101613. [168] D. P. Q. Carneiro, A. A. Cardoso, and F. H. T. Vieira, ‘‘Adaptive resource allocation in 5G CP-OFDM systems using Markovian model-based reinforcement learning algorithm,’’ Neural Comput. Appl., vol. 35, no. 13, pp. 9421–9435, May 2023, doi: 10.1007/s00521-023-08406-2. [169] S. Suknum, C. Thoasiri, and N. Jinaporn, ‘‘Q-learning-based resource allocation in heterogeneous cellular networks,’’ in Proc. Int. Electr. Eng. Congr. (iEECON), Mar. 2022, pp. 1–3, doi: 10.1109/iEECON53204.2022.9741639. [170] V. Aruna, L. Anjaneyulu, and C. Bhar, ‘‘Deep-Q reinforcement learning based resource allocation in wireless communication networks,’’ in Proc. IEEE Int. Symp. Smart Electron. Syst. (iSES), Dec. 2022, pp. 66–72, doi: 10.1109/ISES54909.2022.00025. [171] Z. Zheng, S. Jiang, R. Feng, L. Ge, and C. Gu, ‘‘Survey of reinforcementlearning-based MAC protocols for wireless ad hoc networks with a MAC reference model,’’ Entropy, vol. 25, no. 1, p. 101, Jan. 2023, doi: 10.3390/e25010101. [172] Y.-H. Li, Y.-L. Li, M.-Y. Wei, and G.-Y. Li, ‘‘Innovation and challenges of artificial intelligence technology in personalized healthcare,’’ Sci. Rep., vol. 14, no. 1, p. 18994, Aug. 2024, doi: 10.1038/s41598-024-70073-7. [173] B.-S. Kim, K.-I. Kim, and B. Shah, ‘‘BANSIM: A new discrete-event simulator for wireless body area networks with deep reinforcement learning in Python,’’ J. Syst. Archit., vol. 126, May 2022, Art. no. 102489, doi: 10.1016/j.sysarc.2022.102489. [174] L. Wang, Z.-H. You, D.-S. Huang, and F. Zhou, ‘‘Combining high speed ELM learning with a deep convolutional neural network feature encoding for predicting protein-RNA interactions,’’ IEEE/ACM Trans. Comput. Biol. Bioinf., vol. 17, no. 3, pp. 972–980, May 2020, doi: 10.1109/TCBB.2018.2874267. [175] H. Su, M.-S. Pan, H. Chen, and X. Liu, ‘‘MDP-based MAC protocol for WBANs in edge-enabled eHealth systems,’’ Electronics, vol. 12, no. 4, p. 947, Feb. 2023, doi: 10.3390/electronics12040947. [176] B.-S. Kim, B. Shah, T. He, and K.-I. Kim, ‘‘A survey on analytical models for dynamic resource management in wireless body area networks,’’ Ad Hoc Netw., vol. 135, Oct. 2022, Art. no. 102936, doi: 10.1016/j.adhoc.2022.102936. [177] N. Rajathi, S. Divya, S. A. Elavarasi, G. Saritha, and V. J. Ramya, ‘‘Adaptive intrusion detection in cyber-physical systems using reinforcement learning-based autoencoders,’’ Int. Conf. Integr. Intell. Commun. Syst. (ICIICS), vol. 2024, pp. 1–7, Nov. 2024, doi: 10.1109/iciics63763.2024.10859561. [178] P. V. Dantas, W. Sabino da Silva, L. C. Cordeiro, and C. B. Carvalho, ‘‘A comprehensive review of model compression techniques in machine learning,’’ Int. J. Speech Technol., vol. 54, no. 22, pp. 11804–11844, Nov. 2024, doi: 10.1007/s10489-024-05747-w. [179] I. El Badisy, N. Graffeo, M. Khalis, and R. Giorgi, ‘‘Multi-metric comparison of machine learning imputation methods with application to breast cancer survival,’’ BMC Med. Res. Methodology, vol. 24, no. 1, p. 191, Aug. 2024, doi: 10.1186/s12874-024-02305-3. [180] A. Vettoruzzo, M.-R. Bouguelia, J. Vanschoren, T. Rögnvaldsson, and K. Santosh, ‘‘Advances and challenges in meta-learning: A technical review,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 7, pp. 4763–4779, Jul. 2024, doi: 10.1109/TPAMI.2024. 