Neurological disorders, such as Parkinson disease (PD), multiple sclerosis (MS), Alzheimer disease (AD), and stroke, are significant and growing global health concerns due to their high prevalence. A recent meta-analysis reported a global pooled prevalence of PD of 1.51 cases per 1000 individuals, with higher rates observed among male individuals and a significant increase over the past 4 decades [1]. Similarly, MS currently affects approximately 2.8 million people globally, with rising prevalence and incidence trends observed worldwide [2]. In Europe, the prevalence of AD has been estimated at 5.05% (95% CI 4.73%-5.39%) [3]. The incidence of clinically diagnosed AD per 1000 person-years increases markedly with age: 3.4 for individuals aged 65 to 74 years, 13.8 for those aged 75 to 84 years, and 35.8 for those aged ≥85 years [4,5]. Regarding cerebrovascular impairments, stroke continues to be a leading cause of death and disability, with >12 million new cases per year and >100 million survivors of stroke worldwide [6].

These conditions often lead to alterations in gait and balance, increasing the risk of falls. This risk is especially pronounced in older adults and individuals with neurological impairments. Approximately one-third of adults aged >65 years experience a fall each year [7], and this rate rises to nearly 50% among those aged >80 years [8]. Consequently, falls in this patient population considered fragile can worsen quality of life due to hospitalization, hypomobility, and reduced functional independence [9]. Fall-related injuries not only entail substantial medical costs but also determine patients’ mortality risk. Among neurological disorders, PD, MS, cerebrovascular accidents, and AD are among the most common, causing significant motor disability. It is noteworthy that these patients with neurological disorders are considered recurrent fallers due to their gait features. In particular, it is estimated that 39% of patients with PD experience falls, while 40% to 60% of patients with MS experience recurrent falls [10]. Similarly, patients with stroke are at high risk for falls, with fall frequencies ranging from 27% to 39% in rehabilitation hospitals [10]. Furthermore, older adults with AD are more than twice as likely to experience a fall compared to cognitively healthy peers, with rates ranging from 60% to 80% [11]. However, each neurological disorder is characterized by considerable heterogeneity in gait disturbances due to differences in underlying etiology [12]; for example, neurodegenerative disorders (such as PD, MS, and AD) are associated with distinct gait patterns, characterized by reduced gait speed, shorter step length, and increased muscle weakness [12]. In addition, ambulation in patients with PD is characterized by shorter steps, loss of dissociated arm and trunk movements during gait, and postural instability [13]. By contrast, patients with MS manifest pathological gait features related to spasticity, ataxia, muscle weakness, and sensory and proprioceptive deficits, all of which contribute to progressive disability [14]. AD independently increases fall risk due to its associated impairments in judgment, gait, visuospatial perception, and hazard recognition [15]. Differently from neurodegenerative disorders, patients with stroke are characterized by a marked spatiotemporal asymmetry between steps, which substantially increases fall risk [16]. For these reasons, gait and balance assessment in patients with neurological disorders is essential in all phases of the rehabilitation process.

To this aim, clinical scales (ie, the timed up and go [TUG] test, the Tinetti Scale, the Falls Efficacy Scale–International, and the Berg Balance Scale) as well as innovative technologies, including wearable sensors and nonwearable devices, are used. Specifically, wearable sensors are growing in popularity in both clinical and research contexts due to their suitable features [17]: because of their small size, affordability, and user-friendly nature, wearable sensors represent an optimal solution for health monitoring [18,19]. Unlike traditional laboratory-based monitoring systems, wearable sensors can facilitate the continuous monitoring of motor activity by capturing physiological data during everyday routines [20]. In addition, fall risk assessments with laboratory-based systems are generally performed in supervised conditions, in which the behavior adopted by participants may not be representative of the one adopted in everyday life—during the experimental tasks, the participants might be attempting their “best effort” [21]. Generally, falls occur in unpredictable situations during everyday activities [22]. In this context, wearable sensors enable unsupervised fall risk assessment by providing continuous monitoring during daily functional tasks, thereby offering a reflection of the individual’s real-world fall risk [20,21,23].

