Artificial Intelligence as Assessment Tool in Occupational Therapy: A Scoping Review

Abstract

Occupational therapy (OT) is vital in improving functional outcomes and aiding recovery for individuals with long-term disabilities, particularly those resulting from neurological diseases. Traditional assessment methods often rely on clinical judgment and individualized evaluations, which may overlook broader, data-driven insights. The integration of artificial intelligence (AI) presents a transformative opportunity to enhance assessment precision and personalize therapeutic interventions. Additionally, advancements in human–computer interaction (HCI) enable more intuitive and adaptive AI-driven assessment tools, improving user engagement and accessibility in OT. This scoping review investigates current applications of AI in OT, particularly regarding the evaluation of functional outcomes and support for clinical decision-making. The literature search was conducted using the PubMed and Scopus databases. Studies were included if they focused on AI applications in evaluating functional outcomes within OT assessment tools. Out of an initial pool of 85 articles, 13 met the inclusion criteria, highlighting diverse AI methodologies such as support vector machines, deep neural networks, and natural language processing. These were primarily applied in domains including motor recovery, pediatric developmental assessments, and cognitive engagement evaluations. Findings suggest that AI can significantly improve evaluation processes by systematically integrating diverse data sources (e.g., sensor measurements, clinical histories, and behavioral analytics), generating precise predictive insights that facilitate tailored therapeutic interventions and comprehensive assessments of both pre- and post-treatment strategies. This scoping review also identifies existing gaps and proposes future research directions to optimize AI-driven assessment tools in OT.

1. Introduction

Occupational therapy (OT) is a cornerstone of rehabilitation, dedicated to helping individuals with physical, cognitive, social, or mental impairments regain function and participate fully in daily life [1,2]. OT is able to use the human occupations in order to analyze and classify the participation in daily routines. In this way, the independence and socialization of the individuals are promoted according to their own wishes. It is something that comes in full accordance with the approach of the World Health Organization and the classification of functionality with the main purpose of the well-being of citizens (WHO, 2001) [3]. By focusing on meaningful activities—from self-care and work to leisure—OT interventions aim to maximize functional independence, restore confidence, and ultimately improve quality of life for people recovering from injury or living with disability [1,3,4]. This vital role is recognized across diverse rehabilitation contexts; for example, in recovery for individuals with long-term disabilities, particularly those resulting from neurological diseases, personalized rehabilitation programs routinely include OT to enhance patients’ daily functioning and outcomes [3,4,5]. Such examples underscore the significance of OT within multidisciplinary rehabilitation efforts to address the needs of a growing disabled population.

OT assessments have traditionally relied on standardized tools, clinical observations, and personal interviews to evaluate a client’s abilities and challenges in activities of daily living [6]. These conventional methods—including performance-based measures and questionnaires—provide essential information for treatment planning. However, they also have notable limitations. Assessments often depend on the therapist′s observation and subjective rating, which can introduce bias or inconsistency. Many tools capture only a snapshot of performance in clinical settings, potentially missing fluctuations in a person’s day-to-day function. In some cases, the lack of objective quantification makes it difficult to precisely link observed performance to underlying capacity. This uncertainty can limit clinicians’ confidence in traditional assessment results. Moreover, these methods can be labor-intensive and may not effectively capture long-term progress. Consequently, there is a pressing need for more sensitive, efficient, and objective evaluation tools [7].

The integration of artificial intelligence (AI) and machine learning (ML) into OT has ushered in a new era of data-driven, personalized care [8,9]. AI refers to the simulation of human intelligence by machines, enabling them to perform tasks such as learning, reasoning, and decision-making, often through methods like ML, deep learning (DL), and data mining. AI technologies can analyze extensive datasets to identify patterns and predict outcomes, enhancing the accuracy of assessments and enabling the development of individualized rehabilitation plans [2,8,10,11]. Early evidence suggests that AI-driven approaches can improve assessment precision and tailor interventions to meet individual needs [12]. Human–computer interaction (HCI) plays a crucial role in these advancements by optimizing user interfaces and ensuring seamless integration of AI-powered tools into clinical practice [13]. For instance, AI-powered platforms are increasingly being used to score standardized assessments, offer gamified exercises to enhance patient engagement, and provide real-time feedback for more dynamic treatment planning [14,15]. Furthermore, wearable devices equipped with AI algorithms enable continuous monitoring of service users’ progress, facilitating more responsive and data-informed interventions [16,17,18]. These advancements enhance therapists’ ability to conduct holistic and accurate assessments, thereby contributing to improved outcomes.

