Parkinson disease (PD) is a common neurodegenerative disorder characterized by tremor, bradykinesia, rigidity, and postural or gait disturbances [1,2]. Its high prevalence and disability rates profoundly impair patients’ quality of life and impose substantial burdens on families and health care systems. Since 1990, the number of patients with PD has more than doubled—from approximately 2.8 million to over 6.2 million—and is projected to exceed 12 million by 2040 [3]. In China, accelerated population aging has led to a rising prevalence, with an estimated 1.37% among adults aged 60 years and older and more than 3.62 million patients nationwide [4,5]. Rehabilitation remains essential for improving motor function and slowing functional decline [6-8]. However, conventional face-to-face rehabilitation is time- and resource-intensive, often inaccessible in remote regions, and frequently associated with poor adherence [9,10].

In recent years, the rapid development and application of wearable devices have provided significant advantages for assessment and intervention in neurological and psychiatric disorders, such as stroke and depression [11,12]. Wearable devices encompass a wide range of intelligent electronic devices, including smartwatches, wristbands, insoles, exoskeletons, multisensor wearable devices, and electronic textiles [13-16]. These devices support continuous, real-time, and multidimensional monitoring during daily activities or rehabilitation training, capturing key motor indicators such as gait, tremor, joint mobility, postural control, motor coordination, and overall physical activity levels [17-19].

Against this backdrop, the rapid emergence of artificial intelligence (AI)–enabled wearable devices has created new opportunities for the rehabilitation of PD. In this review, AI-enabled wearable devices refer to noninvasive wearable technologies that incorporate machine learning (ML) or deep learning (DL) methods for data-driven analysis beyond rule-based approaches. Unlike traditional approaches that rely on manual observation or intermittent scale-based assessments, AI-enabled wearable devices enable the capture of large volumes of fine-grained data. Supported by ML and DL techniques, these data can be efficiently processed and intelligently interpreted within a data-driven rehabilitation framework, highlighting the importance of AI model interpretability for clinical and nursing decision-making [20-22]. Moreover, AI-enabled wearable devices facilitate continuous remote monitoring and long-term follow-up, making it possible to track disease progression and changes in motor function in real time within home-based rehabilitation settings. These data can then be leveraged by clinicians and rehabilitation teams to design more precise training programs [23,24]. Such an approach enhances the continuity and personalized rehabilitation, offers a feasible pathway for remote and home-based rehabilitation, and holds promising potential for optimizing health care resource allocation and improving patient adherence.

Although numerous studies and reviews have explored the use of wearable devices in PD rehabilitation, several limitations remain, including (1) most studies involve small sample sizes and short follow-up periods, providing insufficient longitudinal evidence to evaluate long-term rehabilitation outcomes [17,25,26]; (2) current reviews mainly focus on device performance or symptom monitoring, with limited attention to AI data-processing workflows, algorithmic applications, and their potential value for clinical and nursing practice [18,19]; (3) studies differ substantially in device types, monitoring indicators, data-processing strategies, and AI algorithm selection. The absence of a unifying framework to synthesize these variations hinders a comprehensive understanding of the research landscape [27-29]; and (4) evidence in nursing practice remains limited; despite nurses’ central role in rehabilitation, research on AI-enabled tools for nursing decision-making, health education, and adherence support is still scarce [30,31]. Collectively, these issues result in fragmented and unsystematic evidence, limiting its ability to guide clinical rehabilitation nursing practice and theoretical development.

To address these needs, we conducted a scoping review and developed an evidence map to systematically present the current applications of AI-enabled wearable devices in PD rehabilitation and motor function assessment. The scoping review summarized the breadth, depth, and nature of existing evidence [32], while the evidence map visualizes research distribution, patterns, and emerging trends across domains [33]. By integrating these 2 approaches, this study aimed to map and visualize the current evidence on AI-enabled wearable devices for PD rehabilitation and motor function assessment. The objectives were to summarize device types, monitoring indicators, AI algorithms, and application characteristics, and to identify research gaps and future directions to support nursing practice and further digital-health implementation in PD rehabilitation.

The review followed the methodological guidance outlined in the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [34], and the completed checklist is provided in Multimedia Appendix 1. No protocol was registered for this review.