Thus, interest in wearable sensors has been increasing with regard to monitoring fall risk among older adults [21]. Among the prevalent wearable technologies used for health monitoring are inertial measurement units (IMUs), which gather data from accelerometers and gyroscopes. IMUs can be strategically positioned on the body to capture motion data for subsequent analysis, integration, and interpretation [24]. Gait characteristics that can be analyzed include stride length, stance-to-swing phase ratio, and various other spatiotemporal gait parameters [21,25,26]. In addition, data from inertial sensors collected during walking can be used to investigate upper body dynamics through nonlinear analysis techniques, such as recurrence quantification analysis [27]. Some researchers have used inertial sensor data to detect balance alterations by analyzing sway speed in the anterior-posterior and medial-lateral directions of the body’s center of mass projection path in patients with AD [28].

Other wearable sensors include insole-based devices, which have been developed for fall risk assessment and detection. These devices are embedded in the shoe sole to capture data on plantar pressure distribution [21]. The features extracted from these devices are used to objectively quantify the risk of falling in individuals with neurological disorders. However, the selection of the most reliable biomechanical gait parameters related to fall risk remains an open question [21]. In addition, the information acquired from wearable sensors, along with data processing, data fusion, and analysis techniques such as machine learning (ML) models or manual data analysis, is used to extract meaningful insights from the investigated parameters [29].

To this aim, some studies have investigated the role of deep learning (DL) in the context of neurorehabilitation. The DL-based approach can be used to extract data on stride-specific gait parameters from IMUs [30]. This approach could be useful in the field of telemedicine as a potential in-home gait monitoring tool, given that alterations in gait function can serve as a biomarker for predicting fall risk.

In this context, our first objective was to systematically review the literature on the most reliable biomechanical gait parameters, as detected by wearable sensors, for evaluating fall risk in individuals with neurological disorders. Therefore, we selected studies that involved patients with PD, MS, AD, or stroke, as these conditions collectively represent a significant proportion of neurological disorders worldwide and are associated with substantial gait impairment and a high risk of falls. Our second objective was to investigate existing literature on wearable sensor technology, including best practices for device placement as well as data processing and analysis.

We conducted a comprehensive systematic review to explore the existing evidence on fall risk assessment in patients with neurological disorders. We summarized the results of all published studies in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [31]. The PRISMA checklist is available in Multimedia Appendix 1. The protocol has been registered with PROSPERO (CRD42023463944).

We used the population, exposure, comparator, outcomes model to define the search strategy [32] (Table 1). The population, exposure, comparator, outcomes framework was applied to address the following research question: “what are the most reliable biomechanical gait parameters detected with wearable sensors to assess fall risk in neurological disorders?”

We conducted a systematic search for all peer-reviewed articles published up to April 18, 2025, using PubMed, Web of Science, Embase, and IEEE Xplore, which are the most used databases in the fields of health and bioengineering. The full search queries are provided in Multimedia Appendix 2. We included studies involving adults (aged ≥18 y) with stroke, PD, MS, or AD. Specifically, the inclusion criteria were as follows: (1) adult patients with cerebrovascular impairments, PD, MS, or AD; (2) fall risk assessment using wearable sensors; (3) articles written in English; and (4) published in a peer-reviewed journal. We excluded articles describing theoretical models, methodological approaches, algorithms, and basic technical descriptions. We also excluded animal studies, conference proceedings and reviews, studies involving children, and case reports. The list of articles was refined for relevance, revised, and summarized, with the key topics identified from the summaries based on the inclusion and exclusion criteria. Considering the limited literature available, we included the following study designs in the qualitative synthesis: (1) randomized controlled trials, (2) observational studies, (3) cross-sectional studies, (4) case-control studies, and (5) cohort studies. All search results were imported into Rayyan (Rayyan Systems Inc) [33], a web application designed to facilitate the systematic review process, and screened by 2 blinded reviewers (AI and MB). Each reviewer independently screened the articles to ensure consistent application of the inclusion and exclusion criteria. After screening titles and abstracts, blinding was removed. In cases of disagreement, the 2 reviewers discussed the inclusion or exclusion of the undecided articles to reach consensus.

