Automatic Infant Movement Assessment Using Pose-LBP Features and a Cost-Sensitive Subspace kNN Ensemble

Abstract

Background/Objectives: Assessment of infant General Movements (GMs) is essential for early detection of neurological disorders such as cerebral palsy, but current methods depend on expert interpretation. This study proposes an automated and interpretable framework for infant movement classification using pose-based representations from RGB videos. Methods: A pose-driven pipeline was developed to extract 2D skeletal key points using a two-stage tracking strategy. Joint coordinates were normalized using the shoulder center and inter-shoulder distance. Videos were segmented into overlapping temporal windows, and each segment was represented using Pose-LBP histograms and motion ratio features. Classification was performed with a cost-sensitive subspace k-nearest neighbor ensemble (CSS-kNN-E). Performance was evaluated using stratified 10-fold cross-validation on a five-class infant movement dataset. Results: The proposed method achieved 99.16% (±0.48%) accuracy, 99.19% (±0.50%) sensitivity, 99.76% (±0.13%) specificity, and 99.23% (±0.48%) F1-score. The model demonstrated strong discrimination across classes and robustness to class imbalance. Conclusions: The framework provides an accurate and scalable solution for automated infant movement analysis. It reduces dependency on expert evaluation and has strong potential for early clinical screening and decision support.