Several secondary studies have explored AI and related technologies in rehabilitation, each with a distinct focus. For instance, a recent network meta-analysis evaluated eight different cognitive training interventions (including computer-based and virtual reality approaches) for post-stroke cognitive impairment, finding that augmenting traditional therapy with computer-based training yields superior improvements in cognition and daily functioning [19]. Another study conducted a bibliometric analysis of AI-driven gait and biomechanical assessments in neurology, identifying major research themes (ML for gait analysis, wearable sensor technologies, cognitive disorder detection, etc.) and noting the relatively limited collaboration in this emerging field [7]. In pediatric rehabilitation, Guo et al. (2024) reviewed two decades of cerebral palsy treatment devices and observed that most rely on force-based physical stimulation, while suggesting that integrating AI into such equipment could enable more personalized and effective interventions for children with CP [20]. Other reviews have taken broader or adjacent perspectives: Franssen et al. (2019) examined personalized medicine in chronic pulmonary disease, calling for multi-factorial “systems medicine” models and clinical decision support tools to better tailor COPD treatments, and a multidisciplinary overview of cerebellar mutism syndrome in pediatric neuro-oncology synthesized current knowledge of that condition and outlined research priorities to improve its diagnosis, prevention, and management [5].

In the stroke domain, Holguín et al. (2022) bridged advanced neuroinformatics with practice by reviewing how connectome alterations after mild stroke relate to cognitive impairment, and emphasized the OT role in translating these cutting-edge insights into clinical innovation [21]. Notably, one recent review provided a comprehensive survey of AI applications in physical and mental rehabilitation, highlighting technologies from natural language processing and computer vision to robotics that are transforming recovery processes in an occupational therapy context [2]. However, each of these prior studies is limited to particular conditions, populations, or subsets of rehabilitation, and none offers an expansive synthesis of AI’s role specifically within OT. The present scoping review addresses this gap by systematically mapping the breadth of AI applications in occupational therapy, thus uniquely contributing a comprehensive OT-centric perspective on how AI is being applied across diverse practice areas and identifying opportunities for future OT practice and research in the era of AI.

Despite promising technological advancements, the current body of literature exploring AI applications in OT tends to focus predominantly on motor function recovery, while other critical areas, such as cognitive rehabilitation, language rehabilitation, and integration of personal preferences of service users in daily occupations, have not been extensively addressed. Thus, a comprehensive scoping review is necessary to systematically identify these existing gaps and guide future research directions [2,10,11]. Moreover, inconsistencies in AI methodologies and the absence of standardized protocols create significant challenges to the integration of AI into routine OT practice. Many AI-driven tools are still in experimental or prototype stages, with limited validation in clinical settings or on specific populations [2,10,11]. The generalizability of findings is often constrained by small sample sizes and narrowly defined populations [7]. Additionally, ethical concerns related to data privacy, algorithmic bias, and the transparency of AI decision-making processes must be addressed. Ensuring data security and safeguarding against biases are critical to building trust in AI-driven assessments. Ethical considerations also extend to ensuring that AI-generated insights are transparent and interpretable by practitioners, fostering a collaborative human–AI partnership.

Given the rising adoption of AI in various healthcare domains and the limited comprehensive reviews in OT, this study aims to synthesize emerging applications to support evidence-based innovation in OT assessment. While existing secondary studies have explored AI in rehabilitation more broadly, including applications in gait analysis, stroke, cerebral palsy, and mental health, none have specifically focused on the diverse assessment roles of AI within occupational therapy. This scoping review uniquely addresses that gap by providing an OT-centered synthesis of AI applications across functional, cognitive, developmental, and assistive domains. The primary objectives of the review are to (i) examine how AI technologies—including ML, DL, computer vision, and related methods—are being applied to assess functional abilities, occupational performance, personal preferences or needs, and rehabilitation outcomes among individuals with disabilities within OT; (ii) identify the types of AI-driven assessment tools reported in the literature, the target populations, functional domains or client populations they target, and the outcomes of their implementation; and (iii) highlight current advancements, benefits, limitations, and knowledge gaps to inform future research and development in AI-driven OT assessment and support. Ultimately, by clarifying the landscape of AI applications in OT assessments, this review seeks to guide practitioners and stakeholders in harnessing these emerging tools to enhance rehabilitation practices and facilitate improved functional recovery outcomes.

2. Materials and Methods

This study was registered on the Open Science Framework–OSF on 6 March 2025 [22]. The research adhered to the PRISMA-ScR 22-item checklist to ensure a thorough, consistent, and transparent review process [23]. By following the PRISMA-ScR framework, this study aimed to uphold high methodological standards, facilitate transparent reporting, and enhance the reliability and validity of its qualitative synthesis.

2.1. Literature Searches

The literature search was systematically conducted using PubMed and Scopus databases. This dual approach aimed to capture all relevant studies.

The search strategy was performed using the following keywords and phrases, applying the Title/Abstract filter in each database:

“machine learning” OR “deep learning” OR “artificial intelligence” AND assessment OR evaluation OR diagnosis AND “occupational therapy” OR “occupational therapists” OR ergotherapy.

2.2. Eligibility Criteria

2.2.1. Inclusion Criteria

Only peer-reviewed journal articles were considered to ensure quality and reliability. The review covered studies from 1 January 2015 to 8 March 2025, focusing on research involving the use of AI tools for assessment in the OT field. This criterion ensured that the review captured recent and relevant advancements in the field.