After full-text selection, data from the included studies were extracted and organized in a spreadsheet. Information was summarized using the following fields: assigned ID number, study title, year of publication or presentation and first author, study aims and design, study duration, recruitment method and setting, inclusion and exclusion criteria, use of a control group, use of wearable sensors, informed consent, conflicts of interest and funding, type of intervention or evaluation and control, number of participants, baseline characteristics, intervention setting, types of biomechanical parameters related to fall risk, results, and key conclusions. Data extraction was performed independently and in a blinded manner by 2 reviewers. In cases of disagreement, a third reviewer was consulted to resolve discrepancies.

The risk of bias in the included studies was assessed using the National Heart, Lung, and Blood Institute Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies [34]. This tool comprises 14 questions, with each item rated as “yes,” “no,” “cannot determine,” “not reported,” or “not applicable.” On the basis of the responses, the overall quality of each study was categorized as good, fair, or poor. The risk of bias was assessed independently and in a blinded manner by 2 reviewers. In cases of disagreement, a third reviewer was consulted to resolve discrepancies.

A total of 134 articles were identified through database search (PubMed: n=25, 18.7%; Web of Science: n=64, 47.8%; Embase: n=26, 19.4; and IEEE Xplore: n=19, 14.2%). After removing duplicates (57/134, 42.5%), we screened the remaining articles (77/134, 57.5%). Of these 77 articles, 69 (90%) met the inclusion criteria, and 8 (10%) were excluded. After full-text analysis, of the 69 articles, 50 (72%) were excluded, and 19 (28%) studies involving 5859 participants were included in the review synthesis (Figure 1). The included studies were published between 2016 and 2025.

The risk of bias in the included studies was assessed using the National Heart, Lung, and Blood Institute Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies, which comprises 14 questions (Figure 2)

All included studies were rated by the reviewers as having overall good quality (Figure 2), and there was homogeneity in the methodologies and procedures used to assess fall risk in neurological disorders. However, concerns were identified in some domains of the quality assessment tool, specifically items 3, 5, 12, and 13; for example, participation rates of eligible individuals were documented in only 5% (1/19) of the studies [50], a sample-size justification or power analysis was provided only in the study by Sturchio et al [39], repeat assessment of the exposure was described in only 21% (4/19) of the studies [48,49,51,52], and 79% (15/19) of the studies relied on a single measurement or used a cross-sectional design. Moreover, procedures for blinding outcome assessors to exposure status were not reported, and information sufficient to determine loss to follow-up was not provided in the prospective cohort studies. However, these reporting limitations did not affect the overall quality classification assigned to each study.

The included studies involved 2630 patients with neurological disorders, including 226 (8.59%) with MS (n=7, 37% studies), 2305 (87.64%) with PD (n=10, 53% studies), 51 (1.94%) with stroke (n=3, 16% studies), and 48 (1.83%) with AD or cognitive impairment (n=1, 5% study). An additional 3229 participants were healthy controls or older adults. Regarding disease severity, the studies conducted in patients with MS had a mean Expanded Disability Status Scale score of 3.4; and in patients with PD, the mean score on the Unified Parkinson’s Disease Rating Scale, part III (motor examination), was 26.6, while the mean Hoehn and Yahr stage was 2.58 (based on the studies that reported these metrics). Regarding patients with stroke, 2 (11%) of the 19 studies [36,38] included participants who had experienced a cerebrovascular accident at least 5 years earlier, while 1 (5%) included patients who had experienced a stroke at least 6 months before the day of the experiment [40] (Table 2).

