Abstract / Summary
Knee osteoarthritis (KOA) is a prevalent degenerative joint disorder for which non-invasive and cost-effective screening approaches are increasingly needed. Integrating multimodal biomechanical signals (vibroacoustic and neuromuscular) during loaded knee motion offers rich pathological insights, yet its effective utilization is challenged by high-dimensional redundancy and cross-sensor heterogeneity. This study proposes a structured and interpretable multimodal framework for KOA screening, anchored in mechanistically-driven hybrid feature engineering. Comprehensive descriptors spanning time, spectral, and nonlinear domains are extracted from multi-compartment joint signals and quadriceps electromyography. To capture inter-compartment coupling and baseline physiological context, cross-sensor interaction features are systematically constructed and integrated with demographic attributes. To handle the resulting high-dimensional space, a two-stage hybrid feature selection pipeline integrating non-parametric statistical filtering with BorutaShap refinement is embedded to isolate a compact, discriminative feature subset. Evaluated via strict subject-level cross-validation, the ensemble classifier demonstrates that structured multimodal fusion consistently outperforms single-sensor configurations, yielding robust screening performance. Crucially, SHAP-based feature attribution and data-driven dependency exploration provide transparent, multi-level insights into the model’s decision-making process. Rather than claiming definitive causation, this post-hoc analysis translates complex statistical associations into clinically testable biomechanical hypotheses. Overall, the proposed framework demonstrates the potential for reliable, transparent KOA screening and offers a biomechanically grounded perspective for understanding joint abnormalities.