Perception, assessment, and coaching: a systematic review and taxonomy of computer vision-based physical rehabilitation techniques

Abstract

The digital transformation of rehabilitation training has become a public health imperative driven by a global demand that outstrips professional medical resources and is compounded by a deficit in public rehabilitation literacy. As traditional hospital-centric models reach their scalability limits, there is a critical necessity for accessible home-based care solutions to ensure patients do not miss optimal recovery windows. To evaluate computer vision as a potential solution, this paper conducts a systematic review following the PRISMA 2020 reporting framework. We identify that its clinical migration faces a profound “Paradigm Gap” across three critical domains which this study aims to address: (1) the Perception Domain, where algorithms are constrained by inherent reconstruction ambiguities and pathological data scarcity; (2) the Assessment Domain, where a “semantic gap” persists between low-level features and clinical reasoning; and (3) the Coaching Domain, where feedback mechanisms fail to translate summative “Knowledge of Results (KR)” into actionable “Knowledge of Performance (KP).” We formulate and adopt “Perception, Assessment, and Coaching (PAC)” as a novel taxonomy to serve as a logical grid for systematically analyzing existing literature and elucidating technical pathways required to resolve these clinical challenges. This review synthesizes the technological landscape into three evolutionary trajectories: (1) In Perception, research is shifting toward constructing biomechanically consistent digital twins to eliminate visual hallucinations. (2) In Assessment, frontier methods are establishing interpretable clinical reasoning engines to achieve a leap from engineering parameters to Evidence-Based Medicine (EBM) evidence. (3) In Coaching, focus lies in precise movement correction via semantic translation and multimodal strategies to support motor relearning. Furthermore, we explore the potential of Generative AI and Multimodal Large Language Models (MLLMs) in reshaping interaction paradigms (e.g., Visual Self-Modeling) alongside critical discussions on ethical boundaries. Through a systematic literature review, this paper elucidates the task boundaries of rehabilitation vision and constructs the PAC taxonomy. It provides a robust theoretical roadmap and forward-looking guidance for the design of the next generation of clinically valid intelligent rehabilitation systems.

1 Introduction

1.1 Background

The profound transformation of global demographics, driven by an unprecedented wave of aging, is rapidly reshaping the public health landscape. Consequently, Rehabilitation Medicine has evolved from a supplementary service into a fundamental pillar of the global health system. As epidemiological characteristics shift, the medical community faces multifaceted challenges. On one hand, the aging society has significantly exacerbated the disease burden of neurodegenerative disorders such as stroke and Parkinson’s disease. On the other hand, the incidence of sports injuries [e.g., Anterior Cruciate Ligament (ACL) reconstruction] and chronic musculoskeletal disorders (e.g., Osteoarthritis) is surging across all age demographics (1).

Data from the Global Burden of Disease Study 2021 indicate that approximately 2.4 billion individuals worldwide are in urgent need of rehabilitation interventions (2, 3). However, confronted with this exponentially growing population suffering from generalized motor dysfunction, existing healthcare systems are constrained by a dual deficit of “Resources and Cognition” (4, 5). On the supply side, there is a severe, inelastic shortage of professional Physical Therapists (PTs), making the traditional “one-on-one” manual supervision model unsustainable. On the demand side, the public generally lacks scientific “Rehabilitation Literacy,” often conflating rehabilitation with simple mechanical repetition while overlooking the critical roles of neural control and movement quality. The superposition of resource scarcity and cognitive bias causes a vast number of patients to miss optimal recovery windows, confronting the traditional hospital-centric model with a severe “Scalability Bottleneck.”

1.2 Challenges and bottlenecks

To mitigate this public health crisis, the extension of rehabilitation scenarios from clinical settings to the home (Home-based Rehabilitation) has emerged as an inevitable trend. However, mainstream Telerehab solutions remain at a rudimentary level, primarily relying on “video calls” or passive “video watching.” This “supervision vacuum,” detached from professional oversight, renders patients highly susceptible to compensatory injuries caused by incorrect movements and leads to high attrition rates due to the monotony of training (6, 7). Confronted with this dilemma, there is an urgent need for intelligent rehabilitation systems capable of digitizing clinical expertise to achieve scalable and precise guidance.

In recent years, driven by breakthroughs in deep learning, markerless pose estimation technology based on ubiquitous RGB video has provided a novel opportunity for automated home rehabilitation guidance, owing to its low cost and high accessibility. While recent review articles have extensively covered general Human Pose Estimation (HPE) architectures (8, 9) and multi-view reconstruction techniques (10), these works predominantly focus on engineering metrics such as mAP or MPJPE on standard datasets. Although some recent clinical benchmarking studies (11) have attempted to evaluate the angular accuracy of visual systems using Inertial Measurement Units (IMUs) as a reference, most existing research treats perception, assessment, and coaching as fragmented modules. There is a distinct lack of a unified framework that organically integrates biomechanical constraints with motor learning theories. It is precisely due to this misalignment between engineering objectives and clinical requirements that, despite significant progress in computer vision for general Human Action Recognition (HAR), migrating these technologies to rigorous medical rehabilitation scenarios still encounters a profound “Paradigm Gap.” The technological evolution of rehabilitation training systems and the transition from manual supervision to vision-based and generative paradigms are summarized in Figure 1.

Figure 1

1.3 The paradigm gap

1.4 Primary work

1.5 Future outlook & contributions

Building upon this analytical model, this paper further envisions the frontier transformations driven by Artificial Intelligence. With the rapid advancement of Generative AI and Multimodal Large Language Models (MLLMs) (30), these technologies are beginning to permeate and reshape every component of the aforementioned taxonomy. In the latter half of this paper, based on the evolutionary trends in existing literature, we critically explore the application prospects of foundation models in the rehabilitation domain. We analyze how to leverage their powerful semantic reasoning capabilities to construct the system’s “Cognitive Hub,” thereby resolving the interpretability challenges of traditional models. Furthermore, we prospectively discuss the theoretical value and potential risks of Generative AI-driven “Visual Self-Modeling (VSM)” technology in future interaction paradigms.

2 Materials and methods

To ensure a comprehensive, transparent, and reproducible synthesis of the rapidly evolving landscape of vision-based rehabilitation, this review followed a systematic screening protocol grounded in the PRISMA 2020 (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. The methodology was specifically designed to address the “Paradigm Gap” identified in the Introduction by intersecting engineering feasibility with clinical validity. In response to the methodological requirements of a systematic review, the research question, eligibility criteria, search strategy, evidence mapping, and methodological quality appraisal were further structured using a PICOS-informed framework and a PAC-based evidence synthesis strategy.

This review was not prospectively registered in PROSPERO. At the time of study design, the review was conceived as a multidisciplinary taxonomic and evidence-mapping review of computer vision-based rehabilitation systems, integrating engineering studies, clinical validation studies, rehabilitation prototypes, conference proceedings, reviews, and emerging AI-oriented works, rather than as a meta-analysis of a single clinical intervention or therapeutic effect. Nevertheless, the review was conducted according to a predefined search, screening, extraction, and evidence-mapping plan following the PRISMA 2020 reporting framework. The absence of prospective registration is acknowledged as a methodological limitation.

2.1 Research question and PICOS-informed framework

This review was guided by a PICOS-informed framework to clarify the systematic objectives, selection criteria, and scope of evidence synthesis. The overarching research question was: How have computer vision-based techniques been developed, validated, and translated across the Perception, Assessment, and Coaching stages of physical rehabilitation, and what methodological, clinical, and implementation gaps remain for their routine use in rehabilitation practice?

2.2 Search strategy and information sources

• Domain A (Perception): (“Human Pose Estimation” OR “3D Mesh Recovery” OR “Skeleton tracking”) AND (“Biomechanical Constraints” OR “Digital Twin” OR “SMPL” OR “Anatomical Prior”).

• Domain B (Assessment): (“Action Quality Assessment” OR “AQA” OR “Movement evaluation”) AND (“Pathological Synergy” OR “Clinical Metrics” OR “Explainable AI” OR “EBM”).

• Domain C (Coaching): (“Motor Learning” OR “Biofeedback” OR “Instructional feedback”) AND (“Multimodal Large Language Models” OR “MLLM” OR “Visual Self-Modeling” OR “Generative AI”).

(“physical rehabilitation” OR “rehabilitation exercise” OR “stroke rehabilitation” OR “orthopedic rehabilitation” OR “home-based rehabilitation” OR “telerehabilitation”) AND (“computer vision” OR “human pose estimation” OR “markerless motion capture” OR “skeleton tracking” OR “3D pose estimation” OR “action quality assessment”) AND (“movement assessment” OR “clinical validation” OR “biofeedback” OR “coaching” OR “motor learning” OR “exercise feedback”).

This general structure was adapted to the controlled vocabulary and syntax requirements of each database. For example, PubMed/MEDLINE searches incorporated medical and rehabilitation terms, whereas IEEE Xplore and ACM Digital Library searches emphasized pose estimation, skeleton tracking, action quality assessment, and human-computer interaction terms. Web of Science and Scopus were used to capture multidisciplinary studies across engineering, rehabilitation medicine, digital health, and applied artificial intelligence.

All arXiv records and other non-peer-reviewed preprints were treated as frontier or emerging evidence. They were included only when they were directly relevant to rapidly evolving technical directions, such as Generative AI, MLLMs, foundation models, or frontier computer vision methods, and they were not weighted as equivalent to peer-reviewed clinical validation studies. Their evidence status was explicitly recorded during methodological appraisal and is reported in Supplementary Data Sheet 1, Table S2.

2.3 Inclusion and exclusion criteria

To ensure clinical fidelity and technical robustness, the following criteria were applied:

Inclusion Criteria: Studies were eligible if they met at least one core requirement within the PICOS-informed scope and contributed directly to the PAC framework. Specifically, eligible studies included: (1) peer-reviewed journal articles or top-tier conference proceedings (e.g., CVPR, ICCV, PeerJ CS, TNSRE); (2) research presenting vision-based systems specifically for neurological, orthopedic, or geriatric rehabilitation (2, 3, 31); (3) studies providing quantitative validation against “Gold Standard” ground truth (e.g., VICON, IMU) or EBM-compliant metrics (22, 32–34); (4) frontier works investigating foundation models for clinical reasoning (35–38); and (5) studies that contributed to at least one of the three PAC domains, namely Perception, Assessment, or Coaching.

Exclusion Criteria: (1) General Human Action Recognition (HAR) studies focusing solely on action classification without quality analysis (8, 9); (2) research relying exclusively on wearable sensors without a primary computer vision component; (3) Metric-Utility Dissociation: purely algorithmic studies optimizing engineering benchmarks (e.g., MPJPE) on static datasets without biomechanical consistency; (4) Semantic Insufficiency: works providing only summative “Knowledge of Results” (KR) without error attribution (KP) (39, 40); (5) studies unrelated to physical rehabilitation, motor function assessment, or rehabilitation-oriented movement coaching; and (6) studies for which the full text or essential methodological information could not be retrieved.

