Analysis and Classification of Vibroarthrographic Signals using Tuneable ‘Q’ Wavelet Transform

A Computer-aided Diagnosis (CAD) system using vibroarthrography(VAG) signals is an efficient, non-invasive alternative for detection of human knee joint disorders. VAG signals are characterized as nonstationary and aperiodic in nature and thereby extracting features is a challenging task. Secondly, biomedical datasets are mostly skewed due to small datasets especially with respect to abnormal class and thereby there is a threat of overfitting in classification. Therefore, in this study, we propose a new methodology using a nonstationary linear signal processing techniques called Tuneable ‘Q’ Wavelet Transform (TQWT). The VAG signals are decomposed into subband signals using TQWT and entropy information is extracted from each subband. To remove the unbalance between the two classes, an oversampling technique called as Synthetic Minority Oversampling Technique (SMOTE) is employed before training any machine learning model. Recursive feature elimination is utilized to discard the subbands with least contributions, and a Random Forest Classifier is employed for classification. This approach results in an overall 80.8% leave one out classification accuracy with 86.8% accuracy for knee-joint defect detection.