2.2.2. Exclusion Criteria

To focus on contemporary developments, articles published before 2015 were excluded. Additional exclusions encompass conference proceedings, books, non-English publications, studies unrelated to breath analysis, research involving non-human subjects, non-ML approaches, and non-diagnosis/assessment tasks. Review articles and studies with inaccessible full texts were also omitted to ensure thorough analysis and avoid redundancy.

2.3. Data Extraction

The study selection process was independently conducted by two reviewers, C.K. and I.K., who systematically screened titles, abstracts, and full texts for relevance. Identified studies were first compiled into a spreadsheet to remove duplicates. Titles and abstracts were then screened, followed by a full-text review. Any disagreements during this process were resolved through discussion. Once the final set of included studies was determined, data extraction was performed independently by the same two reviewers to ensure consistency and accuracy. A standardized data charting form, piloted by the review team prior to full extraction, was used to capture key variables consistently across studies. Studies were included in the qualitative synthesis if they met specific criteria regarding application domain, technological approach, data sources, sample characteristics, ML models used, validation strategies, and key findings. Descriptive synthesis was used to summarize charted data, identifying patterns and thematic trends across studies in terms of methodological quality, AI techniques, and OT contexts. This rigorous process ensured that the final selection of studies was both relevant and methodologically robust.

3. Results

An initial search identified 85 articles. After removing duplicates and conducting detailed screening, 26 studies were deemed potentially relevant. Applying the inclusion criteria narrowed this down to 13 articles included in the qualitative synthesis. The screening process is outlined in Figure 1, following PRISMA-ScR guidelines.

All included studies applied ML or DL techniques as part of AI-driven tools or models developed to support assessment processes in OT contexts. Studies were excluded if they (i) did not involve assessment, (ii) did not employ ML methods, (iii) were unrelated to OT tasks, or (iv) did not meet language or publication type criteria.

To systematically analyze the application of AI in OT, the included studies were categorized into four primary domains based on their focus areas: (i) Motor Recovery and Physical Rehabilitation, which includes studies utilizing AI to assess and predict motor function improvements in post-stroke and injury rehabilitation; (ii) Developmental and Pediatric Applications, focusing on AI-based assessment tools for children with developmental disabilities and learning challenges; (iii) Cognitive and Mental Health Applications, which explore AI’s role in evaluating engagement, mental health conditions, and cognitive impairments; and (iv) Assistive Technology and Telehealth, encompassing studies that integrate AI into remote therapy, wearable assistive devices, and adaptive technologies to enhance rehabilitation. These categories reflect the diverse ways AI is being implemented in OT, demonstrating its potential to improve assessment accuracy, optimize interventions, and expand accessibility to therapeutic services.

3.1. Motor Recovery and Physical Rehabilitation

Multiple studies leveraged AI techniques to evaluate and predict motor function outcomes in rehabilitation (Table 1). Zhao et al. (2024) developed a video-based model for hand function assessment in spinal cord injury patients, achieving about 55% accuracy in classifying hand prehension ability into five levels [24]. Similarly, Kim et al. (2020) employed a wearable soft robotic glove with ML to quantify finger movement in post-stroke patients; their SVM model differentiated motor impairment levels with roughly 80% accuracy (20% misclassification) [25]. Barth et al. (2023) explored predicting upper limb recovery after stroke using early clinical assessments and ensemble learning [26]. While their bagged-tree models fit the training data well, cross-validation accuracy was modest (~50%), highlighting the complexity of real-world outcome prediction. In assessing impairments, Park et al. (2019) showed an ANN could replicate clinician-rated spasticity grades with 82.2% agreement (κ = 0.743), demonstrating AI’s ability to objectively measure muscle tone [27]. These studies underscore AI’s role in enhancing motor rehabilitation by providing objective, quantitative assessments of upper-limb function, postural control, and neuromuscular impairments.

3.2. Developmental and Pediatric Applications

AI has also been applied to pediatric and developmental domains (Table 1). Chandran et al. (2024) achieved 95% accuracy using a random forest to predict learning developmental capability in children with special needs, effectively identifying “learning capacity disorder” levels [28]. Radhakrishnan et al. (2021) used DL on EEG data (spectrograms) to detect autism spectrum disorder, with a ResNet50 model yielding about 81–82% classification accuracy [29]. In a pediatric self-care assessment, Putatunda (2020) [30] and Zarchi et al. (2018) [31] developed models to classify self-care problems in children with physical disabilities, reporting accuracies of around 82–83%. Notably, Zarchi et al. introduced the SCADI dataset for standardized self-care activity classification [31]. Additionally, Fair-Field and Modayur (2025) demonstrated an AI approach for infant motor development screening: a support vector regression model predicted Alberta Infant Motor Scale (AIMS) scores with perfect sensitivity (1.0) and ~89.5% specificity, facilitating early detection of motor delays [32]. In another study, Ienaga et al. (2022) worked on the assessment of postural control based on the MediaPie Pose model [33]. Collectively, these studies illustrate that machine learning can assist in early diagnosis and evaluation in developmental pediatrics, from autism detection to functional skill assessment, providing clinicians with data-driven tools to identify needs and track developmental progress.

