Deep Learning Pose Estimation for Phenotyping of Co-Occurring Hyperkinetic Movement Disorders

To explore whether routine outpatient video combined with deep learning-based pose estimation and clinically interpretable kinematic features can support multi-label phenotyping of co-occurring hyperkinetic movement disorders (HMDs).

In this exploratory single-centre proof-of-concept study, videos from 21 patients with HMDs and 4 healthy controls were processed with markerless pose estimation (YOLOv8) and 2-dimensional keypoint trajectories transformed into kinematic descriptors spanning statistical, temporal, spectral, and complexity domains. Ten-second windows were aligned to expert annotations for eight hyperkinetic phenomenologies. Conventional supervised classifiers were trained on these tabular descriptors. Window-level predictions were aggregated to the patient level, and label-specific thresholds were tuned on training participants only.

In patient-level multi-label performance reporting, (i) the best single pipeline selected by discrimination (StandardScaler + MLP) achieved a macro-AUPRC of 0.821 ± 0.019 and a macro-receiver operating characteristic area under the curve of 0.830 ± 0.029. (ii) The best single pipeline selected by Hamming accuracy (MinMaxScaler + SVM) reached 0.764 ± 0.041. (iii) Under prespecified nested cross-validation with per-label model selection within training folds (primary analysis), macro-AUPRC was 0.717 ± 0.030, macro-AUROC was 0.767 ± 0.069, Hamming accuracy was 0.764 ± 0.014 and patient-label agreement was 153/200 (76.5%). (iv) Post hoc per-label selection of the best-performing pipeline defined an exploratory upper bound of 172/200 (86.0%).

In this exploratory study, a hybrid pipeline combining deep learning pose estimation with feature-engineered supervised classification produced encouraging patient-level multi-label performance for co-occurring HMDs. These findings are proof-of-concept; external, multicentre, prospective validation is required before clinical or trial use.