Artificial intelligence in stroke rehabilitation: A structured clinical review

Stroke is a major cause of long-term disability, and recovery varies substantially across motor, cognitive, communication, and activities-of-daily-living domains, making continuous assessment and individualized rehabilitation essential. However, current rehabilitation decisions still rely heavily on periodic clinical scales, clinician experience, and local resources, which may not fully capture differences in lesion characteristics, baseline function, comorbidities, and recovery trajectories. Access is also limited by geographic barriers, uneven service capacity, and intermittent follow-up.

Artificial intelligence is increasingly being explored for stroke imaging, early prognosis, rehabilitation-outcome prediction, movement and physiological assessment, robot- and virtual-reality-assisted training, brain–computer interfaces, wearable monitoring, and telerehabilitation, offering new opportunities for more adaptive and continuous care.

Yet acute diagnosis, early prognostication, and rehabilitation-specific outcomes are often discussed together despite addressing different clinical questions, and many rehabilitation applications remain supported mainly by small datasets, internal validation, or early feasibility studies. Technical performance therefore should not be equated with rehabilitation benefit. This creates a need for a structured review that maps AI applications to their clinical decision points, rehabilitation relevance, validation maturity, and readiness for clinical translation."

Weihua Song, author, researcher, Capital Medical University

This study used a structured narrative review design, systematically searching PubMed/MEDLINE, IEEE Xplore, Web of Science, and Scopus for studies on artificial intelligence and stroke rehabilitation published from January 2010 to March 2025, with a PRISMA-style process used to improve transparency. The database searches identified 1,247 records, and after deduplication, title and abstract screening, and full-text assessment, 108 publications were included in the thematic synthesis. The reviewers extracted information on study design, AI technique, sample size, rehabilitation domain, outcomes, model performance, validation methods, and reported limitations. Evidence was organized into four major themes: AI for rehabilitation-outcome prediction, rehabilitation assessment and monitoring, therapeutic intervention and management, and integration of AI with robotics, virtual reality, brain–computer interfaces, wearable technologies, and telerehabilitation. Importantly, the review distinguishes acute diagnostic or prognostic evidence from rehabilitation-specific outcomes and evaluates each application according to validation quality, clinical relevance, and implementation barriers

One of the major applications of artificial intelligence in stroke rehabilitation is the prediction of subsequent functional recovery using clinical scales, imaging, EEG, wearable-sensor data, and demographic information. Existing studies have explored prediction of 90-day functional status, changes in Barthel Index, motor-function scores, language and cognitive recovery, quality-of-life improvement, and long-term mortality, using machine learning, deep learning, and explainable models to identify factors associated with recovery trajectories. Acute imaging and early prognostic models can provide baseline information on lesion burden, initial severity, and treatment response that may inform later rehabilitation planning, whereas rehabilitation-specific prediction focuses more directly on changes in motor function, activities of daily living, participation, and patient-reported outcomes over time. However, most studies remain retrospective, small, or single-center and rely mainly on internal validation, while independent multicenter external validation and prospective clinical-impact studies are still uncommon. Therefore, strong predictive accuracy indicates potential for risk stratification, but does not by itself demonstrate that AI-guided treatment improves rehabilitation outcomes.

Artificial intelligence is increasingly being used to automate the assessment and continuous monitoring of motor function, posture, gait, and rehabilitation progress after stroke, complementing conventional clinical scales and periodic follow-up. Existing systems combine EEG, EMG, wearable sensors, motion tracking, and home-monitoring data to quantify movement patterns, exercise quality, and functional changes, while machine learning can identify abnormal movements or estimate clinical motor scores. Wearable and remote-monitoring technologies can further extend assessment into the home, providing longitudinal information on rehabilitation adherence and progression. However, many studies still focus primarily on technical feasibility, classification accuracy, or short-term surrogate outcomes, and commonly involve small samples, single-center datasets, and limited independent external validation. Heterogeneity in devices, data formats, and assessment schedules also restricts generalizability. Future work should therefore establish more standardized outcome measures and data pipelines and use multicenter, long-term, prospective studies to determine whether AI-based monitoring can meaningfully improve rehabilitation decisions and patient outcomes.

Beyond prediction and monitoring, artificial intelligence is increasingly being explored to support rehabilitation adjustment and the coordinated use of assistive technologies after stroke. Current approaches aim to adapt training intensity, task difficulty, and feedback frequency according to motor performance, cognitive status, and sensor-derived information, while integrating AI with rehabilitation robots, virtual reality, brain–computer interfaces, wearable devices, and telerehabilitation platforms. Robotic systems can provide standardized assisted movement, virtual reality emphasizes interactive task practice and augmented feedback, and BCIs can use neural signals to control external feedback or assistance when voluntary movement is severely limited. However, these technologies currently function mainly as rehabilitation adjuncts, with much of the evidence coming from small samples, short follow-up, and exploratory or single-arm studies; convincing evidence of superiority over conventional rehabilitation remains limited. A key future direction is the development of patient-centered closed-loop systems in which EEG, EMG, and wearable-sensor signals dynamically adjust robotic assistance, extended-reality tasks, and remote feedback while maintaining clinical supervision, interoperability, affordability, and compatibility with routine rehabilitation workflows.

Despite the growing potential of artificial intelligence for stroke rehabilitation prediction, assessment, and assisted intervention, clinical translation remains constrained by challenges in data, modeling, and implementation. Rehabilitation datasets are often small, heterogeneous, and inconsistently sampled, while multimodal clinical and sensor data remain difficult to standardize, increasing the risk of overfitting and limiting cross-institutional generalizability. In addition, the "black-box" nature of complex models can reduce clinical interpretability and trust, and improved predictive accuracy alone does not guarantee better treatment decisions or patient outcomes.

Wearable monitoring, home rehabilitation, and remote-care platforms also raise concerns regarding privacy, security, data sharing, and device reliability. Future research should emphasize multicenter external validation and prospective clinical trials while developing genuinely closed-loop rehabilitation systems that combine clinical information, imaging, EEG, EMG, and wearable-sensor data to dynamically adjust robotic assistance, extended-reality training, and remote feedback.

"Greater attention to explainability, interoperability, cost-effectiveness, and fairness will also be essential, with AI positioned as an adjunct to multidisciplinary care rather than a replacement for clinicians." said Weihua Song.

Authors of the paper include Weihua Song, Shuai Si, Yinghui Wang, Hong Chang, and Xiaoying Tang.