Classification Of Knee Joint Vibration Signal With Machine Learning Algorithms

The knee is the most complex joint and helps the body perform different locomotion functions. Detection of knee joint pathology at an early stage can help clinicians apply appropriate therapeutical or surgical procedures to retard the degenerative process in the affected knee joint. The knee joint vibration arthrographic (VAG) signal recorded by accelerometer or electrostethoscope sensors attached on the surface of the knee cap can be used in clinical practice for noninvasive screening of the impaired knee joints.In this thesis, features were extracted from VAG signals, including the number of atoms derived from the wavelet matching pursuit decomposition (Natom), turns count (TC), form factor (FF), variance of the mean-squared values (VMS), entropy (H), skewness (SK), kurtosis (KU) and fractal dimension (FD). Then, three machine learning algorithms were proposed for the VAG signal classification. The results demonstrate the effectiveness and superiority of these VAG signal analysis methods.The first approach used the bivariate feature distribution estimation and maximal posterior probability decision criterion (MPP) to distinguish the normal VAG signals from the abnormal ones. The MPP provided the total classification accuracy of86.67%and the area (Az) of0.9096under the receiver operating characteristics (ROC) curve, which were superior to the results obtained by either the Fisher’s linear discriminant analysis (accuracy:81.33%, Az.0.8564) or the support vector machine (accuracy:81.33%,Az:0.8533). The second algorithm, multiple classifier system based on recurrent neural network (MCS-RNN), achieved an area (Az) of0.8230under the ROC curve. The third method, dynamic weighted classifier fusion (DWF), reached the overall accuracy88.76%and the area (Az) of0.9515.