Video-Based Data-Driven Models for Diagnosing Movement Disorders: Review and Future Directions

Movement disorders are abnormal, involuntary movements that can heavily impact a person's quality of life. In clinical practice, diagnosis and severity assessments rely mainly on visual clinical inspections (ie, on subjective expert opinion). With clinical videos commonly acquired as part of examinations, novel data-driven models have emerged that use machine learning (ML) and deep learning (DL) algorithms to capture human actions and recognize their characteristics, showing promise as new tools in clinical workflows. This review seeks to provide a comprehensive examination of video-based, data-driven models for movement disorders, including tremor, dystonia, myoclonus, chorea, tics, Parkinson's disease, and ataxia. We explore literature from 2006 to 2024 in a variety of scientific databases, with different video modalities including red-green-blue video of different frame rates, depth video, marker-based approaches, multi-perspective approaches, and multimodal video. We discover a significant trend in studies favoring pose estimation methods, with newer studies incorporating real-time methods and end-to-end DL architectures, and usability is steadily increasing and rapidly approaching expert-level performance. Likewise, we present the main limitations in the current approaches, such as limited public sources of data, lack of standardized metrics, and patient privacy. Taking inspiration from other fields, in medicine and otherwise, we propose possible future research directions including explainable artificial intelligence techniques, privacy-preserving devices and modeling techniques, and better metric guidelines. © 2025 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.