An Explainable AI Framework for Vibroarthrographic Knee Joint Disorder Detection Using Entropy Based Feature Engineering

In this work, a computer-aided diagnosis (CAD) system is proposed to classify knee joint conditions into healthy and non-healthy categories using vibroarthrographic (VAG) signals. To effectively analyze the non-stationary nature of VAG signals, the Tunable Q-factor Wavelet Transform (TQWT) is employed to decompose the signals into multiple sub-band components. From these sub-bands, eight entropy-based features are extracted, yielding a total of 400 features. Feature selection techniques are then applied to identify the most relevant descriptors, yielding 45 optimal features. Six machine learning classifiers, along with an ensemble learning strategy, are utilized for classification. The ensemble-based framework achieves 90.1% classification accuracy and an AUC of 0.91 under leave-one-subject-out validation. Furthermore, an explainable artificial intelligence (XAI) approach based on Shapley Additive exPlanations (SHAP) is incorporated to interpret the model predictions by quantifying the contribution of individual entropy features. This explainability enhances the transparency and reliability of the proposed system, supporting its potential applicability in clinical decision-making.