We classified our results according to the type of motor task (supervised vs unsupervised), the biomechanical parameters extracted, and the technological equipment used as well as the artificial intelligence (AI) approach applied in the included studies.

In this systematic review, we emphasize the important role of wearable technologies in performing objective and quantitative evaluations of fall risk among patients with neurological disorders. While the studies included were generally well conducted, we noted considerable variability in sensor placement, the types of wearable sensors, motor tasks, and assessment methodologies. In addition, the lack of a standardized approach to fall risk assessment using wearable sensors likely reflects the field’s ongoing technological evolution. To the best of our knowledge, this is the first systematic review analyzing the biomechanical features detected from wearable sensors for fall risk assessment in patients with PD, MS, AD, or stroke. Other studies [21] have investigated the role of inertial sensor–based and insole-based wearable sensors to evaluate fall risk in older adults. In line with our results, these studies identified several biomechanical gait parameters, including spatiotemporal parameters, balance features (center of pressure trajectory), and ADLs that were registered to detect fall risk. However, in our systematic review, we found some differences in the assessment of the risk of falling related mainly to the characteristics of the neurological disorders considered. Furthermore, we noticed differences between the studies concerning technological equipment.

Moreover, another systematic review [56] examined the use of wearable sensors to detect freezing of gait and falls in patients with PD. Consistent with our findings, the authors identified the most common sensor placements, such as the thigh, ankle, and lower back, which align with those reported in our review (Figure 3). However, our study presents some key differences: we included a broader range of neurological disorders (eg, MS, AD, and stroke), focused specifically on fall detection and fall risk assessment, and incorporated more recent literature.

Falls are very common in individuals with neurological disorders and are influenced by various aspects and features of gait. According to Zhou et al [57], gait variables related to pace (eg, stride and velocity), rhythm (eg, stride duration as well as stance and swing time), variability (eg, SD of ankle dorsiflexion at heel strike), and spatial parameters (eg, stride length, plantar flexion at toe-off, and ankle dorsiflexion at heel strike) can help classify individuals with neurological disorders as fallers. However, neurological disorders are heterogeneous, and each condition is associated with specific gait features that could lead to an increased fall risk; for instance, in patients with MS, although the location, number, and size of MS lesions vary among individuals, certain common characteristics in gait deficits associated with MS have been observed. Patients with MS [58] may exhibit shorter steps, reduced cadence, prolonged double support time, and increased swing phase time. Moreover, some studies have emphasized that patients with MS have reduced joint motion at the hip and ankle and increased joint motion at the knee, with a hyperextension that is more pronounced during midstance [59]. Falls in this patient population can occur early after disease onset; and some factors increase the risk, such as poor visual acuity, impaired postural control, and altered proprioception. However, the studies included in this systematic review analyzed only specific biomechanical metrics related to clinical tests (eg, 30CST) [23,41] or analyzed spatiotemporal features of gait and turning angle [37,44,52]. According to Tulipani et al [41], the 30CST could be a useful test to discriminate “fallers” and “nonfallers” in patients with MS, as it can detect balance impairments and muscle fatigue, both of which are associated with falls. Another important aspect is that remote fall risk monitoring using wearable sensors has the potential advantage of augmenting clinical visits to gain a broader perspective of patient function, symptom fluctuation, and fall risk. In patients with PD, a critical factor that is associated with falls is freezing of gait. However, it is not the only risk factor involved in falls. People with PD can experience bradykinesia, muscle weakness, and body rigidity, all of which contribute to slower gait and are closely linked to fall rates [60]. According to Pelicioni et al [61], reduced quadriceps strength, poor balance and postural stability, and reduced mobility (as indicated by slower TUG test performance) could explain the increased prevalence of balance-related falls. These results are in line with those reported in the included studies; for example, Greene et al [46] found strong associations between fall counts and several individual gait and mobility parameters, including gait variability, during the TUG test. Gait variability refers to fluctuations in gait patterns, and in patients with PD, it reflects a loss of consistency in the ability to produce a steady gait rhythm, thus presenting a likelihood of a fall. In addition, Ma et al [45] found that patients with PD who experienced falls exhibited greater range of motion of the trunk in the sagittal plane than nonfallers. This aspect was identified by the authors as an independent risk factor for falls in patients with PD; and it could be used to predict falls, especially when it is combined with age, gender, and other gait parameters. From these results, it seems that kinematic gait parameters are accurate in detecting fall risk, as also confirmed by Sturchio et al [39]. The authors found that kinematic data extracted from wearable sensors showed greater predictive accuracy for falls than the Hoehn and Yahr scale, which is considered the most robust clinical predictor of falls in individuals with PD. Regarding patients with AD, Huang et al [50] reported that impaired memory is strongly associated with an increased fall risk in individuals with MCI and AD. Specifically, the authors identified TUG time, stride speed, and heel-strike angles as effective parameters for assessing fall risk in this population. However, our review identified only a single study addressing this critical aspect, highlighting a clear gap in the literature. Future research should further investigate this area to determine the most relevant biomechanical and cognitive features for fall risk assessment in patients with AD. Differently from neurodegenerative disorders (eg, MS, PD, and AD), in patients with stroke, both balance and gait deficits related to falls were observed. According to Weerdesteyn et al [62], stroke-related balance deficits include postural instability and lack of coordination in response to both self-induced and external balance perturbations. Gait deficits related to fall risk include reduced propulsion at push-off, decreased hip and knee flexion during the swing phase, and reduced stability during the stance phase. However, the selected studies [36,38] did not report specific biomechanical features related to falls in patients with stroke. In future studies, it could be essential to select and extract specific biomechanical features related to falls in this patient population.

