JMIR Publications

ABSTRACT

Background:

Accurate and interpretable classification of knee function is essential for evaluating movement quality in clinical and sports contexts but remains challenging without expert assessment or laboratory-based motion capture.

Objective:

This study developed and validated an interpretable ensemble learning framework that classifies knee function into three levels (Good, Moderate, Poor) from single-smartphone, front-view single-leg squat (SLS) videos.

Methods:

A dataset of SLS videos was labeled by physiotherapists into three categories, and kinematic features were extracted from 2D pose estimation. Engineered features, including ratios, nonlinear terms, and interaction variables, were selected using a hybrid statistical and recursive elimination approach. Six classifiers were trained with stratified 5-fold cross-validation and a held-out test set, and a soft voting ensemble was developed to enhance robustness. SHAP and LIME were applied for global and local interpretability.

Results:

The soft voting ensemble combining Logistic Regression and LightGBM achieved high performance on the held-out test set (accuracy = 87.5%, F1-score = 87.6%, AUC = 0.938). SHAP analysis identified the Trunk × Knee interaction as the most influential feature for distinguishing poor classifications, while LIME provided instance-level explanations that clarified decision boundaries.

Conclusions:

Global and local interpretability enhanced model transparency by illustrating why impaired movement patterns were detected, thereby supporting personalized feedback on knee function. This interpretable framework provides a practical and accessible solution for functional knee assessment, with direct applicability to rehabilitation, athletic screening, and remote monitoring. Clinical Trial: The study protocol was reviewed and approved by the Institutional Review Board of Jeonju University