Interpretable machine learning models for classifying lateral epicondylitis based on wrist range of motion and...

BackgroundLateral epicondylitis (LE) is a prevalent overuse injury among workers performing repetitive wrist movements. Altered wrist range of motion (ROM) and reduced muscle strength occur in individuals with LE; however, their combined utility for classifying LE remains unclear. Interpretable machine learning can help identify and explain biomechanical features associated with LE.ObjectiveThis study developed and interpreted machine learning models to classify LE among assembly line workers using wrist ROM and muscle strength data, identifying the most influential biomechanical predictors using SHapley Additive Explanations (SHAP).MethodsData from 45 male assembly line workers (23 with LE, 22 without LE) were used. Predictor variables included wrist ROM (flexion, extension, radial deviation, ulnar deviation) and muscle strength (flexors, extensors, grip strength). Predictors were standardized within each training fold, and Gaussian noise augmentation was applied to training data. Four machine learning models were tuned using grid search and evaluated via repeated stratified 5-fold cross-validation. Discrimination, classification performance, and calibration were assessed, and SHAP was used for model interpretation.ResultsAll four models achieved comparable discrimination (accuracy: 0.83-0.87; ROC-AUC: 0.91-0.92). Support vector machine showed the highest accuracy, whereas logistic regression showed the highest ROC-AUC. SHAP analysis consistently identified wrist radial deviation ROM as the most influential predictor across all models, with lower values linked to higher LE probability.ConclusionInterpretable machine learning models classified LE with good cross-validated discrimination using simple biomechanical measurements. Reduced radial deviation ROM was consistently the strongest predictor, suggesting its relevance in LE assessment in assembly line workers.