Interpretable Multiclass Parkinson’s Disease Classification from Single-Camera Video Using Kinematic and Nonlinear Dynamic Features

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Abstract Parkinson’s disease (PD) is a progressive neurological disorder whose motor symptoms, especially gait impairment, change across disease stages and remain difficult to measure at scale. Motion capture and force plates provide accurate gait analysis but are expensive and impractical for routine clinical use. A major challenge is the lack of large, diverse public datasets, as most clinical gait datasets are proprietary and publicly available datasets are small, imbalanced with limited diversity. To address these challenges, we propose a single-camera, markerless framework for multiclass Parkinsonian gait classification that combines kinematic and nonlinear dynamic features from video-based pose estimation. Unlike prior work focused on binary PD detection or wearable sensors, our approach enables stage-aware gait assessment from standard video. To ensure fair evaluation with limited data, we use a strictly nested cross-validation pipeline where feature selection, preprocessing, and class balancing are performed only within repeated stratified training folds, preventing information leakage. An SVM with RBF kernel achieved the highest mean cross-validation accuracy (0.8366), while a Neural Network achieved similar accuracy (0.8338) with lower variation (std = 0.017). On the held-out test set, both models reached 0.9375 accuracy and 0.8222 macro-F1. A Random Forest provided an interpretable baseline with 0.8750 accuracy. Among the 19 selected features, 16 were nonlinear-dynamic measures, suggesting gait variability in this dataset is better captured by nonlinear dynamics than by kinematics alone, though larger studies are needed. These results demonstrate accurate, interpretable multiclass classification of Parkinsonian gait and provide a foundation for scalable, AI-assisted gait assessment in neuromotor healthcare.

Related topics