Comprehensive Summary
This literature review examined the uses of artificial intelligence (AI) and machine learning (ML) in robotic-assisted rehabilitation in both movement and neurocognitive domains. The authors conducted a search using Scopus and IEEE Xplore, and screened studies that included human participants, ultimately including 201 studies. The studies were organized by the purpose of introducing AI/ML (i.e., movement detection, motion prediction, patient outcomes assessment, compensation detection, and individualized rehabilitation), algorithms used, and input data type. The review indicated ML was largely effective in predicting movements, controlling robotic exoskeletons, classifying gait and hand gestures, and predicting outcomes and recovery in clinical settings with neural networks and deep learning approaches being the most common. The application of the AI/ML resulted in poorer performance on studies that included patient populations less represented in studies (i.e., underrepresented such as children and older adults). Key gaps were mentioned including explainability of the AI/ML, generalization of AI/ML to differing clinical populations, and use of minimal examples having limited reproducibility because of inadequate data and code sharing. In conclusion, while the promise of AI/ML is a strong development toward rehabilitation with robotics, the unique barriers of inclusion, transparency, and validation of AI/ML must be addressed to advance the field.
Outcomes and Implications
This work is significant because it outlined how AI-enabled robotics can change rehabilitation through providing specific, adaptive, personalized therapy which is seen as a distinct advancement in rehabilitation practice compared to traditional methods. Clinically, the review emphasized the potential of machine learning to improve motor recovery, predict motor outcomes and assist real-time changes in therapy, possibly accelerating patients with stroke, cerebral palsy, and neurocognitive disorders toward more functional independence. While clinically applicable models were discussed, it was also recognized that a bulk of the models discussed in the literature have been validated on limited, small or non-representative samples and thus have limited applicability to the clinical space. For these AI and robotic-powered models to be considered "meaningful" in practice, the models must be applied and adapted to more representative patient populations, involve explainable AI algorithms, and be integrated with other electronic health record systems to provide holistic patient care. Although we are still years away from routine application as standard rehabilitation practice, the pace of growth in AI and robotics development, along with the rapidly evolving regulatory frameworks governing the use of healthcare AI shows that personalized AI-enabled rehabilitation therapies could in fact be a regular component of neurorehabilitation practice in the foreseeable future.