Fall risk assessment and fall detection have been investigated using wearable technologies. Among these, accelerometers, IMUs, and insole-based systems were mostly used in the clinical context; for instance, triaxial accelerometers and inertial sensors were widely used in the included studies [24]. They measure body movement by detecting speed changes and calculating the resultant acceleration from the displacement of a mass element. In rehabilitation research, the most commonly used types of accelerometers include strain gauge, capacitive, piezoresistive, and piezoelectric sensors. These accelerometers typically have 1 to 3 sensing axes, enabling motion detection in 1D, 2D, or 3D. From our literature analysis [23,37,41,44,51,52], it emerged that triaxial accelerometers were combined with gyroscopes and EMG sensors to capture a wide range of motion data. In particular, gyroscopes calculate changes in angular motion by detecting Coriolis forces, which are proportional to the rate of angular rotation of the limb. Integrating accelerometer and EMG data from the lower limb provides more insightful findings, especially in the context of neurorehabilitation; for example, Meyer et al [48] collected EMG, accelerometer, and gyroscope data that allowed the investigation of foot drop, which has been identified as one of the most frequent causes of falls in individuals with MS. Interestingly, Tulipani et al [41] found significant associations between accelerometer-derived metrics and clinical assessments, notably in balance confidence and fatigue. These findings indicate that wearable sensors can offer a real-time, objective measure of fall risk that aligns well with traditional clinical evaluations. Moreover, we found that the studies on fall risk assessment in PD (10/19, 53%) primarily focused on integrating advanced sensor technologies and complex data analysis methods to monitor, predict, and manage fall risks in patients with PD. Some of these studies (2/19, 11%) used a full-body wearable motion sensor system consisting of 6 sensors (each including a triaxial accelerometer, a gyroscope, and a magnetometer) to detect kinematic gait data. By contrast, Silva de Lima [63] developed a triaxial accelerometer combined with a barometer, which automatically detected falls and allowed users to manually report falls via a button press. This dual functionality enabled the collection of accurate real-time data on fall events.