The inclusion criteria were defined using the population, concept, and context framework. Eligible studies enrolled participants with a confirmed diagnosis of PD and used noninvasive, body-worn wearable devices integrated with AI-based analytical methods. The focus of included studies was on rehabilitation, motor function assessment, or monitoring. Gray literature such as dissertations and full conference papers was included, whereas preprints and conference abstracts were excluded to ensure methodological consistency and to prioritize evidence that had undergone formal peer review. Only studies published in English or Chinese on or after January 1, 2020, were eligible for inclusion. Detailed inclusion and exclusion criteria are presented in Multimedia Appendix 2.

A comprehensive search was conducted on August 12, 2025, across 9 databases: CNKI (China National Knowledge Infrastructure), Wanfang Data, SinoMed, Cochrane Library, PubMed, Web of Science, CINAHL, Scopus, and Embase. The literature coverage spanned from database inception to the search date. No websites or other nondatabase information sources were searched. Reference lists and forward citations of the included studies were additionally screened to identify further potentially relevant records. To ensure the currency and completeness of the evidence, an updated search using the same strategy was conducted on December 10, 2025.

The reporting of the search process adhered to the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) [35]. The search strategy was developed by a researcher trained in systematic literature retrieval and combined MeSH (Medical Subject Headings) terms with free-text keywords across four core concepts: (1) wearable or body-worn mobile devices; (2) rehabilitation, assessment, or monitoring; (3) PD; and (4) AI or ML. No limits or restrictions, including language, publication date, or study design, were applied at the search level. All eligibility criteria were applied manually during title/abstract screening and full-text review. The search strategy did not undergo peer review. Complete search strategies for each database and information source are provided in Multimedia Appendix 3.

The selection process comprised two stages: (1) title/abstract screening and (2) full-text review. In the first stage, 1 reviewer imported all records into EndNote X9 (version 12062; Clarivate), removed duplicates, and excluded studies outside the eligible date range. Two reviewers then independently screened titles and abstracts to determine provisional inclusion. In the second stage, the same 2 reviewers independently assessed the full texts to establish the final set of included studies. Disagreements were resolved through discussion; unresolved cases were adjudicated by a third reviewer. Cohen κ was calculated to measure the interrater agreement [36]. Agreement was high for title and abstract screening (κ=0.86), whereas it was moderate for full-text review (κ=0.57).

Data charting was conducted independently by 1 reviewer using a standardized extraction form developed a priori. The form included key study characteristics, wearable device details, AI methods, outcomes, and findings relevant to the review questions. Following Joanna Briggs Institute guidance, charting was iterative, and the form was refined as needed. A second reviewer verified all data, with disagreements resolved by a third reviewer. The form was pilot-tested on 5 studies before full implementation.

The data extraction form was developed by the research team based on the objectives of this review and included three core modules: (1) study details, (2) wearable device details, and (3) AI details. When the relevant information was not reported, it was recorded as not available. Each module contained specific data variables, as detailed in Multimedia Appendix 4.

The methodological quality of all included studies was evaluated using the Mixed Methods Appraisal Tool, 2018 version [37]. Two reviewers conducted the assessments independently and resolved discrepancies through discussion; cases in which consensus could not be reached were adjudicated by a third reviewer. The appraisal was used descriptively to identify methodological strengths and limitations of the included studies and was not used as a criterion for study inclusion. The detailed appraisal results are provided in Multimedia Appendix 5 [38-103].

A narrative synthesis approach was used. Extracted data were summarized thematically and presented in descriptive text, tables, and an evidence map to illustrate research trends, methodological characteristics, and application domains of AI-enabled wearable devices in PD.

This scoping review used only published, peer-reviewed literature and therefore did not require approval from an ethics committee or institutional review board. Findings will be submitted to an open access, peer-reviewed journal and presented at relevant medical and engineering conferences.

A total of 1467 records were initially retrieved. Of these, 475 (32.38%) were removed using EndNote X9 due to duplication, publication date, or language incompatibility. The remaining 992 (67.62%) records underwent title and abstract screening, during which 877 (88.41%) were excluded. Among the 115 records reviewed in full text, 47 (40.87%) were unrelated to AI, 7 (6.09%) were unavailable in full text, and 2 (1.74%) lacked extractable data. An additional 7 relevant studies were identified through reference tracing. In total, 66 studies were included in this review [38-103]. The detailed selection process is illustrated in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram (Figure 1).

