Anomaly detection in vibroarthrographic signals using handcrafted signal features and one-class methods

Knee osteoarthritis (OA) alters joint mechanics and may modify vibroarthrographic (VAG) signals generated during movement. This study evaluates whether OA-related changes can be detected as anomalies relative to a reference distribution learned exclusively from healthy participants, without using OA data during model training or threshold selection. VAG signals were obtained from 97 participants, including 48 healthy controls (HC) and 49 patients with clinically and radiologically confirmed knee OA. Signals recorded with a contact microphone positioned on the patella during closed-kinetic-chain (CKC) movement were segmented into cycles and represented by eight handcrafted time-, amplitude-, and frequency-domain features. Four one-class anomaly detection methods were compared: Mahalanobis distance, One-Class Support Vector Machine, robust Euclidean distance, and Isolation Forest. Performance was assessed over 50 participant-level random splits at the cycle, joint, and patient levels. Mahalanobis distance achieved the highest ROC AUC at all aggregation levels, reaching 0.703 ± 0.049 for cycles, 0.731 ± 0.059 for joints, and 0.729 ± 0.075 for patients. Threshold-based evaluation using percentiles derived solely from healthy validation data yielded high specificity but limited sensitivity; the best balanced accuracy at the patient level was 0.627 ± 0.062 with the P90 threshold, with sensitivity of 0.386 ± 0.202 and specificity of 0.868 ± 0.164. These findings demonstrate the feasibility of modelling knee OA as a deviation from a multidimensional healthy VAG pattern and support one-class analysis as a complementary framework for non-invasive functional assessment, while indicating that further development is required to improve sensitivity.