According to a previous review by Ferreira et al [64], studies involving wearable sensors for fall risk assessment mainly used 1 accelerometer, which underlines the importance of the use of acceleration data in interpreting scores from standard clinical scales. Data from the wearable sensors were used to detect prefall and unbalanced situations, enabling the identification of fall risk events. This approach aims to reduce short-term fall risk by allowing real-time daily monitoring of participants and providing immediate feedback when a fall risk event is detected.

In this context, some of the studies (7/19, 37%) performed the fall risk assessment in an unsupervised setting. Compared to laboratory settings, real-time fall risk prediction during ADLs is more suitable for detecting fall risk events. This approach allows continuous monitoring and enables alert systems to notify participants when fall risk events are detected. However, Meyer et al [48] found that the best performance of the ML model used to predict fall risk was achieved with laboratory data, indicating that controlled environments yield more reliable data for assessing fall risk. Supporting this, some studies have found that specific biomechanical features of gait (ie, speed) can vary depending on whether they are analyzed in the home or laboratory setting; for instance, Carcreff et al [65] found that children with cerebral palsy exhibited lower gait speeds in daily life compared to laboratory settings, potentially due to walking barefoot during laboratory-based assessments. Other studies have also observed discrepancies between supervised and unsupervised gait speeds in older adults. Takayanagi et al [66] identified a weak correlation between daily life and laboratory gait speeds, with daily speeds being significantly lower. De la Cámara et al [67] noted that clinical gait speed measurements might not accurately reflect habitual gait speeds and are influenced by physical, mental, and cognitive health. These studies, which used IMUs, faced limitations such as short walking distances in laboratory settings (2.44-10 meters), which do not adequately represent real-world gait. Even in PD, laboratory assessments do not always mirror daily activity accurately. Toosizadeh et al [68] found no significant correlation between laboratory and home gait parameters, possibly due to small sample sizes and methodological differences. Corrà et al [69] found that certain laboratory tests can better represent gait speed measured at home. However, the ability to remotely monitor patients at home is beneficial for clinicians because it allows them to assess patients’ motor capacity in their domestic environment. In this sense, both in-laboratory and daily living assessments using wearable sensors can provide complementary insights into PD, enhancing clinical evaluations and patient management.

The sensitivity of supervised versus unsupervised assessments for detecting motor fluctuations in PD remains unclear. Therefore, to enhance clinical care and design personalized interventions, a better understanding of gait disabilities is needed, particularly through comprehensive daily life performance data from IMUs.

Although wearable sensors have shown considerable potential in monitoring and detecting fall risk, it is important to acknowledge their limitations and the challenges associated with their use. One critical point is that predicting a fall is not equivalent to preventing it. Current systems are largely reactive, aiming to detect gait abnormalities or balance impairments before a fall occurs, but they do not directly intervene to prevent the event itself. This distinction is crucial, especially in the context of a rapidly aging global population, where fall-related injuries are expected to increase substantially. Therefore, future research should focus not only on improving prediction accuracy but also on integrating real-time intervention mechanisms (eg, alerts, assistive devices, or environmental modifications) that could help prevent falls in a practical setting [21].

Another key challenge lies in ergonomics and user acceptance of wearable technologies. Many current devices lack design considerations for comfort and aesthetics, which can lead to discomfort, skin irritation, or reluctance to use the device, particularly among older adults. These limitations negatively affect long-term adherence and reduce the ecological validity of continuous monitoring solutions. Addressing this issue requires a multidisciplinary approach involving engineers, designers, and end users to codevelop devices that are lightweight and easy to wear. In parallel, user-centered studies should investigate how different sensor placements, materials, and feedback systems influence usability and compliance. By overcoming these ergonomic barriers, wearable technologies can move closer to becoming a truly accessible and effective tool for fall risk management in real-life settings [70].