Because this review integrates both engineering and clinical literature, studies were not excluded solely because they lacked randomized clinical trial evidence. However, such studies were differentiated during evidence appraisal to avoid assigning equivalent evidential weight to preliminary prototypes, engineering benchmarks, and clinically validated systems.

2.4 Study selection and data extraction

After duplicate removal, titles and abstracts were screened according to the predefined PICOS-informed eligibility criteria. Full-text articles were then assessed for relevance to computer vision-based physical rehabilitation and mapped to the PAC framework. Reasons for exclusion at the title/abstract and full-text stages were documented and summarized in the PRISMA flow diagram.

During revision, two reviewers independently re-checked the title/abstract screening decisions, full-text eligibility decisions, and data extraction records against the predefined PICOS-informed criteria. Disagreements were resolved through discussion, and a senior reviewer was consulted when consensus was required. Because formal inter-rater reliability statistics were not prospectively recorded during the original screening process, Cohen’s kappa was not calculated for the initial screening and eligibility phases. This absence of prospectively recorded inter-rater reliability statistics is acknowledged as a methodological limitation of the review.

For each included study, key information was extracted using a standardized data extraction form. Extracted items included: author and year, publication type, rehabilitation population, study design, vision modality, algorithmic method, PAC domain, comparator or reference standard, validation metric, clinical outcome or movement-quality indicator, feedback or coaching strategy, real-world deployment context, evidence status, and main limitations. This structured extraction process supported both descriptive synthesis and PAC-based evidence mapping.

2.5 Methodological quality appraisal and evidence grading

To address methodological heterogeneity and avoid discussing all publications as if they provided the same level of evidence, a customized methodological quality appraisal framework was developed for this review. Conventional tools such as QUADAS-2 or ROBINS-I are designed for specific diagnostic or observational study designs and are not fully suitable for a mixed corpus that includes clinical validation studies, engineering benchmarks, rehabilitation prototypes, conference proceedings, and emerging AI-oriented works. Therefore, a customized checklist was used to assess methodological robustness and evidence strength across the included studies.

Each record was classified according to PAC domain, use in review, rehabilitation population or context, study role, reference standard or validation metric, validation context, evidence status, and overall appraisal category. The overall appraisal categories included High, Moderate to high, Moderate/technical, Secondary evidence, Emerging, and Contextual. The complete methodological quality appraisal and evidence weighting are provided in Supplementary Data Sheet 1, Table S2. Studies were not excluded solely on the basis of quality appraisal because the purpose of this review was to map the technological and clinical landscape of vision-based rehabilitation. However, the appraisal results were used to contextualize the strength of evidence and to distinguish clinically validated systems from preliminary or conceptual works.

To avoid conflating heterogeneous evidence types, explicit evidence hierarchies and interpretation boundaries were applied during synthesis. Studies with direct patient-level validation, randomized or controlled designs, biomechanical or clinical reference standards, clinically relevant comparators, or therapist-rated outcomes were interpreted as stronger evidence for clinical readiness. Engineering benchmark studies were used primarily to discuss algorithmic feasibility, technical performance, and methodological limitations. Reviews and contextual/background papers were used to frame the field and support background interpretation, whereas protocols, preprints, benchmark-only studies, and frontier AI studies were treated as emerging or contextual evidence. These records were retained when relevant to the PAC taxonomy or future technical trajectories, but they were conservatively weighted and were not used as primary support for established clinical effectiveness or routine clinical deployment.

In particular, non-peer-reviewed preprints and rapidly evolving frontier AI studies were explicitly labeled as Emerging evidence in the appraisal process. They were used to contextualize future technical trajectories, especially in relation to MLLMs, Generative AI, and foundation-model-assisted rehabilitation, but were not used as primary evidence for clinical effectiveness or routine clinical readiness.

2.6 Selection results and data mapping

The systematic workflow is illustrated in Figure 2. The primary database search identified 427 records, while frontier tracing and Google Scholar snowballing added 31 high-impact records, totaling 458. After removing 130 duplicates and screening 328 titles/abstracts, 228 articles were assessed for full-text eligibility. Finally, 147 publications were included in the PAC-based evidence map and systematic synthesis. The included studies were subsequently mapped according to PAC domain, rehabilitation population, study design, validation method, evidence status, and evidence level. This mapping enabled the review to synthesize not only the technical evolution of computer vision-based rehabilitation but also the methodological robustness, peer-review status, and clinical relevance of the available evidence.

Figure 2

Compared with a purely narrative synthesis, this PAC-based evidence mapping allowed the included publications to be interpreted according to their functional role: Perception studies primarily addressed movement capture and biomechanical reconstruction; Assessment studies focused on movement-quality quantification, clinical metrics, and pathological attribution; and Coaching studies examined feedback delivery, motor learning, and patient-facing interaction strategies. The methodological quality appraisal was used alongside this mapping to avoid assigning equal evidential authority to engineering prototypes, preliminary feasibility studies, non-peer-reviewed preprints, and clinically validated systems.

To further reduce reliance on narrative synthesis alone, we summarized the included records using a descriptive quantitative evidence map. This summary reports the distribution of records across PAC domains, technology-related functional categories, rehabilitation populations or pathological contexts, evidence status, and overall appraisal levels. Because the included publications were highly heterogeneous in terms of study design, technology modality, comparator, population, and outcome metrics, formal meta-analysis was not appropriate. Instead, the descriptive evidence map was used to identify where the available evidence is concentrated and where important gaps remain (Table 1). The detailed record-level coding is provided in Supplementary Data Sheet 1, Table S2.

Overall, the descriptive evidence map indicates that the current literature remains concentrated in Perception-oriented visual measurement and reconstruction, whereas Assessment and especially Coaching studies are less represented. It also shows that many records remain general, benchmark-oriented, or not population-specific, and that high-level clinical validation evidence remains limited.

3 Clinical context and technical challenges

While computer vision technologies have reached a high level of maturity in general Human Action Recognition (HAR), migrating them to medical rehabilitation scenarios continues to confront a significant “Clinical Adaptability Gap.” The essence of rehabilitation transcends mere mechanical repetition; rather, it is a process of remodeling neuromuscular pathways under pathological constraints. To construct visual systems endowed with Clinical Validity, we must transcend pure algorithmic optimization and deeply deconstruct the authentic clinical context in which these algorithms operate. This section dissects the unique challenges facing rehabilitation vision from three core dimensions: pathological characteristics, quantification requirements, and environmental constraints. The mapping between clinical requirements, computer vision challenges, and validity criteria is summarized in Table 2.

3.1 Typical rehabilitation populations and kinematic characteristics

Unlike the pursuit of maximal strength or explosive power common in general fitness populations, the kinematic characteristics of rehabilitation patients are primarily defined by restricted joint Range of Motion (ROM), diminished motor control capabilities, and pathological synergy patterns. Existing vision-based rehabilitation research encompasses a full age spectrum ranging from neurodegenerative pathologies to musculoskeletal injuries.

3.1.1 Neurological rehabilitation: focusing on coordination and pathological synergy

3.1.2 Orthopedic rehabilitation: focusing on geometric alignment and weight bearing

3.1.3 Geriatric & balance rehabilitation: focusing on postural stability

3.2 From clinical scales to visual metrics: multidimensional quantification of movement quality and safety

The core of “Precision Coaching” lies in translating ambiguous clinical observations into computable mathematical metrics. Although existing clinical assessment systems [e.g., Fugl-Meyer Assessment (FMA), Berg Balance Scale (BBS)] have established a set of evaluation standards covering basic parameters such as gait speed, stride length, and symmetry, and computer vision has achieved broad coverage of these metrics in general motion analysis, mere basic kinematic parameters are often insufficient to reveal the pathological essence within the specific clinical context of rehabilitation. To construct a visual system with clinical diagnostic validity, this section focuses on three core and technically challenging dimensions—“Geometric Structural Integrity,” “Temporal Control Quality,” and “Pathological Compensation Mechanisms”—to deeply analyze the mapping gap between existing visual metrics and clinical gold standards.

3.2.1 Geometric metrics: geometric alignment and consistency

3.2.2 Temporal quality metrics: smoothness and jitter suppression

3.2.3 Compensation and safety: multi-label detection and attribution

Compensatory movements are mechanisms where patients utilize incorrect muscle groups to complete tasks (e.g., shoulder hiking instead of arm lifting).

3.3 Real-world constraints and domain gaps in home and remote scenarios

Transferring Computer Vision (CV) algorithms from controlled laboratory settings to authentic home-based rehabilitation scenarios (“In-the-wild”) confronts significant “Sim-to-Real” domain shifts. This has become a focal point of recent literature and represents the primary hurdle that the “Perception Domain” must resolve.

3.3.1 Environmental noise: uncontrolled backgrounds and occlusions

3.3.2 Interaction & viewpoint challenges: patient-centric design

3.3.3 Device & operational load: reducing cognitive barriers

4 A taxonomy and logical architecture for vision-based rehabilitation

To bridge the “Semantic Gap” identified in the Clinical Context and Technical Challenges between raw pixel-level data and high-level clinical decision-making, this review formulates a hierarchical analytical model and taxonomy—“Perception, Assessment, and Coaching (PAC)” (see Figure 3). Serving as a structural grid, this taxonomy delineates the systematic, end-to-end technical pipeline observed in recent literature, which aims to transform uncontrolled visual perception into interpretable reasoning and, ultimately, closed-loop clinical intervention. This section elucidates the underlying logic of this taxonomy, establishing a theoretical blueprint to systematically categorize and review the fragmented existing literature in subsequent sections. Positioning of the PAC taxonomy relative to existing frameworks. To clarify the originality and scope of the PAC taxonomy, Table 3 compares PAC with existing clinical, methodological, computer-vision, action-quality-assessment, motor-learning, and implementation-oriented frameworks. PAC is not intended to replace PRISMA, PICOS, ICF, conventional computer-vision pipelines, action-quality-assessment models, motor-learning theory, or implementation-oriented regulatory frameworks. Rather, it integrates these complementary perspectives into a domain-specific analytical structure for vision-based rehabilitation systems by linking visual measurement, clinical interpretation, and patient-facing feedback.

Figure 3

4.1 The perception domain: from geometric reconstruction to biomechanical digital twins

4.2 The assessment domain: from opaque mapping to clinical attribution reasoning

4.3 The coaching domain: closed-loop guidance and motor relearning

4.4 Operational application of the PAC taxonomy: example-based classification and evaluation

To avoid treating the PAC taxonomy as a purely conceptual framework, this review further operationalizes it as a practical classification and evaluation tool. As shown in Table 4, vision-based rehabilitation systems can be classified according to whether they perform only movement capture, provide clinically meaningful movement-quality assessment, or close the loop through patient-facing feedback and adaptive coaching. This operational mapping helps distinguish pose-estimation backbones (33, 70, 72, 80–82) from rehabilitation assessment systems (32, 42, 60, 85, 94, 95) and from full closed-loop coaching platforms (63, 68, 87, 88, 93, 96).