3357847. [181] R. Roy, S. Mukherjee, M. M. Baral, A. K. Badhan, and M. Ravindra, ‘‘Challenges encountered in the implementation of machine learning in the healthcare industry,’’ in Proc. Int. Conf. Mach. Learn. Big Data Anal., 2023, pp. 377–386. [182] R. N. L. S. Kalpana, A. K. Patro, and D. N. Rao, ‘‘Finding the efficiency of ConvBi-LSTM over anticipation of adversaries in WBANs,’’ Recent Patents Eng., vol. 19, no. 1, pp. 1–18, Jan. 2025, doi: 10.2174/0118722121255695231008171935. [183] T. Thamaraimanalan and S. Ramalingam, ‘‘Enhancing anomaly detection in WBANs using hybrid deep learning and optimization algorithms,’’ Neural Comput. Appl., vol. 37, no. 15, pp. 9223–9243, May 2025, doi: 10.1007/s00521-025-11061-4. [184] A. K. Pipal and R. Jagadeesh Kannan, ‘‘Graph attention layer-based WideResNet for efficient adversarial attack detection in wireless body area networks,’’ Inf. Secur. J., A Global Perspective, vol. 2025, pp. 1–15, Apr. 2025, doi: 10.1080/19393555.2025.2493102. [185] J. C. Hong, N. C. W. Eclov, S. J. Stephens, Y. M. Mowery, and M. Palta, ‘‘Implementation of machine learning in the clinic: Challenges and lessons in prospective deployment from the system for high intensity EvaLuation during radiation therapy (SHIELD-RT) randomized controlled study,’’ BMC Bioinf., vol. 23, no. S12, p. 408, Sep. 2022, doi: 10.1186/s12859-022-04940-3. [186] I. Hussain and A. U. Islam, ‘‘Research direction toward IoT-based machine learning-driven health monitoring systems: A survey,’’ in Proc. ICCVBIC, 2023, pp. 541–555, doi: 10.1007/978-981-19-9819-5_39. [187] S. Vyas and S. Gupta, ‘‘WBAN-based remote monitoring system utilising machine learning for healthcare services,’’ Int. J. Syst. Syst. Eng., vol. 13, no. 1, p. 100, 2023, doi: 10.1504/ijsse.2023.129054. [188] D. E. Mathew, D. U. Ebem, A. C. Ikegwu, P. E. Ukeoma, and N. F. Dibiaezue, ‘‘Recent emerging techniques in explainable artificial intelligence to enhance the interpretable and understanding of AI models for human,’’ Neural Process. Lett., vol. 57, no. 1, p. 16, Feb. 2025, doi: 10.1007/s11063-025-11732-2. [189] A. Amzil, M. Hanini, and A. Zaaloul, ‘‘Modeling and analysis of LoRaenabled task offloading in edge computing for enhanced battery life in wearable devices,’’ Cluster Comput., vol. 28, no. 3, p. 201, Jun. 2025, doi: 10.1007/s10586-024-04925-2. [190] S. Kiani and P. Rezaei, ‘‘Microwave substrate integrated waveguide resonator sensor for non-invasive monitoring of blood glucose concentration: Low cost and painless tool for diabetics,’’ Measurement, vol. 219, Sep. 2023, Art. no. 113232, doi: 10.1016/j.measurement.2023.113232. [191] S. Kiani, P. Rezaei, and M. Fakhr, ‘‘Dual-frequency microwave resonant sensor to detect noninvasive glucose-level changes through the fingertip,’’ IEEE Trans. Instrum. Meas., vol. 70, pp. 1–8, 2021, doi: 10.1109/TIM.2021.3052011. 194776 VOLUME 13, 2025 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques [192] S. Kiani, P. Rezaei, and M. Fakhr, ‘‘Real-time measurement of liquid permittivity through label-free meandered microwave sensor,’’ IETE J. Res., vol. 70, no. 5, pp. 4606–4616, May 2024, doi: 10.1080/03772063.2023.2231875. [193] S. Kiani, P. Rezaei, M. Karami, and R. A. Sadeghzadeh, ‘‘Band-stop filter sensor based on SIW cavity for the non-invasive measuring of blood glucose,’’ IET Wireless Sensor Syst., vol. 9, no. 1, pp. 1–5, Feb. 2019, doi: 10.1049/iet-wss.2018.5044. [194] P. Sohrabi, P. Rezaei, S. Kiani, and M. Fakhr, ‘‘A symmetrical SIW-based leaky-wave antenna with continuous beam scanning from backward-to-forward through broadside,’’ Wireless Netw., vol. 27, no. 8, pp. 5417–5424, Nov. 2021, doi: 10.1007/s11276-021-02798-6. [195] S. Kiani, P. Rezaei, and M. Fakhr, ‘‘A CPW-fed wearable antenna at ISM band for biomedical and WBAN