It is also important to note that data loss and signal noise can affect the reliability of sensor-derived features, especially in real-world settings. To mitigate these issues, careful attention must be given to preprocessing strategies and to the design of data acquisition protocols that minimize artefacts and signal interruptions. Moreover, differences in device types, sensor configurations, and data processing pipelines across studies make cross-study comparisons challenging. Standardization efforts and robust preprocessing techniques are essential to enhance the reproducibility and generalizability of the findings.

In the studies included in this systematic review, a range of AI techniques were applied to analyze data collected via wearable sensors for assessing fall risk in patients with neurological disorders. DL models were prominently featured [37,43,48]. These models are particularly proficient at processing sequential data and capturing temporal dependencies essential for interpreting complex gait patterns over time. This choice reflects a growing trend of using advanced AI methods to detect the complex patterns of movement disorders that are often overlooked by standard clinical assessments. Besides DL, traditional ML techniques also played a significant role; for instance, feature-based models such as support vector machines or random forests were used to classify patients based on the extracted biomechanical features, such as stride length and turning behavior [49,51,52]. Another crucial aspect identified in the included studies was the preprocessing of sensor data using advanced data fusion and signal processing techniques. These methods prepare the raw data for more effective AI analysis by enhancing data quality and extracting meaningful features. Arpan et al [44] used unscented Kalman filter techniques to integrate data from multiple sensors, facilitating a more accurate AI-based prediction of fall events [44]. Some of the included studies (3/19, 16%) explored the implementation of AI models that operate in real time to provide immediate insights or predictive analytics. This approach is particularly useful in a home monitoring context where immediate feedback can prevent falls [63].

Looking ahead, the integration of AI in fall risk assessment could expand in several directions. Tailoring AI models to individual patients’ data over time to adapt to changes in their condition or progression of their neurological disease could improve ongoing monitoring and therapeutic accuracy. By incorporating methodologies that adjust to patient-specific needs and conditions, AI can facilitate more dynamic interventions tailored to the evolving profiles of patients. In a recent study, Espinoza Bernal et al [71] used personalized ML algorithms to monitor and adjust individual rehabilitation strategies for patients with stroke. Using data from wearable sensors during a rehabilitation camp, these algorithms provided high classification accuracy and insights into changes in patient mobility. This capability not only demonstrates the potential of AI to adapt to and predict the needs of patients based on real-time data but also highlights its value in managing conditions that affect mobility and stability, enhancing the precision of therapy and potentially improving patient outcomes [71]. Furthermore, another recent study described neural network models specifically designed to capture the dynamic properties of gait, considering both short-term and long-term dependencies in gait cycles [72]. This approach allowed for the identification of individual-specific gait signatures, which are particularly useful in understanding and accommodating the unique gait dynamics of survivors of stroke, who often exhibit significant variability in their walking patterns due to neurological impairments [72]. The personalization of AI models to adapt to individual patient profiles over time promises to enhance the precision of interventions, although as AI models evolve, there is also a growing need for explainable AI that can make AI decisions more transparent and understandable to clinicians. This will be crucial for building trust and facilitating the integration of AI tools into everyday clinical practice. Future advancements should aim to not only refine the predictive accuracy of fall risk but also ensure that the outcome data are accessible and interpretable for all users involved in the care process. A recent study demonstrated the use of wearable sensors coupled with explainable AI techniques to provide detailed insights into muscle activity changes, aiding clinicians in understanding the predictive outcomes of AI models [73].

