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

Area of Science:

• Biomedical engineering
• Computer vision
• Developmental pediatrics

Background:

• Infant General Movements (GMs) assessment is crucial for early detection of neurological disorders, including cerebral palsy.
• Current GM assessment methods rely heavily on subjective expert interpretation, limiting scalability and consistency.

Purpose of the Study:

• To develop an automated and interpretable framework for classifying infant movements using pose-based representations from RGB videos.
• To reduce dependency on expert evaluation in infant movement analysis for early clinical screening.

Main Methods:

• A pose-driven pipeline extracted 2D skeletal key points from videos using a two-stage tracking strategy.
• Joint coordinates were normalized, and videos were segmented into temporal windows for feature extraction using Pose-LBP histograms and motion ratio.
• Classification was performed using a cost-sensitive subspace k-nearest neighbor ensemble (CSS-kNN-E), evaluated via 10-fold cross-validation.

Main Results:

• The automated framework achieved high performance metrics: 99.16% accuracy, 99.19% sensitivity, 99.76% specificity, and 99.23% F1-score.
• The model demonstrated robust discrimination across different infant movement classes and resilience to class imbalance.

Conclusions:

• The proposed framework offers an accurate, scalable, and automated solution for infant movement analysis.
• This technology has significant potential for early clinical screening and decision support, complementing expert assessments.