applications,’’ Wireless Netw., vol. 27, no. 1, pp. 735–745, Jan. 2021, doi: 10.1007/s11276-020-02490-1. [196] M. Kaseris, I. Kostavelis, and S. Malassiotis, ‘‘A comprehensive survey on deep learning methods in human activity recognition,’’ Mach. Learn. Knowl. Extraction, vol. 6, no. 2, pp. 842–876, Apr. 2024, doi: 10.3390/make6020040. [197] R. Samanta, B. Saha, S. K. Ghosh, and R. Babu Roy, ‘‘Optimizing TinyML: The impact of reduced data acquisition rates for time series classification on microcontrollers,’’ 2024, arXiv:2409.10942. [198] S. Davidashvilly, M. Cardei, M. Hssayeni, C. Chi, and B. Ghoraani, ‘‘Deep neural networks for wearable sensor-based activity recognition in Parkinson’s disease: Investigating generalizability and model complexity,’’ Biomed. Eng. OnLine, vol. 23, no. 1, pp. 1–24, Feb. 2024, doi: 10.1186/s12938-024-01214-2. [199] C. Venkatesh, L. Sivayamini, M. V. Dasu, D. S. Vyshnavi, D. Yasaswini, D. Sreenivasulu, G. Suryashankaravaraprasadad, and C. Nagaraju, ‘‘Cardiac diagnosis system for heart diseases classification based on deep learning and optimization strategies using ECG signals,’’ in Proc. Int. Conf. Comput., Electr. Commun. Eng. (ICCECE), Feb. 2025, pp. 1–8, doi: 10.1109/iccece61355.2025.10940447. [200] K. Pervez, M. Izhar, A. Ahmed, N. Alturki, and S. Abdullah, ‘‘Smart implantable devices for cardiac health: A novel self-powered wireless ECG monitoring system using energy harvesting and machine learning-driven anomaly detection,’’ Smart Health, vol. 37, Sep. 2025, Art. no. 100582, doi: 10.1016/j.smhl.2025.100582. [201] S. Karthika and K. J. Gnanaselvi, ‘‘A review of forensics security hazards and challenges in WBAN and healthcare systems,’’ in Security, Privacy, and Trust in WBANs and E-Healthcare. Singapore: Springer, Nov. 2024, pp. 63–82. [202] R. Abirami and C. Malathy, ‘‘Security requirements and challenges in WBANs and e-health systems,’’ in Security, Privacy, and Trust in WBANs and E-Healthcare, Nov. 2024, pp. 3–22. [203] Premarket Notification 510(K) | FDA. Accessed: Jul. 10, 2025. [Online]. Available: https://www.fda.gov/medical-devices/premarketsubmissions-selecting-and-preparing-correct-submission/premarketnotification-510k ABDU IBRAHIM ADAMU received the Bachelor of Science degree (Hons) in computer science from Kano University of Science and Technology (currently changed to Dangote University of Science and Technology), Kano, Nigeria, in 2014, and the M.Sc. degree in computer science with the Universiti Teknologi MARA, Shah Alam, Selangor, Malaysia, in 2023, where he is currently pursuing the Ph.D. degree in electrical engineering. His research interests include wireless communication, artificial intelligence, machine learning, big data analysis, and cloud computing. PRAVEEN KUMAR DONTA (Senior Member, IEEE) received the Bachelor of Technology and Master of Technology degrees (Hons.) from the Department of Computer Science and Engineering, JNTUA, Ananthapur, in 2012 and 2014, respectively, and the Ph.D. degree from the Department of Computer Science and Engineering, Indian Institute of Technology (Indian School of Mines), Dhanbad, in June 2021. He is currently an Associate Professor (Docent) with the Department of Computer and Systems Sciences, Stockholm University, Sweden. From July 2021 to June 2024, he was with the Distributed Systems Group, TU Wien, as a Postdoctoral Researcher. He was a Visiting Ph.D. Fellow with the Mobile and Cloud Laboratory, University of Tartu, Estonia, from July 2019 to January 2020. His current research is on