Smartphone-Based Interpretable Machine Learning for Classifying Single-Leg Squat Performance Using Trunk, Pelvic, and Knee Kinematics: Cross-Sectional Study

TL;DR

This paper introduces a smartphone-based machine learning system to classify single-leg squat performance into three levels using interpretable models and kinematic features.

Contribution

The novelty lies in using interpretable machine learning with coordination-informed features from smartphone videos to classify movement quality.

Findings

Adaptive boosting achieved 84% accuracy in classifying single-leg squat performance.

Coordination-related features like knee-to-trunk ratio and knee-trunk interaction were key predictors.

SHAP and LIME provided global and local explanations for model decisions.

Abstract

Single-leg squat (SLS) performance is widely used to screen functional movement quality, but practical assessment often relies on expert visual grading or laboratory-based motion capture. In addition, conventional SLS criteria mainly focus on isolated joint deviations and may overlook coordination-related, multisegment movement patterns that characterize impaired performance.

This study aimed to examine the feasibility of an interpretable machine learning framework for classifying SLS performance into 3 levels (good, moderate, and poor) from single-smartphone, frontal-view videos based on trunk, pelvic, and knee kinematics, and to evaluate coordination-informed features and model explainability using Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME).

A dataset of frontal-view SLS videos was labeled by physiotherapists into 3 functional categories (good, moderate, and poor). Videos were processed using 2D pose estimation, and models were trained on 17 engineered kinematic features derived from trunk, pelvic, and knee angles. Following the feature selection, 7 classifiers were trained and evaluated using the 8 selected features with stratified 5-fold cross-validation and a held-out test set. SHAP and LIME were applied for global and local interpretability.

On the held-out test set, adaptive boosting classified SLS performance with an accuracy of 0.84, an F1-score of 0.85, and an area under curve of 0.92. SHAP indicated that the summated angle (trunk + pelvis + knee), coordination-related features (knee × trunk interaction and knee-to-trunk ratio), and knee angle were key contributors to model predictions. LIME provided instance-level explanations that helped interpret individual classifications and decision boundaries.

This study presents an interpretable machine learning framework for classifying SLS performance into 3 levels using frontal-view videos acquired with a single smartphone. By leveraging coordination-informed engineered features and explainable artificial intelligence, the framework enables transparent interpretation of movement performance beyond isolated joint deviations. The proposed workflow uses smartphones for standardized video acquisition, while performance screening is achieved through machine learning. Given its lightweight feature design, this framework has potential for future implementation on modern smartphones and may support rehabilitation planning and injury-prevention strategies in sports and clinical settings.

Figures