3.3. Cognitive and Mental Health Applications

In the cognitive and psychosocial realm, researchers have begun using AI to assess engagement and mental health conditions (Table 1). Suzuki and Suzuki (2023) developed a Classifier of Engagement in Occupation using natural language processing to quantify occupational engagement from context data [34]. By analyzing text (1554 tweets) with a BERT model, they could estimate individuals’ engagement levels, achieving a performance metric of about 0.76 (on a 0–1 scale, indicating moderately good classification fidelity). Huang et al. (2022) focused on psychiatric assessment, using DL on linguistic and acoustic features of speech to evaluate schizophrenia [35]. Their neural network model classified thought disorder severity with 88% accuracy and predicted symptom ratings on the PANSS scale with around 80% accuracy, demonstrating that AI can extract clinically relevant signals from speech. These examples highlight AI’s emerging role in cognitive and mental health within OT, from objectively measuring engagement in meaningful activities to supporting the evaluation of psychiatric disorders, suggesting that ML can augment traditional psychosocial assessments with quantitative insights.

3.4. Assistive Technology and Telehealth

One study exemplified the use of AI in assistive devices and telehealth for OT (Table 1). Ramírez-Sanz et al. (2023) introduced a low-cost telerehabilitation system for patients with Parkinson’s disease, employing computer vision (keypoint detection with an R-CNN model) to remotely assess physical exercises [36]. This proof-of-concept showed that pose-estimation algorithms can be applied in telehealth and HCI settings to monitor and evaluate rehabilitation performance, with reported system processing times on the order of seconds (approximately 7.5 s per session load). This study demonstrates how AI can extend OT beyond traditional settings, enabling remote monitoring of therapy via telehealth and enhancing the evaluation of assistive equipment, thereby improving accessibility and precision in intervention delivery.

Table 1. Characterization of included studies by application domain.

Author	Year	Application Domain	Subjects	Injury/Disease	Outcome	Best Machine Learning/Deep Learning	Validation	Results
Fair-Field and Modayur [32]	2025	Prediction of AIMS	SCOREIT pilot collected AIMS videos of 41 infants aged 3 to 15 months	N/A	AIMS score	Support Vector Regression (SVR)	Leave-one-out	Sensitivity of 1.0 and a specificity of 0.895
Zhao and Zariffa. [24]	2024	Hand Prehension Assessment	17 participants	Spinal cord injury (SCI)	GRASSP Prehension Performance subtest	SlowFast	Leave-One-Subject-Out cross validation (LOSO-CV)	55.10% Accuracy on 5 classes task
Chandran et al. [28]	2024	Capacity Disorder for Specially-Abled Children	92 Testing samples	Specially-Abled Children	Functional level	Random Forest	80% training/20% testing	95.38% Accuracy
Suzuki and Suzuki [34]	2023	Engagement in Occupation	1554 tweets	N/A	Engagement	Bidirectional Encoder Representations from Transformers (BERT)	Training, validation, and test datasets in the proportion of 8:1:1	0.763
Ienaga et al. [33]	2022	Postural Control Assessment	34 typically developing preschoolers and 23 adults	N/A		MediaPipe Pose	N/A	0.8
Kim et al. [25]	2020	Quantitative Measures on the Joint Paralysis	Ten stroke patients at Brunnstrom stage 3 and 4	Stroke	Finger paralysis	Support vector machine (SVM) and a k-means	N/A	A support vector machine revealed a misclassification rate of 20%
Barth et al. [26]	2023	Measurement of upper limb (UL) activity	54 subjects	early after stroke	UL performance	Bagged model	CV	100% Accuracy
Ramirez-Sanz [36]	2023	Assessing Physical Telerehabilitation	76 patients	Parkinson’s disease	Pose estimation	COCO person keypoint detection baselines with keypoint R-CNN models	N/A	7539.65 Loading (ms)
Huang et al. [35]	2022	Assessing Schizophrenia Patients	26 patients	Schizophrenia	Language and Communication (TLC), Positive and Negative Syndrome Scale (PANSS)	Neural Network	Leave one-out CV	TLS’s classification 88% accuracy and PANSS’s 80% accuracy
Radhakrishnan et al. [29]	2021	Detection of Autism Spectrum Disorder	10 typically developing children and 10 autistic children	Autism	Autism Diagnostic Interview-Revised (ADI-R) and Autism Diagnostic Observation Schedule (ADOS)	ResNet50	5-fold CV	81.91% Accuracy
Putatunda [30]	2020	Define and document the development, health and functioning of children	70 children	Disabled children	Self-care problems	Hybrid autoencoder-based method (autoencoders and deep neural networks)	k-fold CV	81.93% accuracy (7 classes)
Park et al. [27]	2019	Clinical Assessment of Spasticity	34 subjects with hemiplegia	Spasticity	Modified Ashworth Scale (MAE) of elbow flexors	Multilayer perceptron (MLP)	648 trials used for training	82.2% Accuracy
Zarchi et al. [31]	2018	Self-care problems classification	70 children	Children with physical and motor disability	Self-care problems	MLP	10-fold CV	83.1% Accuracy