This table-based operationalization also clarifies how the PAC taxonomy can guide future system development. A clinically mature system should not only capture movement accurately, but should also provide clinically interpretable assessment, explicit error attribution, and safe feedback delivery within a realistic rehabilitation workflow. By contrast, systems that report only pose-estimation accuracy or benchmark-level performance should be interpreted as technical foundations rather than complete rehabilitation solutions (10, 12, 89, 97).

This operational use of the PAC taxonomy provides two benefits for evidence synthesis. First, it prevents overestimating the clinical maturity of systems that achieve high pose-estimation accuracy but do not report rehabilitation-specific metrics or clinical validation (10, 33, 80, 81). Second, it identifies the missing link in many current systems: although Perception and Assessment modules are increasingly mature, fewer studies complete the full loop from visual sensing to clinically grounded feedback and adaptive coaching (42, 49, 68, 85, 87, 88). Therefore, the PAC taxonomy can serve not only as a descriptive map of existing literature but also as a design checklist for future vision-based rehabilitation systems. A clinically mature system should ideally demonstrate robust perception, clinically valid assessment, interpretable error attribution, and safe feedback delivery within a realistic rehabilitation workflow (32, 60, 63, 68, 88, 95). This operationalization does not constitute formal clinical validation of the taxonomy, but it demonstrates how the PAC framework can be practically applied to classify existing systems, evaluate their clinical maturity, and guide the design of future rehabilitation solutions.

5 The perception domain

Functioning as the “eyes” of intelligent rehabilitation systems, the core mandate of research within the Perception domain transcends the mere pursuit of minimizing Mean Per Joint Position Error (MPJPE) benchmarks on standard datasets. Instead, the primary objective observed in recent literature is to construct high-fidelity digital twins endowed with Biomechanical Consistency within uncontrolled, in-the-wild home environments. To surmount the significant “Sim-to-Real” domain shift, current methodologies collectively form a robust four-stage pipeline.

5.1 Data acquisition and preprocessing: hardware evolution and environmental challenges

To provide a structured roadmap for surmounting these challenges, a hierarchical taxonomy of Human Pose Estimation (HPE) algorithms for rehabilitation is illustrated in Figure 4. This classification serves as the structural foundation for the subsequent technical discussion, categorizing the literature into three critical stages: (1) 2D coordinate extraction and domain adaptation (detailed in 2D Pose Estimation); (2) 2D-to-3D pose lifting (presented in 3D Pose Estimation and 2D-to-3D Lifting); and (3) the enforcement of physiological validity through biomechanical and physics-aware constraints (formalized in Physiological Validity).

Figure 4

5.2 2D pose estimation: architecture selection and domain adaptation challenges

5.3 3D pose estimation and 2D-to-3D lifting

Core clinical assessment metrics (e.g., Range of Motion, ROM) are inherently 3D spatial parameters. The perspective Foreshortening effect in 2D projections leads to severe measurement errors. Consequently, recovering 3D posture from 2D video (3D Pose Lifting) represents a critical leap “from image to data.”

5.3.1 Evolution of solving the ill-posed problem

5.3.2 Current state of clinical accuracy

Comparative studies using “gold-standard” optical motion capture systems, such as Vicon or OptiTrack, indicate that the Mean Per Joint Position Error (MPJPE) of state-of-the-art visual 3D pose estimation models in standard gait-analysis settings has been reduced to approximately 30–50 mm (33, 98). The HGcnMLP model by Hu et al. (64) reported knee joint angle measurements that were highly consistent with Vicon-based assessment in musculoskeletal gait analysis, with ICC values of 0.84–0.98 and angular errors of approximately . These findings suggest that vision-based pose estimation has potential utility for monitoring large-joint movements such as squats and gait. However, such agreement should not be interpreted as sufficient evidence of clinical validity across all rehabilitation populations, movement tasks, camera viewpoints, or home environments.

Importantly, MPJPE is an engineering accuracy metric rather than a direct clinical validity endpoint. A low MPJPE value on standard datasets does not necessarily guarantee accurate ROM estimation, reliable compensatory-movement detection, agreement with therapist ratings, sensitivity to pathological movement patterns, or robustness in home-based rehabilitation. Therefore, clinical interpretation requires additional validation against biomechanical or clinical reference standards, such as optical motion capture, IMUs, depth-camera systems, goniometry, clinical scales, therapist ratings, or expert annotations.

To clarify this distinction, Table 6 is intended as a technical paradigm table that summarizes the evolution from data-driven pose regression to biomechanically constrained digital-twin modeling, rather than as direct evidence that MPJPE improvements alone establish clinical effectiveness. To further address clinical validity, we added Table 5, which summarizes explicitly coded validation approaches, reference standards, and comparator availability identified within the Perception-domain subset. This table distinguishes records validated against clinical or biomechanical reference standards from records reporting only benchmark-level pose-estimation metrics or lacking an explicitly coded clinical comparator. Because the table is based on conservative record-level coding, low counts in some categories should be interpreted as evidence of limited explicitly reported validation within the coded Perception-domain subset, rather than as evidence that such technologies are absent from the broader rehabilitation literature. The detailed record-level appraisal for all 147 included publications is provided in Supplementary Data Sheet 1, Table S2.

Taken together, current evidence suggests that the clinical readiness of perception systems depends not only on lower pose-estimation error, but also on anatomical consistency, clinically interpretable kinematic outputs, reference-standard validation, and robustness under real-world rehabilitation conditions. This provides the empirical rationale for the subsequent discussion of biomechanical constraints and digital-twin modeling.

5.4 Mathematical formalization: the biomechanical digital twin

5.4.1 Structural consistency: ensuring rigid-body properties

5.4.2 Anatomical constraints: correcting joint violations

5.4.3 Physics-aware constraints: enforcing balance and contact

6 The assessment domain

Leveraging high-fidelity biomechanical digital twins, the primary mandate within the Assessment Domain is to transcend the “Semantic Gap” existing between low-level geometric representations and high-level clinical decision-making. As established in current literature, while general Human Action Recognition (HAR) focuses on identifying “action categories,” the clinical utility of rehabilitation relies on the Quantification of movement quality and the Attribution of pathological mechanisms. Consequently, advanced methodologies must surpass the constraints of opaque feature mapping inherent in traditional end-to-end models, establishing explainable reasoning pipelines aligned with the principles of Evidence-Based Medicine (EBM). This section synthesizes existing technical paradigms following the methodological trajectory of “Temporal Semantic Parsing Quality Quantification Pathological Attribution.”

6.1 Pre-assessment modeling: movement segmentation and phase identification

6.2 Computational paradigms for action quality assessment (AQA)

Research in Action Quality Assessment (AQA) seeks to map complex motor performances into quantifiable metrics. To overcome the lack of explicit decision-making evidence in early models, current literature is primarily bifurcated into two trajectories: Explicit Rule-based and Implicit Learning-based paradigms.

To clarify the applicable boundaries of these trajectories, Table 7 provides a taxonomic comparison of core logic, clinical suitability, and technical trade-offs. While explicit geometric methods offer superior interpretability, high-dimensional feature extraction becomes indispensable for addressing complex neurological features, such as pathological synergies.

6.2.1 Paradigm A: explicit geometric and rule-based methods

6.2.2 Paradigm B: implicit learning and manifold modeling

6.3 Pathological attribution reasoning and multi-label error localization

7 The coaching domain

Building upon high-fidelity perception and precision assessment paradigms, the Coaching Domain represents the pivotal interface for translating computational metrics into therapeutic actions. While the Perception and Assessment domains quantify motor states, the clinical value of rehabilitation technologies ultimately depends on whether these measurements can be converted into safe, interpretable, and patient-facing feedback. Current research guided by Motor Learning Theory explores how kinematic features can be transformed into biofeedback, augmented feedback, or adaptive coaching strategies to support motor relearning and adherence.

However, the Coaching domain should be interpreted more cautiously than the Perception and Assessment domains. Compared with pose-estimation accuracy studies or movement-quality assessment studies, direct clinical evidence for automated coaching remains relatively limited. Many existing works report feasibility, prototype-level feedback generation, case-series evidence, protocols, or evidence synthesized from secondary reviews rather than primary randomized or controlled clinical trials in rehabilitation populations. Therefore, this section synthesizes coaching strategies as promising and theory-informed approaches, while distinguishing them from definitively validated clinical interventions.

7.1 Content translation: the paradigm shift from knowledge of results (KR) to knowledge of performance (KP)

7.1.1 Strategies for semantic translation

7.1.2 Prioritization logic: the hierarchical feedback pyramid

7.2 Intervention strategy and timing: regulation of cognitive load

7.3 Multimodal augmentation: immersive interaction and motor-learning support

7.4 Evidence maturity and clinical interpretation of coaching strategies

Table 9 summarizes major coaching strategies according to their technical modality, theoretical basis, evidence maturity, and clinical interpretation. The table intentionally distinguishes primary clinical evidence from secondary reviews, protocols, prototype studies, and emerging AI-based feedback systems. This distinction is important because the current evidence base does not yet support treating all coaching strategies as equally validated clinical interventions.

Overall, the Coaching domain remains a critical but comparatively under-validated component of vision-based rehabilitation. Existing studies support the plausibility of semantic feedback, adaptive scheduling, sonification, AR/VR visualization, and gamified interaction, but the direct clinical evidence base remains less mature than that for movement capture and movement-quality assessment. Future research should prioritize randomized or controlled clinical studies that evaluate not only execution accuracy and kinematic improvement, but also retention, transfer to daily activities, safety, adherence, pain, satisfaction, quality of life, and functional independence. Such outcomes are essential for determining whether computationally sophisticated coaching systems translate into meaningful patient-centered rehabilitation benefits.

8 Discussion

The systematic analysis of existing literature reveals that vision-based rehabilitation is entering a transformative era. To fully realize the clinical utility of these technologies, we must examine both the evolutionary pathway of emerging computing paradigms and the persistent challenges blocking their practical deployment. This section discusses the future prospects of multimodal foundation models in rehabilitation alongside the clinical implementation, data ecology, regulatory, and ethical barriers that must be addressed to transition from academic prototypes to certified medical devices and routine clinical workflows.

8.1 Future prospects: reshaping rehabilitation computing via multimodal foundation models

The GenAI- and MLLM-related discussion in this section should be interpreted as an emerging and prospective research direction rather than as evidence of established clinical effectiveness. Current applications of foundation models in vision-based rehabilitation remain insufficiently validated in prospective rehabilitation trials. Therefore, claims regarding autonomous rehabilitation agents, generative visual feedback, Visual Self-Modeling, and MLLM-assisted coaching are framed as future hypotheses that require prospective clinical validation, safety evaluation, bias assessment, usability testing, and regulatory review before routine clinical deployment.

