Analysis of vibration signals in the diagnosis of knee cartilage damage using vibroarthrography and machine learning

Cartilage damage in the knee joint is a common issue resulting from trauma, overuse, inflammation, or anatomical abnormalities. If left untreated, it often leads to osteoarthritis (OA), one of the leading causes of disability worldwide. Current diagnostic methods, such as X-ray and magnetic resonance imaging (MRI) remain limited in providing rapid, accessible, and cost-effective assessment of functional joint abnormalities. This study developed a non-invasive expert system for diagnosing cartilage damage based on vibroarthrography (VAG). VAG records vibrations or sounds generated during joint movement, offering a potentially cost-effective, accurate, and scalable diagnostic method. The study utilized vibration signals recorded during diagnostic tests on a specially designed measurement track. Based on these signals, statistical discriminants reflecting the condition of the knee joint were identified. Machine learning models were applied to classify cartilage damage and assess the informational value of individual sensors placed at various anatomical locations. Model performance was evaluated using strict patient-level grouped cross-validation. Additional analyses included feature-importance assessment, calibration analysis, and evaluation of the contribution of individual vibration directions. Analysis of collected data demonstrated that appropriately processed VAG signals, combined with machine learning algorithms, enable stable discrimination between healthy controls and OA patients at the patient level. Ensemble-based classifiers, particularly XGBoost and Random Forest, achieved the highest discriminative performance across both movement protocols. Feature-importance analysis revealed that descriptors associated with signal amplitude and variability contributed most strongly to classification performance. In contrast, directional analysis indicated that vibrations recorded along the Y axis carried the highest diagnostic information. Calibration analysis further confirmed satisfactory agreement between predicted probabilities and observed outcomes. The results suggest that VAG could be a promising, low-cost, and non-invasive supportive screening methodology for the functional assessment of knee joint degeneration. Although demographic factors contributed substantially to classification performance, the observed vibroacoustic patterns indicate the presence of biomechanically meaningful information associated with OA. Further validation in larger, more diverse cohorts is required before clinical implementation.