AIMS: Alberta Infant Motor Scale; SCOREIT: Screening for Infant Motor Status; SVR: Support Vector Regression; GRASSP: Graded Redefined Assessment of Strength, Sensibility, and Prehension; SCI: Spinal Cord Injury; SlowFast: a DL model for video action recognition; LOSO-CV: Leave-One-Subject-Out Cross-Validation; BERT: Bidirectional Encoder Representations from Transformers; SVM: Support Vector Machine; UL: Upper Limb; CV: Cross-Validation; COCO: Common Objects in Context; R-CNN: Region-based Convolutional Neural Network; TLC: Thought, Language, and Communication Scale; PANSS: Positive and Negative Syndrome Scale; ADI-R: Autism Diagnostic Interview-Revised; ADOS: Autism Diagnostic Observation Schedule; ResNet50: Residual Neural Network with 50 layers; MLP: Multilayer Perceptron; N/A: Not Available.

Table 1. Characterization of included studies by application domain.

Author	Year	Application Domain	Subjects	Injury/Disease	Outcome	Best Machine Learning/Deep Learning	Validation	Results
Fair-Field and Modayur [32]	2025	Prediction of AIMS	SCOREIT pilot collected AIMS videos of 41 infants aged 3 to 15 months	N/A	AIMS score	Support Vector Regression (SVR)	Leave-one-out	Sensitivity of 1.0 and a specificity of 0.895
Zhao and Zariffa. [24]	2024	Hand Prehension Assessment	17 participants	Spinal cord injury (SCI)	GRASSP Prehension Performance subtest	SlowFast	Leave-One-Subject-Out cross validation (LOSO-CV)	55.10% Accuracy on 5 classes task
Chandran et al. [28]	2024	Capacity Disorder for Specially-Abled Children	92 Testing samples	Specially-Abled Children	Functional level	Random Forest	80% training/20% testing	95.38% Accuracy
Suzuki and Suzuki [34]	2023	Engagement in Occupation	1554 tweets	N/A	Engagement	Bidirectional Encoder Representations from Transformers (BERT)	Training, validation, and test datasets in the proportion of 8:1:1	0.763
Ienaga et al. [33]	2022	Postural Control Assessment	34 typically developing preschoolers and 23 adults	N/A		MediaPipe Pose	N/A	0.8
Kim et al. [25]	2020	Quantitative Measures on the Joint Paralysis	Ten stroke patients at Brunnstrom stage 3 and 4	Stroke	Finger paralysis	Support vector machine (SVM) and a k-means	N/A	A support vector machine revealed a misclassification rate of 20%
Barth et al. [26]	2023	Measurement of upper limb (UL) activity	54 subjects	early after stroke	UL performance	Bagged model	CV	100% Accuracy
Ramirez-Sanz [36]	2023	Assessing Physical Telerehabilitation	76 patients	Parkinson’s disease	Pose estimation	COCO person keypoint detection baselines with keypoint R-CNN models	N/A	7539.65 Loading (ms)
Huang et al. [35]	2022	Assessing Schizophrenia Patients	26 patients	Schizophrenia	Language and Communication (TLC), Positive and Negative Syndrome Scale (PANSS)	Neural Network	Leave one-out CV	TLS’s classification 88% accuracy and PANSS’s 80% accuracy
Radhakrishnan et al. [29]	2021	Detection of Autism Spectrum Disorder	10 typically developing children and 10 autistic children	Autism	Autism Diagnostic Interview-Revised (ADI-R) and Autism Diagnostic Observation Schedule (ADOS)	ResNet50	5-fold CV	81.91% Accuracy
Putatunda [30]	2020	Define and document the development, health and functioning of children	70 children	Disabled children	Self-care problems	Hybrid autoencoder-based method (autoencoders and deep neural networks)	k-fold CV	81.93% accuracy (7 classes)
Park et al. [27]	2019	Clinical Assessment of Spasticity	34 subjects with hemiplegia	Spasticity	Modified Ashworth Scale (MAE) of elbow flexors	Multilayer perceptron (MLP)	648 trials used for training	82.2% Accuracy
Zarchi et al. [31]	2018	Self-care problems classification	70 children	Children with physical and motor disability	Self-care problems	MLP	10-fold CV	83.1% Accuracy