The paradigm of vision-based rehabilitation may be reshaped by the future integration of Multimodal Large Language Models (MLLMs) and Generative AI. Recent medical AI research suggests that foundation models may contribute to clinical decision support and interpretability, but their direct clinical utility in rehabilitation remains insufficiently validated. In this emerging and prospective paradigm, MLLMs may serve as a potential “Cognitive Hub” to support the transition from discriminative analysis to generative, patient-facing assistance. We therefore outline below a theoretical architecture of prospective autonomous rehabilitation agents, while emphasizing that these systems remain at an early translational stage.

To illustrate the theoretical necessity of this shift, we analyze the post-stroke movement “Hand-to-Opposite-Shoulder” as a conceptual archetype. This movement, involving complex joint decoupling and Pathological Synergy, serves as a suitable litmus test for evaluating the transition from geometric regression to semantic reasoning.

8.1.1 Reshaping perception: from geometric reconstruction to semantic intuition

Traditional rehabilitation AI has historically been confined to Geometric Reconstruction, outputting coordinates that often fail to capture the Functional Impairments underlying motor deficits (125, 126). In the era of foundation models, perception is evolving toward Intent-based Semantic Sensing. Leveraging the long-horizon temporal reasoning of modern MLLMs (127), future systems may be able to assist in estimating motor intent and characterizing the dynamic gap between intended and observed movement, although this capability remains to be validated in prospective rehabilitation studies.

For the “Hand-to-Opposite-Shoulder” task (59), the literature suggests a shift toward Zero-shot Parameter Estimation, where systems extract mission-critical features directly from RGB video based on semantic instructions. This allows for the Semantic Tagging of obstacles, identifying high-level functional barriers such as [Elbow Flexion Dominance] or [Ipsilateral Trunk Lean] that transcend simple joint angles (126).

To reconcile the tension between data scarcity and privacy mandates, future frameworks should adopt Federated Learning and Privacy-Preserving Analytics (e.g., GDPR/HIPAA-compliant pipelines), enabling collaborative training across institutional boundaries without raw data exposure.

8.1.2 Reshaping assessment: causal attribution of intent-reality conflict

The critical frontier in assessment is the transition from opaque scoring to a “Diagnostic Nexus” powered by Chain-of-Thought (CoT) reasoning. This paradigm focuses on resolving the Intent-Reality Conflict—the gap between a patient’s motor intent and their actual execution—by explaining why a movement fails to meet clinical targets through transparent diagnostic chains (52, 128). By articulating the underlying clinical logic, these frameworks provide physiotherapists with actionable insights that align with established Evidence-Based Medicine (EBM) pathways.

8.1.3 Reshaping guidance: towards generative sensorimotor loops

The evolution of the Coaching domain involves upgrading interaction from mere “error reporting” to “motor-learning support” within a closed sensorimotor loop.

8.1.3.1 Intelligent Decision-Making and Metaphorical Cueing

By leveraging MLLMs, future agents can generate instructions using an “External Focus” strategy, which has been proven to optimize motor performance (40). Examples observed in pilot studies include the use of metaphorical cues (e.g., “pulling a seatbelt”) to promote automated motor control and reduce cognitive load.

8.1.3.2 Generative Visual Biofeedback: Visual Self-Modeling (VSM)

8.2 Barriers to clinical translation: clinical implementation, data ecology, and ethical safety

While the proposed framework illustrates a clear evolutionary path for vision-based rehabilitation, transitioning these academic prototypes into approved Software as a Medical Device (SaMD) remains a formidable challenge. The primary hurdle lies in shifting the paradigm from “algorithmic accuracy” to “clinical fidelity.” We dissect below the root challenges frequently overlooked in current literature and discuss potential pathways for large-scale clinical adoption.

8.2.1 Clinical implementation and workflow integration

Beyond algorithmic accuracy, the transition of vision-based rehabilitation systems from experimental prototypes to routine clinical practice requires a broader implementation pathway. A first unresolved challenge is the lack of multicenter validation. Many existing systems have been evaluated in single-center, small-sample, or laboratory-controlled settings, whereas real-world deployment requires robustness across hospitals, rehabilitation protocols, therapist practices, camera positions, lighting conditions, home environments, and patient demographics (12, 33, 68, 132). Without such external validation, performance reported on controlled datasets may not translate into reliable clinical use.

A second challenge concerns generalization across pathological populations. Movement abnormalities differ substantially among stroke, Parkinson’s disease, osteoarthritis, sarcopenia, postoperative orthopedic rehabilitation, cerebral palsy, and low back pain populations. Models trained on healthy participants or mimicked impairments may fail to recognize disease-specific features such as abnormal synergy, tremor, spasticity, compensatory trunk movement, pain-avoidance strategies, or fatigue-related movement degradation (31, 60, 68, 94, 95, 129). Therefore, future studies should move beyond generic pose-estimation accuracy and evaluate whether computer vision outputs remain clinically meaningful across disease phenotypes, severity levels, and rehabilitation stages.

Integration into clinical workflows remains a practical bottleneck. For physiotherapists and rehabilitation physicians, an algorithmic score is insufficient unless it can be translated into clinically actionable information, such as the affected joint, movement phase, compensatory pattern, severity level, recommended correction, and longitudinal progress. Effective deployment therefore requires therapist-facing interfaces, electronic documentation compatibility, shared decision-making support, exercise prescription adjustment, and follow-up monitoring. Systems should reduce rather than increase clinician workload, particularly in settings already facing rehabilitation workforce shortages and limited access to therapy services (4, 5, 7, 24, 63). From the perspective of the PAC taxonomy, successful clinical translation requires not only robust Perception, but also clinically validated Assessment and safe, workflow-compatible Coaching.

8.2.2 Privacy, consent, and data protection in home video rehabilitation

Privacy and data protection represent critical barriers for camera-based rehabilitation. Unlike wearable sensors that primarily record inertial signals, vision-based systems may capture identifiable facial information, body shape, caregivers, household environments, and other sensitive contextual data. Even skeletonized trajectories may retain re-identification risks when combined with temporal movement patterns or demographic information. Privacy-preserving video processing, on-device inference, data minimization, secure transmission, anonymization, and transparent consent procedures should therefore be treated as core design requirements rather than optional technical features (24, 46, 97).

For home-based video rehabilitation, consent procedures should be more specific than those used for conventional clinical data collection. Participants should be informed whether the system performs continuous or repeated video capture, whether raw video, skeleton trajectories, or derived kinematic features will be stored, how long data will be retained, whether data may be reused for algorithm development, and how patients can withdraw consent or request data deletion. In addition, because home-based cameras may inadvertently capture caregivers, family members, or household environments, future systems should implement bystander-protection mechanisms, privacy zones, automatic face or background blurring, and clear policies for secondary data use.

8.2.3 Regulatory readiness and SaMD certification

Regulatory readiness is equally important for clinical translation. Systems that only estimate pose may function as low-risk measurement tools, but systems that generate clinical interpretations, detect abnormal movement patterns, recommend exercise modifications, or deliver automated coaching may approach the functional scope of Software as a Medical Device (SaMD). Such systems require evidence of analytical validity, clinical validity, usability, risk control, explainability, and post-deployment monitoring. This is particularly important for MLLM- or agent-assisted rehabilitation systems, which may generate fluent but unsafe exercise advice if hallucination control, uncertainty estimation, and clinician escalation mechanisms are not implemented (35, 37, 38, 49, 97).

More specific regulatory frameworks should also be considered. In the European Union, systems that provide clinical assessment, therapeutic recommendations, or adaptive exercise guidance may fall within the scope of the Medical Device Regulation (MDR 2017/745), and AI-enabled rehabilitation systems may also need to consider requirements related to the EU AI Act. In the United States, such systems may be evaluated within FDA digital health and Software as a Medical Device pathways. Accordingly, future vision-based rehabilitation systems should provide evidence for clinical performance evaluation, risk management, human factors and usability testing, cybersecurity, post-market surveillance, and change-management plans for adaptive or continuously updated algorithms. In addition to regulatory pathways, future systems should align, where applicable, with relevant medical-device software, safety, and usability standards, such as ISO 14971 for risk management, IEC 62304 for medical device software life-cycle processes, and IEC 62366-1 for usability engineering, depending on the intended use, risk classification, deployment setting, and jurisdiction.

8.2.4 Data ecology: from mimicked data to pathological digital phenotypes

8.2.4.1 Research perspective

To overcome the barriers identified in Table 10, we posit that future research must shift toward constructing Pathological Digital Phenotypes. By utilizing physics engines (e.g., MuJoCo) for Counterfactual Synthesis, researchers can simulate pathological states with varying parameters (e.g., joint stiffness). This Sim-to-Real strategy, when constrained by established biomechanical rules (134) and augmented by generative techniques to produce realistic synthetic postures (135), can help ensure that synthetic data remains anatomically valid while providing the high-volume knowledge required for foundation model training.

8.2.5 Clinical safety and ethics: biomechanical firewall and embodied trust

Given the scarcity of real-world data, Generative AI has become an essential tool for data augmentation and personalized feedback. However, this introduces the risk of “Medical Hallucinations.” If Generative Visual Biofeedback (VSM) deviates from physiological constraints (136), it may inadvertently induce patients to attempt anatomically impossible or hazardous movements.

8.2.5.1 Ethical guardrails

Based on the security analysis of current technologies, we propose the implementation of a “Biomechanical Firewall.” This involves reverse-embedding physiological constraints (e.g., ) into the generative loop as Runtime Filters. Any AI-generated instruction or video violating a patient’s personalized Range of Motion (ROM) must be intercepted and reverted to a safe state. Critically, this mechanism functions within a Human-in-the-loop (HITL) workflow: when the system detects high-risk compensatory patterns that exceed safety thresholds, it triggers an immediate alert for human clinical review, ensuring that AI-generated feedback remains under the supervision of qualified physiotherapists. Furthermore, leveraging the Chain-of-Thought (CoT) reasoning of MLLMs to provide anatomically consistent explanations (37) can make the complex features of AI more transparent. This causal interpretability serves as a logical cornerstone for establishing Embodied Trust among patients, clinicians, and AI agents, representing a key milestone for SaMD certification (97).

8.3 Limitations

This review has several limitations that should be considered when interpreting its findings. First, although the review followed the PRISMA 2020 reporting framework and was conducted according to a predefined search, screening, extraction, and evidence-mapping plan, it was not prospectively registered in PROSPERO. This limits the extent to which the review protocol can be independently verified. In addition, formal inter-rater reliability statistics, such as Cohen’s kappa, were not prospectively recorded during the original screening and eligibility phases. Although two reviewers independently re-checked the screening decisions, full-text eligibility decisions, and data extraction records during revision, the lack of prospectively recorded agreement statistics remains a methodological limitation.

The evidence base summarized in this review may also be affected by publication bias. Studies reporting successful computer vision-based rehabilitation systems, positive validation results, or promising prototype performance are more likely to appear in the published literature than studies reporting failed clinical validation, poor usability, low adherence, or negative implementation outcomes. Therefore, the current literature may present an overly optimistic view of the maturity and clinical readiness of vision-based rehabilitation technologies.

