Multimodal Wearable Sensor-Based Two-Stage KOA Grading With Explainable Machine Learning

Abstract

Knee osteoarthritis (KOA) screening and objective severity assessment are critical for personalized clinical interventions, yet single-modal data often fails to capture the complex and heterogeneous pathological gait changes associated with the disease. This paper proposes a wearable multimodal sensor fusion framework based on machine learning classifiers and explainable artificial intelligence (XAI) for a two-stage hierarchical KOA assessment. Surface electromyography (sEMG), inertial measurement unit (IMU), and plantar pressure signals were simultaneously acquired during gait, and the clinical task was decomposed into two independent stages: Stage1 for screening healthy subjects and KOA patients, and Stage 2 for grading mild and severe KOA among clinically confirmed KOA participants. To evaluate modal complementarity and eliminate algorithm-specific bias, we conducted systematic ablation experiments across all sensor configurations using the descriptive average performance of support vector machine (SVM), k-nearest neighbors (KNN), and random forest (RF) classifiers,which were independently trained and evaluated rather than combined as an ensemble model. Furthermore, repeated out-of-fold SHapley Additive exPlanations (OOF-SHAP) analysis was employed to provide stable interpretations of global feature importance and participant-level feature contributions. Ablation results demonstrate that the multimodal fusion improves performance over single- and dual-modal configurations, achieving $83.8~\pm ~2.6$ % accuracy for screening and $72.5~\pm ~2.9$ % accuracy for severity grading. OOF-SHAP analysis showed that gait/spatial features were important for KOA screening, while sEMG and gait-phase features contributed to severity grading.The proposed framework provides an interpretable and non-invasive approach for KOA assessment with potential clinical applications.