Detecting fall risk in individuals with neurological disorders has a primary role in the rehabilitation context. In this review, we identified several biomechanical parameters related to fall risk, as measured by wearable sensors. Data extracted from sensors can provide valuable insights into motor deficits associated with fall risk and guide the development of personalized fall prevention programs. Of note, conventional clinical scales may be not able to capture specific fall-related movements detected by wearable sensors [74]. This discrepancy highlights the importance of body-worn sensors, combining daily life sensor data with clinical assessment outcomes to enhance fall prediction accuracy. Therefore, daily life data recording should be preferred over in-laboratory data collection, as individuals with neurological disorders are more susceptible to the effects of test instructions or external distractions in controlled settings. However, lower activity levels in individuals with neurological disorders may affect the quality of the data collected. Moreover, both the location and method of sensor attachment are crucial when collecting sensor-based data from individuals with neurological disorders. Using multiple sensors might provide more detailed information but is less practical for this patient group. In addition, researchers or trained staff need to be present to ensure the correct placement and wearing of the sensors. From a practical standpoint, the sensor’s location should be chosen carefully to minimize discomfort during the assessment and avoid interfering with natural movement. Technical considerations, such as battery life, data transmission, and storage capacity, must also be factored in when selecting an appropriate sensor for research or clinical practice.

Future studies should incorporate additional fall risk factors (ie, cognitive status, comorbidities, and sensory functions) that could improve the accuracy of fall risk prediction. They should also consider using multiple sensors to capture more detailed biomechanical information. Innovative technologies such as wearable sensors hold promise for enhancing motion analysis and fall risk prediction due to their increasing simplicity, speed, and ease of interpretation compared to traditional methods. However, there remains a scarcity of studies with practical clinical applications for gait analysis in populations with neurological disorders, particularly in patients with stroke or AD. This gap may be attributed to practical challenges, such as the lack of widely available specialized devices in hospital settings and economic constraints. In addition, the absence of direct collaboration between clinicians and bioengineering professionals contributes to this issue. Increased collaboration between physiotherapists, kinesiologists, and bioengineers is essential for sharing practical insights that inform the development of specific motor training programs based on functional assessments. Technological advancements could address these challenges by promoting the use of smart tools, such as smartphones, which are becoming more prevalent in clinical settings, as supported by some studies [13,75,76]. In parallel, advancements in wearable sensor technology are being complemented by the rapid evolution of AI, particularly through its integration with big data, cloud computing, and Internet of Things platforms. Wearable sensors stand to benefit significantly from these advancements, as AI algorithms enable the efficient processing of large and complex datasets; the extraction of informative features; and the generation of accurate, data-driven decisions. These capabilities extend to a range of functions, including gesture recognition, movement analysis, object detection, and continuous health monitoring [77]. Moreover, the integration of multimodal sensors to collect various data types such as motion, physiological, and environmental signals offers promising opportunities for enhancing fall risk assessment. While traditional data fusion methods often struggle with the complexity and heterogeneity of such data, DL provides powerful tools to automatically extract and model relationships across multiple sensor modalities [78].

This systematic review has some limitations that need to be acknowledged. One limitation is the absence of quantitative analysis. In particular, we found considerable heterogeneity among the included studies in terms of sensors used, sensor placement, and biomechanical features extracted; therefore, conducting a quantitative analysis was not feasible. This limitation underscores the need for caution in generalizing the findings and emphasizes the importance of interpreting the results within the context of individual study characteristics. The selected studies also presented other limitations, including small sample sizes, lack of a control group or normative data, and lack of long-term follow-up evaluations. Future large-scale longitudinal studies are warranted to further explore and validate key aspects of fall risk assessment and fall detection in patients with neurological disorders.

Despite these limitations, our review provides a comprehensive qualitative synthesis that systematically summarized and interpreted the findings of individual studies to elucidate common themes, patterns, and discrepancies in the literature. As such, it offers valuable insights into the biomechanical features and wearable sensors used to perform fall risk assessment, identifying key implications for clinical practice and considerations for future investigation.

In this systematic review, we highlighted the critical role of wearable technologies in providing an objective and quantitative assessment of fall risk in patients with neurological disorders. Although the included studies were well executed, we identified significant heterogeneity in sensor placement, types of wearable systems, motor tasks, and evaluation methods. In addition, we observed a tendency among the included studies to primarily focus on technological and engineering aspects, often at the expense of the clinical context. However, the absence of a standardized fall risk assessment using wearable systems is likely attributable to ongoing technological advancements, which can be considered a strength.