Related Experiment Video
Updated: Aug 6, 2026
06:52
Deep-Learning Based Multi-Joint Synchronous Tracking for Objective Quantification of Hindlimb Locomotor Kinematics in Rats
Published on: April 3, 2026
Deep Learning Pose Estimation for Phenotyping of Co-Occurring Hyperkinetic Movement Disorders
Laura Cif1,2, Diane Demailly2,3, Gabriella A Horvàth4
1Service of Neurology, Department of Clinical Neurosciences, Lausanne University Hospital (CHUV), University of Lausanne (UNIL), Lausanne, Switzerland.
Annals of Clinical and Translational Neurology
|July 25, 2026
Summary
Deep learning pose estimation combined with kinematic features shows promise for identifying multiple co-occurring hyperkinetic movement disorders (HMDs) in patients. Further validation is needed for clinical use.
Area of Science:
- Neurology
- Biomedical Engineering
- Artificial Intelligence in Medicine
Background:
- Hyperkinetic movement disorders (HMDs) often co-occur, complicating diagnosis and management.
- Accurate phenotyping is crucial for understanding HMDs and developing targeted treatments.
- Current phenotyping methods can be subjective and time-consuming.
Purpose of the Study:
- To investigate the feasibility of using routine outpatient videos with deep learning-based pose estimation for multi-label phenotyping of co-occurring HMDs.
- To assess the performance of clinically interpretable kinematic features derived from video analysis in classifying HMDs.
- To establish a proof-of-concept for an automated HMD phenotyping pipeline.
Main Methods:
- Exploratory single-center study using video data from 21 HMD patients and 4 healthy controls.
- Markerless pose estimation (YOLOv8) to extract 2D keypoint trajectories.
- Transformation of trajectories into statistical, temporal, spectral, and complexity-based kinematic features.
- Supervised classification models trained on extracted features for multi-label HMD phenotyping.
Main Results:
- The best-performing pipeline achieved a macro-AUPRC of 0.717 ± 0.030 and macro-AUROC of 0.767 ± 0.069 under nested cross-validation.
- Patient-level classification demonstrated a 76.5% agreement between predicted and actual labels.
- An exploratory analysis suggested an upper bound of 86.0% for patient-label agreement with post-hoc pipeline selection.
Conclusions:
- A hybrid approach combining deep learning pose estimation and feature-engineered classification shows encouraging results for multi-label phenotyping of co-occurring HMDs.
- The findings represent a proof-of-concept, highlighting the potential of automated video analysis in HMD assessment.
- External, multicenter, prospective validation is essential prior to clinical implementation or use in trials.