learning-driven distributed computing continuum systems, casual and conscious continuum systems, and intelligent data protocols. He is an ACM Professional Member. He is an Editorial Board Member of IEEE INTERNET OF THINGS JOURNAL,Computing (Springer), ETT Wiley, POLS One,Measurement, and Computer Communications (Elsevier). DARMAWATY MOHD ALI received the degree (Hons.) in electrical, electronic, and systems engineering from the Universiti Kebangsaan Malaysia (UKM), in 1999, the master’s degree from the Universiti Teknologi Malaysia (UTM), and the Ph.D. degree from Universiti Malaya (UM), in 2012. She is an Associate Professor with the Faculty of Electrical Engineering, Universiti Teknologi MARA (UiTM). She started her first job as a Product Engineer. Her research interests include wireless access technology and the provision of quality of service (QoS) in wireless networks. SOHAIL SARANG (Senior Member, IEEE) received the B.Eng. degree in telecommunication engineering from Hamdard University, Karachi, Pakistan, in 2014, the M.Sc. degree in electrical and electronics engineering from the Universiti Teknologi PETRONAS, Malaysia, in 2018, and the Ph.D. degree in electrical and computer engineering from the Faculty of Technical Sciences, University of Novi Sad, Serbia. He is a Postdoctoral Researcher with the Department of Electrical Engineering, Faculty of Technical Sciences, University of Novi Sad. His research interests include energy harvesting communications, lowpower sensor networks, battery-free IoT, MAC protocols, and machine learning-driven communication algorithms and protocols. VOLUME 13, 2025 194777 A. I. Adamu et al.: Systematic Literature Review of Advanced Machine Learning Techniques GORAN M. STOJANOVIĆ (Member, IEEE) received the B.Sc., M.Sc., and Ph.D. degrees in electrical engineering from the Faculty of Technical Sciences (FTS), University of Novi Sad (UNS), Serbia, in 1996, 2003, and 2005, respectively. He is currently a Full Professor with FTS, UNS. He has 27 years of experience in research and development. He has over 18 years of experience in writing, implementing, and coordinating EU-funded projects (Horizon Europe, H2020, EUREKA, ERASMUS, and CEI), with a total budget exceeding 22.86 MEUR. He supervised 14 Ph.D. students, 40 M.Sc. students, and 60 diploma students at FTS-UNS. He is the author/co-author of 280 articles, including 180 in peer-reviewed journals with impact factors, five books, three patents, and two chapters in a monograph. He was a keynote speaker at 14 international conferences. His research interests include sensors, flexible electronics, textile electronics, edible electronics, and microfluidics. SUZI SEROJA SARNIN received the bachelor’s degree in electrical and electronics from the Universiti Teknologi Malaysia, in 1999, and the M.Sc. degree in microelectronics and the Ph.D. degree in electrical engineering from Universiti Kebangsaan Malaysia, in 2005 and 2018, respectively. From 1999 to 2001, she was a Quality Control Engineer with Memory Tech (M) Sdn. Bhd. In March 2001, she was a Contract Lecturer with the Universiti Teknologi MARA. She continued to work for the Department of Electrical Engineering, Universiti Teknologi MARA, as a Senior Lecturer. She has collaborated actively with researchers in several other electrical engineering disciplines and industries. Her research interests include wireless communication, multiple access, space-time coding and coding theories, signal processing, and the Internet of Things. 194778 VOLUME 13, 2025