Objective: To highlight the scope of using Artificial Intelligence (AI) for generating videos of movement disorders phenomenology for medical education in movement disorders

Background: Familiarizing medical students and neurology trainees with the different phenomenology of movement disorders is crucial to medical training (1). Relying solely on clinical encounters may not be practical, as not all types of cases would be seen in clinics. To address this, videos of real patients with movement disorders have traditionally been used for teaching, helping trainees recognize both common and rare presentations. However, this also raises important questions about patient privacy and ethical considerations. (2) Generative AI can simulate these videos and replicate movement disorder phenomenology.

Method: We used two generative AI platforms – Sora Turbo (OpenAI, Inc., San Francisco, California, United States) and Veo 2 (Google LLC, Mountain View, California, United States) – for text-to-video generation in a stepwise manner. Initially, the AI engine was fed text prompts to generate videos of common movement disorder phenomenologies. For Sora, in a second step, it was trained by uploading videos of real patients with identities hidden and asked to analyze these to generate simulated videos of the same phenomenology. A text prompt also asked for a graph similar to accelerometry that reflected the movement amplitude and rhythmicity against time. Different phenomenologies were used with varying levels of complexity. These videos were analyzed by a movement disorder specialist to judge for accuracy of the presentation.

Results: The initial videos generated by both the AI platforms were found to be an inaccurate representation of the given phenomenology. AI ‘hallucinations’ were noted with random and bizarre movements. Google Veo2 had better text-to-video results in terms of accuracy of the movement phenomenology. Sora could be ‘trained’ with the real videos of the given phenomenology, and the subsequent video and accelerometry graph generated were found to more accurately represent the movement disorders’ classic characteristics. When a more complex phenomenology was requested, for eg, Myoclonus-dystonia, even after training, the video generated was found to be inaccurate.

Conclusion: AI may have a role in education in movement disorders, which allows for protecting patient privacy. However, it needs further development for more accuracy and applicability.

References: 1. Jog, Mandar S., and Linda Grantier. “Methods for Digital Video Recording, Storage, and Communication of Movement Disorders.” Movement Disorders, vol. 16, no. 6, 2001, pp. 1196–200, https://doi.org/10.1002/mds.1199.
2. Zainal Abidin H, Razali HYH. Ethical deliberations on video recording of patients in healthcare facilities- a scoping review. Med J Malaysia. 2024 Nov;79(6):785-793. PMID: 39614799.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/artificial-intelligence-for-education-in-movement-disorders/