Publication

Abstract :

Gait analysis is a critical tool for diagnosing neurological and musculoskeletal disorders in children, yet traditional approaches depend on costly motion capture systems and large datasets, restricting their use in resource-limited settings. This study introduces a data-efficient machine learning framework for classifying pediatric gait anomalies, utilizing a small dataset of 64 labeled videos *sourced from from a private hospital*. Pose estimation was performed using YOLOv8 on smartphone-captured videos, from which key gait features were extracted. A BaggingClassifier, optimized via stratified k-fold cross-validation, was subsequently applied. The model achieved an accuracy of 81.82% and a recall of 93.33% on the test set, demonstrating robust performance despite limited data. Stratified cross-validation enhanced model stability, reducing accuracy variability across folds. This scalable, cost-effective approach harnesses ubiquitous smartphone technology, offering a practical solution for early gait anomaly detection in underserved regions and advancing accessible healthcare diagnostics.