AIMS: Alberta Infant Motor Scale; SCOREIT: Screening for Infant Motor Status; SVR: Support Vector Regression; GRASSP: Graded Redefined Assessment of Strength, Sensibility, and Prehension; SCI: Spinal Cord Injury; SlowFast: a DL model for video action recognition; LOSO-CV: Leave-One-Subject-Out Cross-Validation; BERT: Bidirectional Encoder Representations from Transformers; SVM: Support Vector Machine; UL: Upper Limb; CV: Cross-Validation; COCO: Common Objects in Context; R-CNN: Region-based Convolutional Neural Network; TLC: Thought, Language, and Communication Scale; PANSS: Positive and Negative Syndrome Scale; ADI-R: Autism Diagnostic Interview-Revised; ADOS: Autism Diagnostic Observation Schedule; ResNet50: Residual Neural Network with 50 layers; MLP: Multilayer Perceptron; N/A: Not Available.

3.5. Summary of Included Studies by Application Domain

Figure 2 presents a visual summary of the distribution of included studies across key OT application domains. The majority of studies (5 out of 13) focused on motor recovery and physical rehabilitation, underscoring AI’s growing role in assessing and predicting physical outcomes, such as upper limb recovery, spasticity levels, and postural control, particularly in stroke or spinal cord injury populations.

A substantial number of studies were dedicated to developmental and pediatric applications (five studies), reflecting emerging interest in using AI for early diagnosis and functional assessment in children with developmental delays, autism spectrum disorders, or self-care difficulties. This indicates a shift toward preventive and supportive interventions facilitated by data-driven tools in child-focused OT.

Cognitive and mental health applications were represented in two studies, emphasizing the use of AI for assessing psychiatric conditions, thought disorders, and occupational engagement through speech and language analysis. Though fewer in number, these studies demonstrate innovative uses of NLP and ML in domains traditionally dominated by subjective measures.

Finally, one study addressed assistive technology and telehealth, demonstrating AI’s potential to support remote rehabilitation, motion tracking, and patient monitoring—critical advancements for expanding OT access and continuity of care.

This distribution highlights both the current emphasis on physical rehabilitation and the emerging integration of AI in broader, more holistic OT practice areas.

4. Discussion

The scoping review identified 13 studies that applied AI techniques to various OT domains. Collectively, these studies demonstrate that AI can assist OTs in multiple ways–from automating assessment scoring to predicting patient outcomes. AI innovations, including ML, computer vision, and natural language processing (NLP), have been used to provide more precise assessments, create tailored interventions, and enhance rehabilitation progress monitoring. Most included studies reported positive results, suggesting that AI tools can improve evaluation accuracy or efficiency in areas such as motor function measurement, developmental screening, and assistive device optimization. While initial findings are encouraging, these AI applications remain largely in experimental stages, often tested on limited datasets under controlled conditions. This underscores the need for further research before AI-driven assessments can be widely implemented in everyday OT practice.

The scoping review also highlighted the evolution of AI applications in OT over the years. Early AI implementations (2018–2020) were primarily focused on motor recovery and physical rehabilitation, with models designed to classify motor impairments or predict rehabilitation outcomes. However, since 2021, AI applications have expanded into broader domains, including cognitive and developmental assessment and assistive technology. A noticeable shift toward pediatric and mental health applications emerged post-2020, reflecting the growing interest in AI’s role in screening developmental disorders and evaluating cognitive impairments. Additionally, AI has increasingly been integrated into assistive technologies, including telehealth platforms and wearable rehabilitation devices. This shift suggests that AI in OT, which initially concentrated on physical rehabilitation, has progressively expanded to include more holistic approaches, encompassing social participation and cognitive evaluation.

A review of the included studies revealed that many were conducted by multidisciplinary teams, often explicitly involving occupational therapists, especially in projects focused on self-care, cognitive assessment, and rehabilitation support. For example, occupational therapists played a key role in studies by Chandran et al. (2024, India) on learning capacity prediction [28], Suzuki and Suzuki (2023, Japan) on engagement classification [34], Putatunda (2020, India) with the Care2Vec model, and Ramírez-Sanz (2023, Spain) on telerehabilitation [36]. Other studies involved related professionals, such as physical therapists (e.g., Zhao et al. 2024, Canada [24]; Barth et al. 2023, USA [26]), or broader clinical experts, as seen in Kim et al. (2020, Korea) [25] and Huang et al. (2022, Taiwan) [35], where psychiatric nurses and therapists collaborated on AI models for schizophrenia assessment. Geographically, the studies spanned across Asia (India, Japan, Korea, Taiwan), North America (Canada, USA), and Europe (Spain), reflecting a growing global interest in AI-enhanced occupational therapy. While countries like Japan and Korea appear to focus on posture and motor function through pose estimation and robotics [25,33], others, such as India and Canada, contributed more in developmental, diagnostic, and classification tools. These variations underscore how local clinical priorities and research capacities may influence the focus and integration of AI in OT practices worldwide.

