Vibroarthrography combined with machine learning showed potential as a noninvasive method for identifying osteoarthritis-related cartilage damage, but substantial demographic confounding complicates interpretation of its diagnostic performance.
Researchers evaluated 97 participants, including 48 healthy controls and 49 patients with symptomatic knee osteoarthritis or cartilage damage who underwent surgery. The final analysis included 95 healthy knees and 49 affected knees. Osteoarthritis status was established through clinical and radiologic evaluation, with cartilage damage confirmed intraoperatively during arthroscopy or total knee arthroplasty. Controls had no history of knee pain, injury, or known joint disease and had negative findings on orthopedic examination.
Vibroarthrography (VAG) signals were recorded using an accelerometer over the central patella while participants completed 10 repetitions under closed kinetic chain and open kinetic chain conditions. Statistical features from 3 acceleration axes were evaluated using 8 machine-learning classifiers with 5-fold stratified grouped cross-validation, which kept data from individual participants separate between training and testing sets.
Ensemble classifiers generally showed the strongest discrimination. Under closed kinetic chain (CKC) conditions, Random Forest achieved an AUC of 0.917, XGBoost 0.912, and AdaBoost 0.910. XGBoost had 79% accuracy, 69% sensitivity, and 88% specificity. Under open kinetic chain (OKC) conditions, Random Forest achieved an AUC of 0.857 and XGBoost 0.846. Both had 78% accuracy, 68% sensitivity, and 86% specificity. The researchers noted that variability across cross-validation folds was expected given the limited cohort and strict patient-level validation.
Simple signal characteristics, including root-mean-square, crest factor, median, and mean, were among the most informative features. The y-axis features contributed most duriparticipants separateng CKC movements, whereas z-axis features contributed most during OKC movements.
However, these axes represented the sensor coordinate system rather than anatomical directions. Differences among axes could therefore reflect sensor orientation, movement and loading patterns, soft-tissue transmission, or fixation and motion artifacts rather than direct evidence of directional changes in vibration propagation.
Demographic differences represented an important source of potential confounding. Controls had a mean age of 27 years and body mass index of 22.94, compared with 55 years and 30.32, respectively, among patients with osteoarthritis. In a separate analysis, classifiers using only age and body mass index (BMI) achieved accuracies ranging from 84% to 90%, comparable to or higher than the VAG-based models. The researchers therefore noted that age and BMI may account for a substantial proportion of group separation and called for larger, more balanced studies directly comparing demographic-only, VAG-only, and combined models.
Other limitations included the relatively small cohort and use of a single sensor configuration and acquisition setup. Because the accelerometer was attached to the skin over the patella rather than directly to bone, signals represented a soft-tissue-coupled surrogate of joint vibration and could be affected by skin movement and signal attenuation. The researchers also called for validation across different sensors and mounting configurations, standardized measurement protocols, and testing across clinical environments.
The findings support further evaluation of VAG as a noninvasive supportive screening method for knee joint degeneration, but do not establish its incremental diagnostic contribution beyond demographic characteristics. “Further validation across different sensor types, mounting configurations, and measurement setups is required,” wrote first author Robert Karpiński, PhD, of Lublin University of Technology and The John Paul II Catholic University of Lublin, and colleagues.
The authors reported no competing interests.
Source: Scientific Reports