Another important limitation is the substantial heterogeneity of the included publications. The reviewed corpus included peer-reviewed clinical studies, engineering benchmarks, conference papers, systematic or scoping reviews, rehabilitation prototypes, protocols, and selected frontier preprints. The included populations ranged from stroke, Parkinson’s disease, osteoarthritis, sarcopenia, and orthopedic rehabilitation patients to healthy participants performing simulated rehabilitation tasks. Similarly, the reported outcomes varied widely, including MPJPE, joint angle error, ICC, RMSE, action quality scores, therapist ratings, usability, adherence, and selected clinical indicators. Because of this methodological, clinical, and outcome-level heterogeneity, a formal meta-analysis was not performed. Instead, this review used PAC-based evidence mapping and methodological appraisal to provide a structured qualitative synthesis.

The search strategy primarily focused on English-language literature and major international databases. Relevant studies published in Chinese, Japanese, Korean, German, or other languages may therefore have been underrepresented. This may introduce language-related selection bias, particularly because computer vision-based rehabilitation research is also active in non-English-speaking regions.

The rapidly evolving nature of Generative AI, MLLMs, and foundation-model-assisted rehabilitation represents another limitation. Selected frontier studies and preprints were included to contextualize emerging directions, but these records were treated as emerging evidence and were not weighted as equivalent to peer-reviewed clinical validation studies. Accordingly, recommendations related to MLLMs, autonomous rehabilitation agents, and generative visual feedback should be interpreted cautiously and updated as new peer-reviewed evidence becomes available.

Patient-reported outcomes were underrepresented in the reviewed literature. Many studies emphasized algorithmic or biomechanical metrics, such as pose-estimation error, ROM, ICC, RMSE, or action quality scores, whereas fewer studies systematically reported pain, perceived exertion, satisfaction, confidence, quality of life, functional independence, or long-term adherence. This limits the ability to determine whether technically accurate systems translate into meaningful patient-centered rehabilitation benefits. Future studies should integrate patient-reported outcomes alongside biomechanical, clinical, and usability endpoints.

Finally, population coverage remains incomplete. Although this review covered neurological, orthopedic, and geriatric rehabilitation populations, pediatric rehabilitation and populations with cognitive impairments, such as dementia or traumatic brain injury, were less extensively represented. These populations may require different feedback timing, interface design, caregiver involvement, safety monitoring, and motivational strategies. Future work should therefore validate vision-based rehabilitation systems across broader patient groups and real-world clinical contexts.

9 Conclusion

This paper has systematically reviewed the evolution of computer vision in rehabilitation training, establishing a structured taxonomic perspective through the unified “Perception, Assessment, and Coaching (PAC)” framework. Our analysis indicates a clear technological trajectory: (1) In the Perception Domain, the focus has shifted from simple geometric reconstruction to Biomechanical Digital Twin that ensure anatomical validity via physical constraints; (2) In the Assessment Domain, the research paradigm is evolving from empirical data fitting toward Interpretable Pathological Attribution; (3) In the Coaching Domain, feedback mechanisms are transitioning from rudimentary error reporting to an adaptive, multimodal Sensorimotor Loop designed to support motor relearning based on motor learning principles.

Notably, the deep penetration of Generative AI and MLLMs is fundamentally reconstructing these Domains. From intent-driven semantic sensing and CoT-based clinical reasoning to VSM-based generative feedback, large models are driving the evolution of rehabilitation systems from assistive tools into Autonomous Rehab Agents possessed of clinical logic. Although challenges such as pathological data distortion and clinical fidelity validation remain, the technical pathways established in this framework provide a clear roadmap for addressing the global shortage of rehabilitation resources and the deficit in rehabilitation literacy. Ultimately, the convergence of computer vision and rehabilitation medicine serves as a pivotal driver for the democratization of precision healthcare, providing a vital response to the challenges of a global aging society.

References

• 1.

  CuiALiHWangDZhongJChenYLuH. Global, regional prevalence, incidence and risk factors of knee osteoarthritis in population-based studies. EClinicalMedicine. (2020) 29:100587. 10.1016/j.eclinm.2020.100587

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2.

  FeiginVLAbateMDAbateYHAbd ElHafeezSAbd-AllahFAbdelalimA, et al. Global, regional, and national burden of stroke and its risk factors, 1990–2021: a systematic analysis for the global burden of disease study 2021. Lancet Neurol. (2024) 23:973–1003. 10.1016/S1474-4422(24)00369-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3.

  WangHLinJZhangSZhaoFZhangXWangL, et al. Global, regional and national burdens of stroke and its subtypes: unraveling the correlations with the global aging trend. Neuroepidemiology. (2025). 60:483–502. 10.1159/000546317

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4.

  JesusTSLandryMDDussaultGFronteiraI. Human resources for health (and rehabilitation): six rehab-workforce challenges for the century. Hum Resour Health. (2017) 15:8. 10.1186/s12960-017-0182-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5.

  PhanseVA. Critical gaps in physical therapy in the United States of America: exploring the shortage and schedule a classification. progress in medical sciences. PMS-E115. Prog Med Sci. (2022) 6:E115. 10.47363/PMS/2022(6)E115.

  • CrossRef
  • Google Scholar
  • View reference in article

• 6.

  NikolaevVANikolaevAA. Recent trends in telerehabilitation of stroke patients: a narrative review. NeuroRehabilitation. (2022) 51:1–22. 10.3233/NRE-210330

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7.

  ZhangZ-y.TianLHeKXuLWangX-q.HuangL, et al. Digital rehabilitation programs improve therapeutic exercise adherence for patients with musculoskeletal conditions: a systematic review with meta-analysis. J Orthop Sports Phys Ther. (2022) 52:726–39. 10.2519/jospt.2022.11384

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8.

  ChenYTianYHeM. Monocular human pose estimation: a survey of deep learning-based methods. Comput Vis Image Underst. (2020) 192:102897. 10.1016/j.cviu.2019.102897

  • CrossRef
  • Google Scholar
  • View reference in article

• 9.

  ZhengCWuWChenCYangTZhuSShenJ, et al. Deep learning-based human pose estimation: a survey. ACM Comput Surv. (2023) 56:1–37. 10.1145/3603618

  • CrossRef
  • Google Scholar
  • View reference in article

• 10.

  DesmaraisYMottetDSlangenPMontesinosP. A review of 3D human pose estimation algorithms for markerless motion capture. Comput Vis Image Underst. (2021) 212:103275. 10.1016/j.cviu.2021.103275

  • CrossRef
  • Google Scholar
  • View reference in article

• 11.

  StenumJRossiCRoemmichRT. Two-dimensional video-based analysis of human gait using pose estimation. PLoS Comput Biol. (2021) 17:e1008935. 10.1371/journal.pcbi.1008935

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12.

  LamWWTangYMFongKN. A systematic review of the applications of markerless motion capture (MMC) technology for clinical measurement in rehabilitation. J Neuroeng Rehabil. (2023) 20:57. 10.1186/s12984-023-01186-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13.

  XuJYinSPengY. Human-centric fine-grained action quality assessment. IEEE Trans Pattern Anal Mach Intell. (2025). 47:6242–55. 10.1109/TPAMI.2025.3556935.

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14.

  ZhouKCaiRWangLShumHPLiangX. A comprehensive survey of action quality assessment: method and benchmark. arXiv [Preprint]. arXiv:2412.11149 (2024). 10.48550/arXiv.2412.11149

  • CrossRef
  • Google Scholar
  • View reference in article

• 15.

  ZhaoXGuoCZouQ. Human pose estimation with gated multi-scale feature fusion and spatial mutual information. Vis Comput. (2023) 39:119–37. 10.1007/s00371-021-02317-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 16.

  YangJZhangB-TLiuF-LFuHLaiY-KGaoL. Single-image 3D human reconstruction with 3D-aware diffusion priors and facial enhancement. In: Proceedings of the SIGGRAPH Asia 2025 Conference Papers (2025). p. 1–13. 10.1145/3757377.3763839

  • CrossRef
  • Google Scholar
  • View reference in article

• 17.

  ZhuYPicardD. Decanus to legatus: synthetic training for 2D–3D human pose lifting. In: Proceedings of the Asian Conference on Computer Vision (2022). p. 2848–65. 10.1007/978-3-031-26316-3_16

  • CrossRef
  • Google Scholar
  • View reference in article

• 18.

  GrafRLerchlTNispelKMöllerHAtadMMcGinnisJ, et al. Rule-based key-point extraction for mr-guided biomechanical digital twins of the spine. In: International Workshop on Digital Twin for Healthcare. Springer (2025). p. 109–18. 10.1007/978-3-032-07694-6_11

  • CrossRef
  • Google Scholar
  • View reference in article

• 19.

  RothS. The next frontier in healthcare: perspectives and discussion on building trust and societal acceptance of digital humans in the essential framework of impact biomechanics. Front Bioeng Biotechnol. (2025) 13:1693334. 10.3389/fbioe.2025.1693334

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20.

  SadéeCTestaSBarbaTHartmannKSchuesslerMThiemeA, et al. Medical digital twins: enabling precision medicine and medical artificial intelligence. Lancet Digit Health. (2025). 7:100864. 10.1016/j.landig.2025.02.004

  • CrossRef
  • Google Scholar
  • View reference in article

• 21.

  Galvan-SosaDMatsudaKOkazakiNInuiK. Empirical exploration of the challenges in temporal relation extraction from clinical text. J Nat Lang Process. (2020) 27:383–409. 10.5715/jnlp.27.383

  • CrossRef
  • Google Scholar
  • View reference in article

• 22.

  BayleNLempereurMHutinEMotavasseliDRemy-NerisOGraciesJ-M, et al. Comparison of various smoothness metrics for upper limb movements in middle-aged healthy subjects. Sensors. (2023) 23:1158. 10.3390/s23031158

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23.

  MocciaCMoiranoGPopovicMPizziCFariselliPRichiardiL, et al. Machine learning in causal inference for epidemiology. Eur J Epidemiol. (2024) 39:1097–108. 10.1007/s10654-024-01173-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24.

  WillinghamTBStowellJCollierGBackusD. Leveraging emerging technologies to expand accessibility and improve precision in rehabilitation and exercise for people with disabilities. Int J Environ Res Public Health. (2024) 21:79. 10.3390/ijerph21010079

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25.

  DongYWangSSunJWangMChengL. Three-dimensional human body reconstruction using dual-view normal maps. Symmetry. (2024) 16:1647. 10.3390/sym16121647

  • CrossRef
  • Google Scholar
  • View reference in article

• 26.

  QiaoCRolfeEDLMakESenguptaAPowellRWatsonLP, et al. Prediction of total and regional body composition from 3D body shape. NPJ Digit Med. (2024) 7:298. 10.1038/s41746-024-01289-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27.

