Classification of knee-joint vibroarthrographic signals using time-domain and time-frequency domain features and least-squares support vector machine

Analysis of knee-joint vibration sounds, also known as vibroarthrographic (VAG) signals, could lead to a noninvasive clinical tool for early detection of knee-joint pathology. In this paper, we employed the wavelet matching pursuit (MP) decomposition and signal variability for time-frequency domain and time-domain analysis of VAG signals. The number of wavelet MP atoms and the number of significant turns detected with the fixed threshold from signal variability analysis were extracted as prominent features for the classification over the data set of 89 VAG signals. Compared with the Fisher linear discriminant analysis, the nonlinear least-squares support vector machine (LS-SVM) is able to achieve higher overall accuracy of 73.03%, and the area of 0.7307 under the receiver operating characteristic curve.