All included studies were published between 2020 and 2025, and the publication volume remained relatively stable over this 6-year period. The highest number of studies was published in 2025 (16/66, 24.24%; Table 1). The included studies originated from multiple countries and regions (Figure 2), with Switzerland contributing the largest proportion (27/66, 40.91%). With respect to publication type, the vast majority were journal articles (61/66, 92.42%).

Regarding study populations, sample sizes ranged from 2 to 1057 participants, with a mean sample size of 55.06 (SD 132.78), yielding a total of approximately 3579 participants. Overall, most studies involved relatively small samples with substantial heterogeneity, and the majority included only patients with PD (47/66, 71.21%). In terms of the application objectives of wearable devices, most studies focused on motor function assessment (54/66, 81.82%), whereas rehabilitation-oriented applications were relatively scarce, being reported in only 8 (12.12%) studies. Overall, this distribution suggests that intervention-focused research remains limited. Detailed characteristics of the included studies are presented in Table 1 and Multimedia Appendix 6 [38-103]. For categories in which studies could contribute to more than 1 classification, percentages may exceed 100%; detailed coding rules are provided in the table footnotes.

A total of 5 major types of wearable devices were identified across the included studies. Sensor module was the predominant device type, accounting for 84.85% (56/66) of studies, followed by smart insoles (7/66, 10.61%) and smartwatches (5/66, 7.58%; Table 2). Regarding device origin, commercially available devices were more common (41/66, 62.12%) than noncommercial devices (21/66, 31.82%); nevertheless, a substantial proportion of studies still relied on customized or research-grade devices. A total of 14 distinct wearing locations were reported (Figure 3). The most common placements were the shank (25/66, 37.88%) and the wrist (23/66, 34.85%), followed by the foot (19/66, 28.79%) and the ankle (16/66, 24.24%). Many studies used multiple wearing locations within a single study, reflecting the diversity of measurement strategies adopted.

With regard to sensor configuration, devices integrating multiple sensors predominated, accounting for 75.76% (50/66) of the included studies. The collected biosignals were predominantly inertial signals. Accelerometer data were acquired in the vast majority of studies (61/66, 92.42%), followed by gyroscope data (50/66, 75.76%), whereas magnetic field, pressure, and electromyography signals were used less frequently (Table 2).

With respect to application settings, wearable devices were more frequently deployed in clinical environments (31/66, 46.97%), with an additional 34.85% (23/66) of studies spanning both clinical and home environments, whereas studies conducted exclusively in home settings were relatively uncommon (12/66, 18.18%). Regarding the duration of monitoring or intervention, most studies used either single-session assessments (32/66, 48.48%) or repeated short-term testing designs (20/66, 30.30%), while studies involving long-term continuous monitoring were comparatively limited (9/66, 13.64%). Overall, wearable device research in PD is dominated by sensor module devices, with devices most commonly worn on the lower limbs and wrist. Data collection is primarily passive, and studies largely focus on short-term applications conducted in clinical settings. Additional technical characteristics of the wearable devices are provided in Table 2 and Multimedia Appendix 7 [38-103].

The AI applications in the included studies were categorized into four purposes: (1) monitoring or assessment (35/66, 53.03%), (2) state recognition or functional screening (17/66, 25.76%), (3) prediction (8/66, 12.12%), and (4) rehabilitation and feedback (6/66, 9.09%; Table 3). Among the studies, 62.12% (41/66) used ML algorithms only, 28.79% (19/66) used DL algorithms only, and 9.09% (6/66) combined ML and DL approaches. These studies used algorithms to address classification problems (59/66, 89.39%), regression problems (15/66, 22.73%), and clustering problems (3/66, 4.55%).

A total of 30 distinct AI algorithms were identified across the included studies. The most frequently used were support vector machines (SVMs; 28/66, 42.42%), followed by convolutional neural networks (CNNs; 23/66, 34.85%), random forests (19/66, 28.79%), logistic regression (15/66, 22.73%), decision trees (14/66, 21.21%), long short-term memory (LSTM) networks (11/66, 16.67%), and multilayer perceptrons (8/66, 12.12%; Table 3).