  KolarikMSarnovskyMParalicJBabicF. Explainability of deep learning models in medical video analysis: a survey. PeerJ Computer Science. (2023) 9:e1253. 10.7717/peerj-cs.1253

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 28.

  LuSLiuMYinLYinZLiuXZhengW. The multi-modal fusion in visual question answering: a review of attention mechanisms. PeerJ Comput Sci. (2023) 9:e1400. 10.7717/peerj-cs.1400

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29.

  Di MitriDSchneiderJDrachslerH. Keep me in the loop: real-time feedback with multimodal data. Int J Artif Intell Educ. (2022) 32:1093–118. 10.1007/s40593-021-00281-z

  • CrossRef
  • Google Scholar
  • View reference in article

• 30.

  SchneiderJ. Explainable generative AI (GenXAI): a survey, conceptualization, and research agenda. Artif Intell Rev. (2024) 57:289. 10.1007/s10462-024-10916-x

  • CrossRef
  • Google Scholar
  • View reference in article

• 31.

  HeRYouZZhouYChenGDiaoYJiangX, et al. A novel multi-level 3D pose estimation framework for gait detection of parkinson’s disease using monocular video. Front Bioeng Biotechnol. (2024a) 12:1520831. 10.3389/fbioe.2024.1520831

  • CrossRef
  • Google Scholar
  • View reference in article

• 32.

  CapecciMCeravoloMGFerracutiFIarloriSMonteriuARomeoL, et al. The kimore dataset: kinematic assessment of movement and clinical scores for remote monitoring of physical rehabilitation. IEEE Trans Neural Syst Rehabil Eng. (2019) 27:1436–48. 10.1109/TNSRE.2019.2923060

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33.

  DillSAhmadiAGrimmerMHaufeDRohrMZhaoY, et al. Accuracy evaluation of 3D pose reconstruction algorithms through stereo camera information fusion for physical exercises with mediapipe pose. Sensors. (2024) 24:7772. 10.3390/s24237772

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34.

  IonescuCPapavaDOlaruVSminchisescuC. Human3. 6m: large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Trans Pattern Anal Mach Intell. (2013) 36:1325–39. 10.1109/TPAMI.2013.248

  • CrossRef
  • Google Scholar
  • View reference in article

• 35.

  SahaHNBhattacharyaDCDuttaSBeraABasuraySChangdarS, et al. Transforming healthcare with state-of-the-art medical-LLMs: a comprehensive evaluation of current advances using benchmarking framework. Comput Mater Contin. (2026) 86:1–56. 10.32604/cmc.2025.070507.

  • CrossRef
  • Google Scholar
  • View reference in article

• 36.

  ShoitanRMoussaMMTawfikNChoYAbdallahMS. Exploring generative artificial intelligence: a comprehensive guide. PeerJ Comput Sci. (2026) 12:e3276. 10.7717/peerj-cs.3276

  • CrossRef
  • Google Scholar
  • View reference in article

• 37.

  SinghalKTuTGottweisJSayresRWulczynEAminM, et al. Toward expert-level medical question answering with large language models. Nat Med. (2025) 31:943–50. 10.1038/s41591-024-03423-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38.

  ZhuJCaiH. Large language models and their impact in medical imaging education. PeerJ Comput Sci. (2025) 11:e3433. 10.7717/peerj-cs.3433

  • CrossRef
  • Google Scholar
  • View reference in article

• 39.

  AoyagiYOhnishiEYamamotoYKadoNSuzukiTOhnishiH, et al. Feedback protocol of ‘fading knowledge of results’ is effective for prolonging motor learning retention. J Phys Ther Sci. (2019) 31:687–91. 10.1589/jpts.31.687

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40.

  ChenTTMakTCNgSSWongTW. Attentional focus strategies to improve motor performance in older adults: a systematic review. Int J Environ Res Public Health. (2023) 20:4047. 10.3390/ijerph20054047

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 41.

  OsawaKYouYSunYWangT-QZhangSShimodozonoM, et al. Telerehabilitation system based on openpose and 3D reconstruction with monocular camera. J Robot Mechatron. (2023) 35:586–600. 10.20965/jrm.2023.p0586

  • CrossRef
  • Google Scholar
  • View reference in article

• 42.

  YaoLLeiQZhangHDuJGaoS. A contrastive learning network for performance metric and assessment of physical rehabilitation exercises. IEEE Trans Neural Syst Rehabil Eng. (2023) 31:3790–802. 10.1109/TNSRE.2023.3317411

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 43.

  YeungEHChenYFokWWLauGK. Validation of consumer-grade digital camera-based human activity evaluation for upper limb exercises and development of a therapist-guided, automated telerehabilitation framework and platform for stroke rehabilitation. arXiv [Preprint]. arXiv:2311.13088 (2023). 10.48550/arXiv.2311.13088

  • CrossRef
  • Google Scholar
  • View reference in article

• 44.

  YoshimuraNMoralesJMaekawaTHaraT. Openpack: a large-scale dataset for recognizing packaging works in IoT-enabled logistic environments. In: 2024 IEEE International Conference on Pervasive Computing and Communications (PerCom). IEEE (2024). p. 90–7. 10.1109/PerCom59722.2024.10494448

  • CrossRef
  • Google Scholar
  • View reference in article

• 45.

  BidulkaLGholamiMZhengJMcKeownMJWangZJ. Escape: energy-based selective adaptive correction for out-of-distribution 3d human pose estimation. Neurocomputing. (2025) 611:128605. 10.1016/j.neucom.2024.128605

  • CrossRef
  • Google Scholar
  • View reference in article

• 46.

  ReddyLAnandKKaushikSRodrigoCMcKayJLKesarTM, et al. Classifying simulated gait impairments using privacy-preserving explainable artificial intelligence and mobile phone videos. PLoS Digit Health. (2025) 4:e0001004. 10.1371/journal.pdig.0001004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 47.

  KohKOppizziGKehsGZhangL-Q. Abnormal coordination of upper extremity during target reaching in persons post stroke. Sci Rep. (2023) 13:12838. 10.1038/s41598-023-39684-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 48.

  GildeaKMercadal-BaudartCBlythmanRSimmsC. Temporally optimized inverse kinematics for 6DOF human pose estimation. In: ESB 2022: 27th Congress of the European Society of Biomechanics (2022).

  • Google Scholar

• 49.

  TangJAbediAColellaTJKhanSS. Rehabilitation exercise quality assessment and feedback generation using large language models with prompt engineering. In: International Joint Conference on Artificial Intelligence. Springer (2025). p. 60–75. 10.1007/978-981-95-0568-5_5

  • CrossRef
  • Google Scholar

• 50.

  LiangSHeJWangCLiaoLZhangGChenY, et al. SDPose: exploiting diffusion priors for out-of-domain and robust pose estimation. arXiv [Preprint]. arXiv:2509.24980 (2025). 10.48550/arXiv.2509.24980

  • CrossRef
  • Google Scholar

• 51.

  SongY-PWuXYuanZQiaoJ-JPengQ. Posturehmr: posture transformation for 3D human mesh recovery. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2024). p. 9732–41. 10.1109/CVPR52733.2024.00929

  • CrossRef
  • Google Scholar

• 52.

  HussainIJanyR. Interpreting stroke-impaired electromyography patterns through explainable artificial intelligence. Sensors. (2024) 24:1392. 10.3390/s24051392

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 53.

  PornpipatsakulKChuengwutigoolWChaichaowaratRFoongchomcheayA. Bridging exercise monitoring system using RGB camera for stroke rehabilitation. In: TENCON 2023–2023 IEEE Region 10 Conference (TENCON). IEEE (2023). p. 960–65. 10.1109/TENCON58879.2023.10322445

  • CrossRef
  • Google Scholar

• 54.

  SimoesWReisLAraujoCMaia jrJ. Accuracy assessment of 2D pose estimation with mediapipe for physiotherapy exercises. Procedia Comput Sci. (2024) 251:446–53. 10.1016/j.procs.2024.11.132

  • CrossRef
  • Google Scholar

• 55.

  VineethNDarshanPDeepuKDaragajAPrabhuAG. Assessment technology for physiotherapy practices using deep learning. In: 2024 International Conference on Advancements in Smart, Secure and Intelligent Computing (ASSIC). IEEE (2024). p. 1–5. 10.1109/ASSIC60049.2024.10507911

  • CrossRef
  • Google Scholar

• 56.

  KuangZWangJSunDZhaoJShiLZhuY. Hierarchical contrastive representation for accurate evaluation of rehabilitation exercises via multi-view skeletal representations. IEEE Trans Neural Syst Rehabil Eng. (2024). 33:201–11.10.1109/TNSRE.2024.3523906.

  • CrossRef
  • Google Scholar

• 57.

  KakuALiuKParnandiARajamohanHRVenkataramananKVenkatesanA, et al. Strokerehab: a benchmark dataset for sub-second action identification. Adv Neural Inf Process Syst. (2022) 35:1671–84. 10.5555/3600270.3600392

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58.

  UllahAANajamSJalalA. A novel Parkinson patients physical action system via PoseNet and transformer. In: 2025 4th International Conference on Communication, Computing and Digital Systems (C-CODE). IEEE (2025). p. 1–6. 10.1109/C-CODE67372.2025.11204160

  • CrossRef
  • Google Scholar

• 59.

  CuiCSunarMSEg SuG. Deep vision-based real-time hand gesture recognition: a review. PeerJ Comput Sci. (2025) 11:e2921. 10.7717/peerj-cs.2921

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 60.

  AhmedTThopalliKRikakisTTuragaPKelliherAHuangJ-B, et al. Automated movement assessment in stroke rehabilitation. Front Neurol. (2021) 12:720650. 10.3389/fneur.2021.720650

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61.

  PengYLeeJWatanabeS. I3D: transformer architectures with input-dependent dynamic depth for speech recognition. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2023). p. 1–5. 10.1109/ICASSP49357.2023.10096662

  • CrossRef
  • Google Scholar

• 62.

  FeichtenhoferCFanHMalikJHeK. Slowfast networks for video recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019). p. 6202–11. 10.1109/ICCV.2019.00630

  • CrossRef
  • Google Scholar

• 63.

  BieblJTRykalaMStrobelMKaur BollingerPUlmBKraftE, et al. App-based feedback for rehabilitation exercise correction in patients with knee or hip osteoarthritis: prospective cohort study. J Med Internet Res. (2021) 23:e26658. 10.2196/26658

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64.

  HuRDiaoYWangYLiGHeRNingY, et al. Effective evaluation of HGcnMLP method for markerless 3D pose estimation of musculoskeletal diseases patients based on smartphone monocular video. Front Bioeng Biotechnol. (2023) 11:1335251. 10.3389/fbioe.2023.1335251

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65.

  KryeemARazSEluzDItahDHel-OrHShimshoniI. Personalized monitoring in home healthcare: an assistive system for post hip replacement rehabilitation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023). p. 1868–77. 10.1109/ICCVW60793.2023.00201

  • CrossRef
  • Google Scholar

• 66.

