Smartphone-Based Interpretable Machine Learning for Classifying Single-Leg Squat Performance Using Trunk, Pelvic, and...

Area of Science:

• Biomechanics
• Machine Learning
• Sports Science

Background:

• Single-leg squat (SLS) performance is crucial for assessing functional movement quality.
• Current assessment methods like visual grading or motion capture have limitations.
• Conventional criteria often miss coordination patterns indicative of impaired performance.

Purpose of the Study:

• To assess the feasibility of an interpretable machine learning framework for SLS performance classification.
• To classify SLS performance into good, moderate, and poor levels using smartphone videos.
• To evaluate coordination-informed features and model explainability using SHAP and LIME.

Main Methods:

• A dataset of frontal-view SLS videos was collected and labeled by physiotherapists.
• 2D pose estimation processed videos, extracting 17 kinematic features from trunk, pelvis, and knee.
• Classifiers were trained on 8 selected features, with SHAP and LIME used for interpretability.

Main Results:

• Adaptive boosting achieved 0.84 accuracy, 0.85 F1-score, and 0.92 AUC on the test set.
• Key predictive features included summated angle, knee-trunk interaction, knee-to-trunk ratio, and knee angle.
• SHAP and LIME provided insights into global and local model predictions.

Conclusions:

• An interpretable machine learning framework for SLS classification using smartphone videos was developed.
• The framework offers transparent movement interpretation beyond isolated joint deviations.
• This approach supports rehabilitation planning and injury prevention strategies.