Furthermore, most of the included studies were classified as experimental and quantitative research, focusing on the development and evaluation of AI-based tools in occupational therapy. However, one study by Park et al. was specifically identified by the authors of the study as a prospective observational study, reflecting a more clinical, real-world data collection approach [27]. This variation highlights the methodological diversity across studies applying AI in OT contexts.

AI model performance varied widely across studies, with accuracy levels ranging from moderate (55%) to high (>90%), depending on the complexity of the task. While some models, such as DL systems for autism detection, achieved promising diagnostic accuracy (~82%), others, such as video-based hand function assessment models, reported only modest performance (F1 ≈ 0.55). Many studies achieved near-perfect accuracy on limited test sets, raising concerns about overfitting and lack of external validation. In general, AI models that focused on simpler, well-defined tasks, such as identifying motor impairment levels, demonstrated higher accuracy than those tackling more complex, multi-class classifications related to engagement or rehabilitation outcomes. Importantly, none of the reviewed studies provided conclusive evidence that AI systems outperform experienced clinicians in real-world settings. Instead, AI models were found to approximate clinical judgments and automate routine measurement tasks, reinforcing their role as supportive, rather than replacement, tools for OTs.

A key challenge identified in this scoping review is the validation of AI models. Most studies relied on internal validation techniques, such as cross-validation (CV) and holdout testing, due to limited sample sizes. While CV helps mitigate overfitting risks, it does not fully assess a model’s generalizability across diverse clinical populations. Notably, very few studies conducted external validation, such as testing on independent datasets or evaluating performance in multi-center trials. This raises concerns that AI models, despite promising initial results, may fail when applied in real-world OT practice. Additionally, outcome validation often relies on comparisons with standardized clinical assessments, which themselves have inherent variability.

Although direct comparisons between studies were not feasible due to variations in target populations, AI methods, and outcome measures, the trends observed in our scoping review align with broader findings from related reviews. For instance, prior work has emphasized the predominance of AI in motor rehabilitation and gait analysis, yet our review expands this by highlighting AI’s emerging use in cognitive, developmental, and engagement-related assessments specifically within occupational therapy. Unlike existing reviews that often focus on single clinical conditions or technologies, our findings demonstrate a broader, OT-specific adoption of AI across domains. This supports our assertion that AI is gradually evolving from specialized applications toward more holistic, occupation-centered assessments, reinforcing the need for interdisciplinary, clinically integrated tools that match OT’s person-centered practice.

The overall findings suggest that AI holds significant promise in OT, but its real-world applicability remains constrained by data limitations, validation challenges, and a lack of standardized implementation frameworks. The transition from experimental AI models to clinically integrated solutions will require addressing issues related to data diversity, external validation, and interpretability. Ensuring that AI tools are transparent, explainable, and aligned with clinical workflows will be crucial for widespread adoption in OT practice. As research progresses, AI’s role in OT is expected to expand further, providing data-driven, efficient, and personalized interventions that enhance therapeutic outcomes and support rehabilitation professionals.

4.1. Open Issues in Primary Studies

It is important to acknowledge the limitations present in the individual studies reviewed, as these open issues represent valuable opportunities for future research. First, most studies had small sample sizes and narrowly defined populations, which restricts generalizability. For instance, if an AI model is developed using data from a small group of stroke patients in a single hospital, its applicability to broader stroke populations, especially in different settings or severity levels, remains uncertain. Many studies were pilot projects or proofs-of-concept, often lacking statistical power or demographic diversity to ensure wide-ranging applicability.

Second, many studies relied heavily on specific clinical settings or specialized equipment, such as controlled laboratory conditions or sensor-based assessments. While these environments optimize data quality, they do not necessarily reflect real-world OT settings, where factors like lighting, background noise, and patient adherence vary. Consequently, AI models trained under such conditions may not generalize effectively to home-based rehabilitation or community healthcare contexts.

Third, a number of studies lacked long-term follow-up or repeated measures, making it difficult to assess the stability of AI-generated predictions over time. For example, if an AI model predicts post-stroke recovery trajectories, but there is no follow-up at later stages, we cannot determine whether those predictions accurately reflect real-world outcomes or if their accuracy deteriorates over time. Longitudinal validation is crucial to ensure that AI-powered assessments remain reliable across different rehabilitation phases.

Additionally, the lack of standardized protocols across studies complicates the comparison and synthesis of findings. Each study often defined its own assessment criteria, data collection methods, and AI evaluation metrics. This fragmentation prevents direct performance comparisons and hinders the establishment of best practices. As a result, some studies may have reported seemingly high accuracy due to specific dataset characteristics, while others tackling more complex real-world tasks may have shown lower performance. Without standardized benchmarking, drawing broad conclusions about AI effectiveness remains challenging.

