New AI diagnoses 8 movement disorders from pediatric clinic videos
AI achieves 84% accuracy detecting dystonia, tremor, tics from standard recordings without markers
A multinational team led by Laura Cif and colleagues has built a deep learning framework that can simultaneous phenotype hyperkinetic movement disorders from standard video recordings, eliminating the need for specialized motion-capture equipment. The system combines markerless pose estimation (extracting body keypoints from video) with kinematic descriptors and a pretrained tabular foundation model. Initial training was performed on 21 adult patients with confirmed hyperkinetic disorders and 4 healthy controls under a standardized protocol.
External validation used a real-world pediatric cohort of 12 children with monogenic combined movement disorders. Without retraining the core model, only a lightweight calibration of the final decision layer was applied using a small clinician-selected subset (5 patients). On the held-out 7 children, Hamming accuracy rose from 80.4% to 83.9%, and the Jaccard index improved from 54.8% to 63.3%. For disorders with high clinician agreement, performance reached 90% Hamming accuracy and 78.6% Jaccard index, demonstrating the framework's viability for clinical deployment.
- Detects 8 disorders simultaneously from routine videos without markers: dystonia, tremor, myoclonus, chorea, athetosis, ballismus, stereotypies, tics
- Cross-cohort transfer from 21 adults to 12 children using only 5 clinician-labeled cases for calibration
- Achieved 90% Hamming accuracy and 78.6% Jaccard index on high-agreement diagnoses
Why It Matters
Enables scalable, objective movement disorder diagnosis from standard clinic videos, reducing specialist workload and improving pediatric care access.