Table 1.

Summary of studies included in the research.

Author	Aim	Study Design/Intervention	Treatment Period	Sample Size	Outcomes Measures	Main Findings
Murakami et al. 2023 [54]	To evaluate how using a robot hand integrated with AI and EMG technology affects upper extremity rehabilitation in chronic stroke patients.	Randomized Controlled Trial.	4 weeks.	20 patients.	FMA, MAL-14 AOU, MAS; H reflex and reciprocal inhibition.	The group that actively participated in the intervention demonstrated notable enhancements in FMA, MAL-14 AOU, and wrist MAS immediately after the intervention and also four weeks after. There were no notable enhancements observed in FMA for the control group.
Rupprechter et al. 2021 [55]	To assess a new computer vision technique using deep learning to measure the degree of walking problems in PD.	Methodological development study.	The study did not apply to treatment intervention, as it concentrated on creating and evaluating a gait assessment method.	The study utilized footage from 729 gait evaluations in which trained clinicians gave ratings.	The model’s ability to predict gait severity ratings were compared to clinician ratings, and the model’s predictions were also correlated with manual ratings.	The computer vision model achieved an accuracy of 50%, accurately estimating UPDRS ratings within one point of clinician ratings in 95% of cases. The model’s predictions showed a strong correlation with clinician diagnoses.
Yang et al. 2022 [56]	To create and assess an AI model to identify PD and monitor how it advances through analyzing night-time breathing patterns.	Development and evaluation study.	The study does not involve any treatment; rather, it focuses on the development and evaluation of AI models.	Data from 7671 individuals, encompassing information from various hospitals and multiple public datasets, was used to assess the model.	The AI model was evaluated based on its capacity to identify PD and to gauge the severity and advancement of PD.	The AI model can reliably identify PD and forecast its severity and progression. An attention layer is used for explainability and is capable of conducting remote PD assessments in homes without physical contact utilizing radio waves.
Gandolfi et al. 2023 [57]	To assess if ML can effectively forecast the recovery of UL function in patients recovering from sub-acute strokes and to pinpoint the key factors influencing these forecasts utilizing XAI techniques.	Retrospective study.	Patients received intensive, multidisciplinary upper limb rehabilitation for 2 h every day, 6 days a week, throughout their hospitalization. The mean period from stroke onset to release was around 37.71 days.	The ultimate dataset included 95 entries from a starting group of 192 individuals.	FMA-UE, TCT, MI, BI.	ML models outperformed standard statistical approaches in predicting UL recovery and the development of the illness. Baseline motor impairment was the most important characteristic. XAI techniques delivered reliable and clear findings, improving the comprehension of predictive variables.
Moobs et al. 2024 [58]	To determine the effectiveness of a novel two-tier ML model in detecting aberrant arm motions during walking in people with ABI.	Observational study.	Not specified.	42 ABI participants and 34 healthy controls.	Concordance between ML model predictions and clinician evaluations.	The ML model predictions were in close concordance with those of experienced human assessors, with no statistically significant variances between the networks. The models did not accurately forecast scores with minor impacts.
Varghese et al. 2024 [59]	To create reliable ML models for detecting and monitoring movement disorders using smart devices due to the lack of comprehensive datasets containing both movement data and clinical annotations for such disorders.	Cross-sectional study.	3 years.	504 participants, including individuals with PD, DD, and HC.	The outcome measure included the balanced accuracy of ML models in distinguishing between PD vs. HC and PD vs. DD, along with the detailed collection of clinical annotations and movement data.	The ML models obtained a mean balanced accuracy of 91.16% for distinguishing between PD and HC and 72.42% for distinguishing PD from DD. The research emphasizes the efficiency of the models but also acknowledges difficulties in differentiating between similar disorders.
Yoo et al. 2024 [60]	To forecast the restoration of walking ability post-SCI upon leaving a rehab center, utilizing ML methods to analyze crucial predictive factors and propose an ML-driven tool to aid in predicting gait recovery.	Retrospective Study.	Information was gathered between June 2008 and December 2021.	353 patients with traumatic or non-traumatic SCI.	The primary outcome was the FAC_DC.	The prediction of FAC_DC was accurate using random forest and decision tree algorithms, yielding RMSE values of 1.09 and 1.24 for all participants, 1.20 and 1.06 for traumatic SCI, and 1.12 and 1.03 for non-traumatic SCI. The primary factor for predicting gait recovery was found to be the initial FAC.
Hossain et al. 2023 [61]	To assess how stroke survivors perceive their body position using a robotic arm matching task and to evaluate the effectiveness of various ML methods and a task score in distinguishing between stroke survivors and non-stroke individuals based on movement data.	Cross-sectional study.	Not specified.	429 individuals who have had a stroke confirmed by neuroimaging (less than 35 days after the stroke) and 465 healthy individuals.	Parameters like trial-to-trial variability, spatial contraction/expansion ratio, systematic spatial shifts, and absolute error were used to measure performance in the arm position matching task. Task scores were additionally computed to evaluate overall effectiveness.	For the ML and deep learning models, the classification performance metrics were as follows: accuracy 82.4%, precision 85.6%, recall 76.5%, and F1 score 80.6%. Random Forest surpassed all other models in terms of numerical accuracy, scoring 83%. Both sensitivity and specificity were higher for ML models compared to the overall task score. Variability was the most dominant feature in classifying performance in the arm position matching task.

Legend: artificial intelligence (AI), electromyography (EMG), upper extremity (UE), Fugl–Meyer assessment (FMA), motor activity log-14 amount of use score (MAL-14 AOU), modified Ashworth scale (MAS), Parkinson’s disease (PD), Unified Parkinson’s Disease Rating Scale (UPDRS), machine learning (ML), upper limb (UL), explainable artificial intelligence (XAI), upper-extremity score on the Fugl–Meyer Assessment (FMA-UE), Trunk Control Test (TCT), Motricity Index (MI), Barthel Index (BI), acquired brain injury (ABI), differential diagnoses (DD), healthy controls (HC), spinal cord injury (SCI), decision support system (DSS), functional ambulation category at discharge (FAC_DC), root mean squared error (RMSE).