Moreover, AI model explainability was often limited. Few studies provided insights into why the AI made certain predictions, making it difficult for occupational therapists to trust or validate the outputs. In clinical practice, transparency is essential: OTs need to understand the rationale behind AI-generated recommendations to make informed decisions that align with patient-centered care.

4.2. Limitations

At the scoping review level, certain methodological limitations must also be acknowledged. We included only English-language, peer-reviewed studies, which may have introduced publication bias. Studies with significant AI-driven improvements may be more likely to be published, while unsuccessful or inconclusive findings may remain unpublished. This could skew perceptions of AI effectiveness in OT. Additionally, we focused on explicitly OT-related research, meaning that relevant studies from adjacent fields (such as general rehabilitation or physiοtherapy) may not have been captured. Future reviews should consider broadening the inclusion criteria to encompass interdisciplinary AI applications. Another limitation is that rehabilitation studies specifically aligned with injury-related outcomes were excluded, potentially omitting valuable insights into AI applications in post-injury recovery and adaptation strategies. Expanding future searches to include gray literature, additional databases, and non-injury-related rehabilitation studies could provide a more comprehensive understanding of AI’s role in OT.

4.3. Future Directions

To overcome these limitations and advance AI’s role in OT, future research should prioritize several key areas. A critical next step is assembling larger, more representative datasets that encompass diverse OT clients in terms of age, culture, and medical history. Many of the current limitations stem from data scarcity. Multi-center collaborations, data-sharing initiatives, and the development of open-access repositories will be essential for ensuring that AI models are trained on large, heterogeneous samples rather than single-site data. This will improve model robustness, reduce overfitting, and enhance generalizability. As the field matures, validation must go beyond internal testing. Future research should prioritize prospective validation in real-world OT settings, such as integrating AI assessment tools into actual rehabilitation programs and evaluating their impact on clinical decision-making. Independent external validation using datasets from different sources is also needed to verify that AI models perform consistently across populations. Without external validation, even highly accurate models may fail when applied in new clinical environments.

AI models should be designed and tested with practitioners in mind. Occupational therapists must be actively involved in AI development to ensure that models align with OT ethics, workflows, and patient needs. AI should enhance—not replace—therapist expertise by providing actionable insights that therapists can interpret and adapt. Future research should assess user adoption factors, such as how easily OTs can integrate AI tools into daily practice and whether AI recommendations align with clinical reasoning. Many current AI models function as black boxes, generating predictions without clear explanations. Future research should focus on explainable AI (XAI) techniques that help OTs understand how predictions are made. For example, models that highlight which movement patterns led to a fall risk prediction will foster trust and allow therapists to validate or override AI-generated recommendations. Transparent models will be crucial for clinical acceptance and ethical AI integration in OT. OT assessments are inherently holistic, often combining physical, cognitive, and environmental factors. Future AI models should reflect this by integrating multimodal data sources–including motion tracking, speech analysis, sensor data, and patient-reported outcomes. Advances in sensor fusion, context-aware AI, and real-time analytics could enable more comprehensive and dynamic AI-assisted evaluations. As AI adoption increases, ethical frameworks and privacy safeguards must be developed. Future studies should explore best practices for ensuring secure data handling, informed consent, and transparency in AI decision-making. It will also be essential to monitor for unintended biases in AI models to ensure equitable access and outcomes for diverse populations.

5. Conclusions

OT is a person-centered profession that, with scientific evidence, focuses on engaging the individual in self-selected, meaningful, and purposeful activities and occupations. It promotes the health and well-being of the individual through activities of daily living, regardless of the level of functioning or symptoms of difficulty [4]. This scoping review contributes new knowledge by systematically mapping how AI methods—including machine learning, computer vision, and natural language processing—are currently being applied to OT assessments across various domains. AI techniques have been applied across various OT domains—including motor rehabilitation, activity monitoring, cognitive assessment, and assistive technology—with promising results. Early studies indicate that AI can automate components of assessment, such as scoring functional tests from video and predicting rehabilitation outcomes, offering objective insights that complement clinical expertise. Additionally, AI-driven systems can enhance decision-making by identifying subtle patterns and risk factors that might be difficult for therapists to detect manually. However, despite this potential, AI in OT remains in an early stage, with challenges related to validation, dataset diversity, and ethical considerations. This review highlights a clear need for occupational therapists to become involved in the development, implementation, and evaluation of AI tools, ensuring that these technologies align with the values of holistic, client-centered care. The key findings underscore both the promise and the current limitations of AI applications. To fully harness these technologies, continued refinement, rigorous evaluation, and user-centered integration will be essential. AI should not replace therapists’ expertise but serve as a supportive tool, enhancing precision, efficiency, and personalization in rehabilitation. Ultimately, the next decade will be crucial in translating these advancements into safe, effective, and ethically grounded AI applications that genuinely enrich OT practice. By embracing AI’s potential while addressing its limitations, occupational therapists can leverage these technologies to improve patient care, enabling timely, data-driven, and proactive interventions that empower individuals to engage more fully in meaningful activities.