  YuCXiaoBGaoCYuanLZhangLSangN, et al. Lite-hrnet: a lightweight high-resolution network. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021). p. 10440–50. 10.1109/CVPR46437.2021.01030

  • CrossRef
  • Google Scholar

• 67.

  XuYZhangJZhangQTaoD. Vitpose: simple vision transformer baselines for human pose estimation. Adv Neural Inf Process Syst. (2022) 35:38571–84. 10.5555/3600270.3603065

  • CrossRef
  • Google Scholar

• 68.

  HeSMengDWeiMGuoHYangGWangZ. Proposal and validation of a new approach in tele-rehabilitation with 3D human posture estimation: a randomized controlled trial in older individuals with sarcopenia. BMC Geriatr. (2024b) 24:586. 10.1186/s12877-024-05188-7

  • CrossRef
  • Google Scholar

• 69.

  AbediAMalmirianMKhanSS. Cross-modal video to body-joints augmentation for rehabilitation exercise quality assessment. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer (2023). p. 320–7. 10.1007/978-3-031-74640-6_24

  • CrossRef
  • Google Scholar

• 70.

  Aguilar-OrtegaRBerral-SolerRJiménez-VelascoIRomero-RamírezFJGarcía-MarínMZafra-PalmaJ, et al. Uco physical rehabilitation: new dataset and study of human pose estimation methods on physical rehabilitation exercises. Sensors. (2023) 23:8862. 10.3390/s23218862

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 71.

  LugaresiCTangJNashHMcClanahanCUbowejaEHaysM, et al. Mediapipe: a framework for building perception pipelines. arXiv [Preprint]. arXiv:1906.08172 (2019). 10.48550/arXiv.1906.08172

  • CrossRef
  • Google Scholar

• 72.

  BazarevskyVGrishchenkoIRaveendranKZhuTZhangFGrundmannM. Blazepose: on-device real-time body pose tracking. arXiv [Preprint]. arXiv:2006.10204 (2020). 10.48550/arXiv.2006.10204

  • CrossRef
  • Google Scholar

• 73.

  MarusicATapusA. Skeleton-based transformer for classification of errors and better feedback in low back pain physical rehabilitation exercises. In: 2025 International Conference On Rehabilitation Robotics (ICORR). IEEE (2025). p. 1274–80. 10.1109/ICORR66766.2025.11063192

  • CrossRef
  • Google Scholar

• 74.

  MartınezGH. Openpose: whole-body pose estimation (Ph.D. dissertation) (2019).

  • Google Scholar

• 75.

  FangH-SLiJTangHXuCZhuHXiuY, et al. Alphapose: whole-body regional multi-person pose estimation and tracking in real-time. IEEE Trans Pattern Anal Mach Intell. (2022a) 45:7157–73. 10.1109/TPAMI.2022.3222784

  • CrossRef
  • Google Scholar

• 76.

  PereiraBCunhaBVianaPLopesMMeloASSousaAS. A machine learning app for monitoring physical therapy at home. Sensors. (2023) 24:158. 10.3390/s24010158

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 77.

  CunhaBMaç aesJAmorimI. Smartphone-based markerless motion capture for accessible rehabilitation: a computer vision study. Sensors. (2025) 25:5428. 10.3390/s25175428

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 78.

  XuBFangSLiZYangSXieDPuS. Prime: 3D human pose and body shape recovery with perspective projection. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE (2023). p. 1–5. 10.1109/ICASSP49357.2023.10094299

  • CrossRef
  • Google Scholar

• 79.

  ShettyKBirkholdAJaganathanSStrobelNKowarschikMMaierA, et al. Pliks: a pseudo-linear inverse kinematic solver for 3D human body estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2023). p. 574–84. 10.1109/CVPR52729.2023.00063

  • CrossRef
  • Google Scholar

• 80.

  PavlloDFeichtenhoferCGrangierDAuliM. 3D human pose estimation in video with temporal convolutions and semi-supervised training. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019). p. 7753–62. 10.1109/CVPR.2019.00794

  • CrossRef
  • Google Scholar

• 81.

  LiWLiuHTangHWangPVan GoolL. MHFormer: multi-hypothesis transformer for 3D human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022). p. 13147–56. 10.1109/CVPR52688.2022.01280

  • CrossRef
  • Google Scholar

• 82.

  ChenTFangCShenXZhuYChenZLuoJ. Anatomy-aware 3D human pose estimation with bone-based pose decomposition. IEEE Trans Circ Syst Video Technol. (2021) 32:198–209. 10.1109/TCSVT.2021.3057267

  • CrossRef
  • Google Scholar

• 83.

  FangLMouLGuYHuYChenBChenX, et al. Global–local multi-stage temporal convolutional network for cataract surgery phase recognition. Biomed Eng Online. (2022b) 21:82. 10.1186/s12938-022-01048-w

  • CrossRef
  • Google Scholar

• 84.

  FiltjensBVanrumsteBSlaetsP. Skeleton-based action segmentation with multi-stage spatial-temporal graph convolutional neural networks. IEEE Trans Emerg Top Comput. (2022) 12:202–12. 10.1109/TETC.2022.3230912

  • CrossRef
  • Google Scholar

• 85.

  DebSIslamMFRahmanSRahmanS. Graph convolutional networks for assessment of physical rehabilitation exercises. IEEE Trans Neural Syst Rehabil Eng. (2022) 30:410–9. 10.1109/TNSRE.2022.3150392

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 86.

  YangJLyuMQiZShiY. Deep learning based image quality assessment: a survey. Procedia Comput Sci. (2023) 221:1000–5. 10.1016/j.procs.2023.08.080

  • CrossRef
  • Google Scholar

• 87.

  SunJZhengMXianJXieRNiXZhangS, et al. Research on rehabilitation exercise guidance system based on action quality assessment. In: 2024 17th International Convention on Rehabilitation Engineering and Assistive Technology (i-CREATe). IEEE (2024). p. 1–5. 10.1109/i-CREATe62067.2024.10776381

  • CrossRef
  • Google Scholar

• 88.

  SeredinOKopylovASurkovEMityugovNTokarevABagchiP, et al. Automated control of rehabilitation process in physical therapy using a novel human skeleton-based balanced time warping algorithm. Sensors. (2025) 25:6696. 10.3390/s25216696

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 89.

  HribernikMUmekATomažičSKosA. Review of real-time biomechanical feedback systems in sport and rehabilitation. Sensors. (2022) 22:3006. 10.3390/s22083006

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 90.

  KnippenbergEVerbruggheJLamersIPalmaersSTimmermansASpoorenA. Markerless motion capture systems as training device in neurological rehabilitation: a systematic review of their use, application, target population and efficacy. J Neuroeng Rehabil. (2017) 14:61. 10.1186/s12984-017-0270-x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 91.

  SchaffertNJanzenTBMattesKThautMH. A review on the relationship between sound and movement in sports and rehabilitation. Front Psychol. (2019) 10:244. 10.3389/fpsyg.2019.00244

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 92.

  JohanssonGMÖhbergF. Augmented feedback in post-stroke gait rehabilitation derived from sensor-based gait reports—a longitudinal case series. Sensors. (2025) 25:3109. 10.3390/s25103109

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 93.

  LucianiBPedrocchiATropeaPSeregniABraghinFGandollaM. Augmented reality for upper limb rehabilitation: real-time kinematic feedback with HoloLens 2. Virtual Real. (2025) 29:57. 10.1007/s10055-025-01124-1

  • CrossRef
  • Google Scholar

• 94.

  CaiSWeiXSuEWuWZhengHXieL. Online compensation detecting for real-time reduction of compensatory motions during reaching: a pilot study with stroke survivors. J Neuroeng Rehabil. (2020) 17:58. 10.1186/s12984-020-00687-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 95.

  ItoDKawakamiMHosoiYKamimotoTYamadaYTakemuraR, et al. Development of a quantitative assessment for abnormal flexor synergy index in patients with stroke: a validity and responsiveness study. J Neuroeng Rehabil. (2024) 21:229. 10.1186/s12984-024-01534-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 96.

  OwakiDSekiguchiYHondaKIzumiS-I. Two-week rehabilitation with auditory biofeedback prosthesis reduces whole body angular momentum range during walking in stroke patients with hemiplegia: a randomized controlled trial. Brain Sci. (2021) 11:1461. 10.3390/brainsci11111461

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 97.

  KellyCJKarthikesalingamASuleymanMCorradoGKingD. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. (2019) 17:195. 10.1186/s12916-019-1426-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 98.

  ZhuXBoukhennoufaILiewBGaoCYuWMcDonald-MaierKD, et al. Monocular 3D human pose markerless systems for gait assessment. Bioengineering. (2023) 10:653. 10.3390/bioengineering10060653

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 99.

  MiaoSLiuZWangDShenXShenN. Applying hybrid deep learning models to assess upper limb rehabilitation. IEEE Access. (2024). 12:154337–48. 10.1109/ACCESS.2024.3482115

  • CrossRef
  • Google Scholar

• 100.

  RahmanJUHanifMHaiderUQaisarSMAyouniS. Action recognition via shallow CNNs on intelligently selected motion data. Comput Mater Contin. (2026) 86:96. 10.32604/cmc.2025.071251.

  • CrossRef
  • Google Scholar

• 101.

  ZhengCZhuSMendietaMYangTChenCDingZ. 3D human pose estimation with spatial and temporal transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2021). p. 11656–65. 10.1109/ICCV48922.2021.01145

  • CrossRef
  • Google Scholar

• 102.

  YinHParmarPXuDZhangYZhengTFuW. A decade of action quality assessment: largest systematic survey of trends, challenges, and future directions. arXiv [Preprint]. arXiv:2502.02817 (2025).

  • Google Scholar

• 103.

  HalilajERajagopalAFiterauMHicksJLHastieTJDelpSL. Machine learning in human movement biomechanics: best practices, common pitfalls, and new opportunities. J Biomech. (2018) 81:1–11. 10.1016/j.jbiomech.2018.09.009

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 104.

  LiZSedlarJCarpentierJLaptevIMansardNSivicJ. Estimating 3D motion and forces of human–object interactions from internet videos. Int J Comput Vis. (2022b) 130:363–83. 10.1007/s11263-021-01540-1

  • CrossRef
  • Google Scholar

• 105.

  PavlakosGChoutasVGhorbaniNBolkartTOsmanAATzionasD, et al. Expressive body capture: 3D hands, face, and body from a single image. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019). p. 10975–85. 10.1109/CVPR.2019.01123

  • CrossRef
  • Google Scholar

• 106.

  HsuC-HJangJ-SR. Enhancing 3D human pose estimation with bone length adjustment. In: Proceedings of the Asian Conference on Computer Vision (2024). p. 3723–38. 10.1007/978-981-96-0885-0_14

  • CrossRef
  • Google Scholar

• 107.

