Hybrid Framework for Cartilage Damage Detection from Vibroacoustic Signals Using Ensemble Empirical Mode Decomposition and CNNs

This study proposes a hybrid analytical framework for detecting chondromalacia using vibroacoustic (VAG) signals from patients with knee osteoarthritis (OA) and healthy controls (HCs). The methodology combines nonlinear signal decomposition, feature extraction, and deep learning classification. Raw VAG signals, recorded with a custom multi-sensor system during open (OKC) and closed (CKC) kinetic chain knee flexion-extension, underwent preprocessing (denoising, segmentation, normalization). Ensemble Empirical Mode Decomposition (EEMD) was used to isolate Intrinsic Mode Functions (IMFs), and Detrended Fluctuation Analysis (DFA) computed local (α₁) and global (α₂) scaling exponents as well as breakpoint location. Frequency-energy features of IMFs were statistically assessed and selected via Neighborhood Component Analysis (NCA) for support vector machine (SVM) classification. Additionally, reconstructed α₁/α₂-based signals and raw signals were converted into continuous wavelet transform (CWT) scalograms, classified with convolutional neural networks (CNNs) at two resolutions. The SVM approach achieved the best performance in CKC conditions (accuracy 0.87, AUC 0.91). CNN classification on CWT scalograms also demonstrated robust OA/HC discrimination with acceptable computational times at higher resolutions. Results suggest that combining multiscale decomposition, nonlinear fluctuation analysis, and deep learning enables accurate, non-invasive detection of cartilage degeneration, with potential for early knee pathology diagnosis.