Artificial intelligence in stroke rehabilitation: A structured clinical review

Stroke is a major cause of long-term disability. Recovery can vary greatly in movement, thinking, communication, and everyday activities. This makes regular assessment and personalized rehabilitation important. However, rehabilitation decisions still depend heavily on clinical rating scales used at certain times, clinicians' experience, and the resources available in each area. These methods may not fully reflect differences in brain damage, a person's abilities before treatment, other health conditions, and how recovery changes over time. Access to rehabilitation is also limited by distance, uneven availability of services, and gaps between follow-up visits.

Artificial intelligence (AI) is increasingly being studied for brain imaging after stroke, early predictions of recovery, estimates of rehabilitation outcomes, movement and body-function assessment, robot-assisted and virtual reality training, brain-computer interfaces, wearable monitoring devices, and remote rehabilitation. These technologies may create new ways to provide care that adapts to patients and tracks their progress more continuously.

However, studies often discuss early diagnosis, predictions made soon after a stroke, and rehabilitation outcomes together, even though they address different clinical questions. Many AI tools for rehabilitation have only been tested using small datasets, data from the same study centers used to develop the models, or early studies of whether the tools are practical to use. Good technical performance does not necessarily mean that rehabilitation will improve. A structured review is therefore needed to show how AI applications relate to clinical decisions, how useful they are for rehabilitation, how thoroughly they have been tested, and how ready they are for use in clinical practice.

Weihua Song, author, researcher, Capital Medical University

This study used a structured review of existing research. The researchers systematically searched PubMed/MEDLINE, IEEE Xplore, Web of Science, and Scopus for studies on AI and stroke rehabilitation published between January 2010 and March 2025. They used a process based on PRISMA guidelines to make their methods clearer. The searches found 1,247 records. After removing duplicate records and screening titles, abstracts, and full papers, the researchers included 108 publications in their review. They collected information about study design, AI methods, sample sizes, rehabilitation areas, outcomes, model performance, validation methods, and reported limitations. They grouped the evidence into four main topics: AI for predicting rehabilitation outcomes; rehabilitation assessment and monitoring; treatment and care management; and the use of AI with robotics, virtual reality, brain-computer interfaces, wearable technologies, and remote rehabilitation. Importantly, the review separates evidence about early diagnosis or prognosis from evidence about rehabilitation outcomes. It also examines each application based on the quality of its testing, its clinical value, and the barriers to putting it into practice.

One major use of AI in stroke rehabilitation is predicting how a person's abilities may recover over time. These predictions can use clinical rating scales, medical images, electroencephalography (EEG), data from wearable sensors, and demographic information. Studies have explored predictions of functional ability 90 days after stroke, changes in Barthel Index scores, movement abilities, language and thinking skills, quality of life, and long-term risk of death. Researchers have used machine learning, deep learning, and explainable AI models to identify factors linked to different recovery patterns. Imaging and early prediction models can provide information about the amount of brain damage, the severity of the stroke at the beginning, and the response to treatment. This information may help guide later rehabilitation planning. Predictions designed specifically for rehabilitation focus more directly on changes in movement, everyday activities, participation, and outcomes reported by patients over time. However, most studies look back at existing records, involve small groups or single centers, and test models mainly using data from the same setting where they were developed. Independent testing across multiple centers and studies that measure the real effects on clinical care are still uncommon. Accurate predictions may help identify patients at different levels of risk, but they do not by themselves prove that AI-guided treatment improves rehabilitation outcomes.

AI is increasingly being used to automate the assessment and ongoing monitoring of movement, posture, walking, and rehabilitation progress after stroke. These tools can add to traditional clinical rating scales and occasional follow-up visits. Current systems combine EEG, electromyography (EMG), wearable sensors, motion tracking, and data collected at home to measure movement patterns, the quality of exercises, and changes in function. Machine learning can also detect unusual movement patterns or estimate clinical scores for movement ability. Wearable devices and remote monitoring can extend assessments into the home, providing information over time about whether patients follow their rehabilitation plans and how their abilities change. However, many studies mainly examine whether the technology works, how accurately it classifies data, or how it performs on short-term measures that stand in for meaningful recovery. They often include small groups of patients, data from a single center, and little independent testing at other sites. Differences in devices, data formats, and assessment schedules also make it harder to apply the results to other settings. Future research should use more consistent outcome measures and data-processing methods, along with long-term, prospective studies across multiple centers. These studies should determine whether AI-based monitoring can improve rehabilitation decisions and patient outcomes in meaningful ways.

Beyond prediction and monitoring, AI is increasingly being studied as a way to adjust rehabilitation and coordinate different assistive technologies after stroke. Current methods aim to change training intensity, task difficulty, and how often feedback is provided based on movement performance, thinking abilities, and information from sensors. They may combine AI with rehabilitation robots, virtual reality, brain-computer interfaces (BCIs), wearable devices, and remote rehabilitation platforms. Robots can provide consistent help with movement. Virtual reality supports interactive practice and gives patients additional feedback. BCIs can use signals from the brain to control external feedback or assistance when a person has very limited ability to move voluntarily. However, these technologies are currently used mainly to support rehabilitation rather than replace standard care. Much of the evidence comes from small groups, short follow-up periods, and early studies that explore whether a treatment is feasible or test it in a single group without a comparison group. Strong evidence that these technologies work better than conventional rehabilitation remains limited. An important future goal is to develop patient-focused systems that continuously adjust treatment based on feedback. In these systems, EEG, EMG, and wearable-sensor signals could help change the level of robotic assistance, tasks using extended reality, and feedback delivered remotely. Such systems would still need clinical supervision and would have to work with other technologies, remain affordable, and fit into routine rehabilitation services.

Despite the growing potential of AI to predict recovery, assess progress, and support rehabilitation after stroke, several problems still make it difficult to use these tools in clinical practice. Rehabilitation datasets are often small, differ greatly from one another, and collect information at inconsistent times. It is also difficult to standardize different types of clinical information and sensor data. These problems increase the risk that models will learn patterns that do not apply to new patients and make it harder to use them reliably across different hospitals and institutions. In addition, complex AI models can work like a "black box," making it difficult for clinicians to understand how they reach their conclusions. This can reduce trust in the models. Better prediction accuracy alone does not guarantee better treatment decisions or improved patient outcomes.

Wearable monitoring, home-based rehabilitation, and remote-care platforms also raise concerns about privacy, data security, information sharing, and whether devices work reliably. Future research should focus on independent testing across multiple centers and prospective clinical trials. Researchers should also develop rehabilitation systems that continuously adjust treatment using clinical information, medical images, EEG, EMG, and wearable-sensor data to change robotic assistance, extended-reality training, and remote feedback as needed.

"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.

The paper's authors include Weihua Song, Shuai Si, Yinghui Wang, Hong Chang, and Xiaoying Tang.