  KimJ-WChoiJ-YHaE-JChoiJ-H. Human pose estimation using mediapipe pose and optimization method based on a humanoid model. Appl Sci. (2023) 13:2700. 10.3390/app13042700

  • CrossRef
  • Google Scholar

• 108.

  AytekinAILiCLuvizonDDabralROswaldMHabermannM, et al. Physics-based human pose estimation from a single moving RGB camera. In: Proceedings of the Computer Vision and Pattern Recognition Conference (2025). p. 3891–900.

  • Google Scholar

• 109.

  ParkMOhSJeongTYuS. Multi-stage temporal convolutional network with moment loss and positional encoding for surgical phase recognition. Diagnostics. (2022) 13:107. 10.3390/diagnostics13010107

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 110.

  AverellEKnoxDvan WijckF. A real-time algorithm for the detection of compensatory movements during reaching. J Rehabil Assist Technol Eng. (2022) 9:20556683221117085. 10.1177/20556683221117085

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 111.

  GoldbraikhAShubiORubinOPughCMLauferS. MS-TCRNet: multi-stage temporal convolutional recurrent networks for action segmentation using sensor-augmented kinematics. Pattern Recognit. (2024) 156:110778. 10.1016/j.patcog.2024.110778

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 112.

  VakanskiAJunH-p.PaulDBakerR. A data set of human body movements for physical rehabilitation exercises. Data. (2018) 3:2. 10.3390/data3010002

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 113.

  BruceXLiuYChanKCChenCW. Egcn++: a new fusion strategy for ensemble learning in skeleton-based rehabilitation exercise assessment. IEEE Trans Pattern Anal Mach Intell. (2024) 46:6471–85. 10.1109/TPAMI.2024.3378753

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 114.

  ZhangKZhangPTuXLiuZXuPWuC, et al. Sr-pose: A novel non-contact real-time rehabilitation evaluation method using lightweight technology. IEEE Trans Neural Syst Rehabil Eng. (2023) 31:4179–88. 10.1109/TNSRE.2023.3324960

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 115.

  ReissTHoshenY. Attribute-based representations for accurate and interpretable video anomaly detection. arXiv [Preprint]. arXiv:2212.00789 (2022). 10.48550/arXiv.2212.00789

  • CrossRef
  • Google Scholar

• 116.

  CherianAWangJ. Generalized one-class learning using pairs of complementary classifiers. IEEE Trans Pattern Anal Mach Intell. (2021) 44:6993–7009. 10.1109/TPAMI.2021.3092999

  • CrossRef
  • Google Scholar

• 117.

  MesquitaGC’oiasARDubrawskiABernardinoA. Frame-level real-time assessment of stroke rehabilitation exercises from video-level labeled data: task-specific vs. foundation models. arXiv [Preprint]. arXiv:2506.03752 (2025). 10.48550/arXiv.2506.03752

  • CrossRef
  • Google Scholar

• 118.

  SherifOHamdiA. Error-guided pose augmentation: enhancing rehabilitation exercise assessment through targeted data generation. arXiv [Preprint]. arXiv:2506.09833 (2025). 10.48550/arXiv.2506.09833

  • CrossRef
  • Google Scholar

• 119.

  GadhviRDesaiPSiddharthK. Posepilot: an edge-AI solution for posture correction in physical exercises. In: Iberian Conference on Pattern Recognition and Image Analysis. Springer (2025). p. 208–19. 10.1007/978-3-031-99568-2_17

  • CrossRef
  • Google Scholar

• 120.

  AntonjMCornianiGPielaKEusebiLCasadioMSciuttiA, et al. Implementing and testing the novel rehab-pal system on unimpaired population: towards rehabilitation engagement at home with a socially assistive robot for pediatric adherence. In: 2025 International Conference On Rehabilitation Robotics (ICORR). IEEE (2025). p. 1178–84. 10.1109/ICORR66766.2025.11063217

  • CrossRef
  • Google Scholar

• 121.

  WangLChenXDengQYouMXuYLiuD, et al. Effectiveness of a digital rehabilitation program based on computer vision and augmented reality for isolated meniscus injury: protocol for a prospective randomized controlled trial. J Orthop Surg Res. (2023) 18:936. 10.1186/s13018-023-04367-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 122.

  NaqviWMNaqviIWMishraGVVardhanVD. The future of telerehabilitation: embracing virtual reality and augmented reality innovations. Pan Afr Med J. (2024) 47:157. 10.11604/pamj.2024.47.157.42956.

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 123.

  VenkatesanMMohanHRyanJRSchürchCMNolanGPFrakesDH, et al. Virtual and augmented reality for biomedical applications. Cell Rep Med. (2021) 2:100348. 10.5281/zenodo.4976835

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 124.

  YeungAWKTosevskaAKlagerEEibensteinerFLaxarDStoyanovJ, et al. Virtual and augmented reality applications in medicine: analysis of the scientific literature. J Med Internet Res. (2021) 23:e25499. 10.2196/25499

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 125.

  HossainDScottSHCluffTDukelowSP. The use of machine learning and deep learning techniques to assess proprioceptive impairments of the upper limb after stroke. J Neuroeng Rehabil. (2023) 20:15. 10.1186/s12984-023-01140-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 126.

  UkeyJRogersCUhlrichSAkcakayaMSethiA. Enhancing stroke recovery assessment: a machine learning approach to real-world hand function analysis. Int J Med Inform. (2025) 204:106077. 10.1016/j.ijmedinf.2025.106077

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 127.

  SonodaYKurokawaRNakamuraYKanzawaJKurokawaMOhizumiY, et al. Diagnostic performances of gpt-4o, claude 3 opus, and gemini 1.5 pro in “diagnosis please” cases. Jpn J Radiol. (2024) 42:1231–5. 10.1007/s11604-024-01619-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 128.

  LeeSLeeELeeK-SPyunS-B. Explainable artificial intelligence on safe balance and its major determinants in stroke patients. Sci Rep. (2024) 14:23735. 10.1038/s41598-024-74689-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 129.

  KohKOppizziGBaghiRKehsGJZhangL-Q. Loss of joint individuation and abnormal synergy post stroke in upper limb movements. Neurorehabil Neural Repair. (2025) 39:715–27. 10.1177/15459683251340914

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 130.

  LeeS-HSongW-K. Mitigating trunk compensatory movements in post-stroke survivors through visual feedback during robotic-assisted arm reaching exercises. Sensors. (2024) 24:3331. 10.3390/s24113331

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 131.

  RenYZhouYYangJShiJLiuDLiuF, et al. Customize-a-video: one-shot motion customization of text-to-video diffusion models. In: European Conference on Computer Vision. Springer (2024). p. 332–49. 10.1007/978-3-031-73024-5_20

  • CrossRef
  • Google Scholar

• 132.

  MilosevicBLeardiniAFarellaE. Kinect and wearable inertial sensors for motor rehabilitation programs at home: state of the art and an experimental comparison. Biomed Eng Online. (2020) 19:25. 10.1186/s12938-020-00762-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 133.

  SethiDBhartiSPrakashC. A comprehensive survey on gait analysis: history, parameters, approaches, pose estimation, and future work. Artif Intell Med. (2022) 129:102314. 10.1016/j.artmed.2022.102314

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 134.

  FalisseAPittoLKainzHHoangHWesselingMVan RossomS, et al. Physics-based simulations to predict the differential effects of motor control and musculoskeletal deficits on gait dysfunction in cerebral palsy: a retrospective case study. Front Hum Neurosci. (2020) 14:40. 10.3389/fnhum.2020.00040

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 135.

  DindorfCDullyJKonradiJWolfCBeckerSSimonS, et al. Enhancing biomechanical machine learning with limited data: generating realistic synthetic posture data using generative artificial intelligence. Front Bioeng Biotechnol. (2024) 12:1350135. 10.3389/fbioe.2024.1350135

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 136.

  ScandelliFTemporitiFVecchiSManesNPozziARajevichLet al. Action observation and motor imagery during hand immobilization period accelerate motor and functional recovery in patients with surgical fixation for distal radial fractures: a randomized controlled trial. Arch Phys Med Rehabil. (2025). 10.1016/j.apmr.2025.12.015

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 137.

  LinT-YMaireMBelongieSHaysJPeronaPRamananD, et al. Microsoft coco: common objects in context. In: European Conference on Computer Vision. Springer (2014). p. 740–55. 10.1007/978-3-319-10602-1_48

  • CrossRef
  • Google Scholar

• 138.

  KayWCarreiraJSimonyanKZhangBHillierCVijayanarasimhanS, et al. The kinetics human action video dataset. arXiv [Preprint]. arXiv:1705.06950 (2017).

  • Google Scholar

• 139.

  MehtaDRhodinHCasasDFuaPSotnychenkoOXuW, et al. Monocular 3D human pose estimation in the wild using improved CNN supervision. In: 2017 International Conference on 3D Vision (3DV). IEEE (2017). p. 506–16.

  • Google Scholar

• 140.

  MahmoodNGhorbaniNTrojeNFPons-MollGBlackMJ. Amass: archive of motion capture as surface shapes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2019). p. 5442–51.

  • Google Scholar

• 141.

  LiuJShahroudyAPerezMWangGDuanL-YKotAC. Ntu rgb+ d 120: a large-scale benchmark for 3d human activity understanding. IEEE Trans Pattern Anal Mach Intell. (2019) 42:2684–701. 10.1109/TPAMI.2019.2916873

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 142.

  AntunesJBernardinoASmailagicASiewiorekDP. AHA-3D: a labelled dataset for senior fitness exercise recognition and segmentation from 3D skeletal data. In: BMVC (2018). p. 332.

  • Google Scholar

• 143.

  ParmarPMorrisB. Action quality assessment across multiple actions. In: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE (2019). p. 1468–76. 10.1109/WACV.2019.00161

  • CrossRef
  • Google Scholar

• 144.

  GuptaAChaharTGoswamiMPPalitRPandeyD. Barbell exercise classification and repetition counting. In: International Conference on Data, Electronics and Computing. Springer (2024). p. 257–67.

  • Google Scholar

• 145.

  GaoYVedulaSSReileyCEAhmidiNVaradarajanBLinHC, et al. Jhu-isi gesture and skill assessment working set (jigsaws): a surgical activity dataset for human motion modeling. In: MICCAI Workshop: M2cai. Vol. 3 (2014). p. 3.

  • Google Scholar

• 146.

  WoznowskiPBurrowsADietheTFafoutisXHallJHannunaS, et al. Sphere: a sensor platform for healthcare in a residential environment. In: Designing, Developing, and Facilitating Smart Cities: Urban Design to IoT Solutions. Springer (2016). p. 315–33. 10.1007/978-3-319-44924-1_14

  • CrossRef
  • Google Scholar

• 147.

  PinteaSLZhengJLiXBankPJvan HiltenJJvan GemertJC. Hand-tremor frequency estimation in videos. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops (2018).

  • Google Scholar