As shown in Figure 4, SVMs and CNNs were predominantly applied in the “monitoring or assessment” and “state recognition or functional screening” categories. Random forests, decision trees, and logistic regression models were distributed across application purposes but were mainly used for monitoring- or assessment-related tasks. Temporal and probabilistic models, such as LSTM networks and hidden Markov models, were more often observed in state recognition or functional screening tasks, although their overall adoption remained limited. In contrast, algorithmic applications in the “prediction” and “rehabilitation or feedback” categories were comparatively sparse, indicating that prospective and closed-loop applications remain at an early stage of development.

A total of 4 different model validation approaches were used across the included studies, with approximately 31.82% (21/66) using more than 1 method (Table 3). Leave-one-out cross-validation was the most frequently used approach (37/66, 56.06%), followed by k-fold cross-validation (27/66, 40.91%), hold-out validation (18/66, 27.27%), and external validation (4/66, 6.06%). A total of 25 performance metrics were used to evaluate model performance. The most commonly reported metrics were sensitivity (43/66, 65.15%), accuracy (41/66, 62.12%), specificity (27/66, 40.91%), F₁-score (27/66, 40.91%), area under the curve (25/66, 37.88%), precision (24/66, 36.36%), and correlation coefficient (14/66, 21.21%). Detailed characteristics of the AI algorithms applied in each study are provided in Multimedia Appendix 8 [38-103].

Based on a synthesis of the included evidence, we developed a conceptually integrated “challenges and opportunities” framework grounded in the evidence base. This framework categorizes the challenges and opportunities of AI-enabled wearable devices in clinical applications into six dimensions: (1) evidence and data quality; (2) technical limitations; (3) usability, adherence, and equity; (4) economic and policy barriers; (5) privacy, security, and data governance; and (6) clinical translation and workflow integration. The framework is intended to inform future research and practice rather than to represent an empirically validated implementation pathway, as illustrated in Figure 5 and detailed in Multimedia Appendix 9.

This study has several limitations. First, no prospective protocol was registered; however, this is an acceptable practice for scoping reviews, and we clarify this point to enhance methodological transparency. Second, although a systematic search and screening process was conducted, the scope of databases, search strategies, and inclusion criteria may have led to the omission of eligible studies, posing a risk of incompleteness or potential bias. In addition, most included studies had small sample sizes, single-center designs, short follow-up durations, and limited external validation, alongside substantial heterogeneity across study populations, task paradigms, signal sources, algorithmic approaches, and outcome measures. This heterogeneity may limit cross-study comparability and the generalizability of findings. Given the predominantly exploratory and proof-of-concept nature of the included studies, the closed-loop rehabilitation framework and related recommendations proposed in this review should be interpreted as conceptual and practice-informing rather than empirically validated implementation pathways. Finally, the search was restricted to English and Chinese peer-reviewed publications, with limited inclusion of gray literature or other languages, which may introduce time-lag, linguistic, and publication biases. In particular, studies reporting null or negative findings, early-stage failures, or limited model performance may be less likely to be published and therefore underrepresented, potentially leading to an overestimation of the maturity, effectiveness, and translational readiness of AI-enabled wearable technologies.

This study systematically summarized recent evidence on AI-enabled wearable devices for motor function assessment and rehabilitation in people with PD. By combining a scoping review with an evidence map and adopting a rehabilitation-oriented perspective, we synthesized evidence across multiple dimensions, including application aims, sensing modalities, AI methods, and validation practices, thereby identifying key translational gaps between proof-of-concept studies and real-world rehabilitation workflows. Compared with previous reviews that primarily focused on monitoring functions or device performance, this study places greater emphasis on rehabilitation applications and nurse-led translation into practice, and further clarifies the key actionable gaps that limit real-world implementation. On this basis, a conceptually integrated “challenges and opportunities” framework is proposed to inform the design, evaluation, and reporting of devices and algorithms, and to highlight considerations for developing workflow-integrated, decision-support wearable systems. From a real-world perspective, these findings may support continuity of rehabilitation across clinical, home, and community settings by enabling nurses to deliver continuous monitoring, personalized follow-up, and timely intervention, ultimately improving the efficiency and accessibility of PD rehabilitation management.