Empirical Wavelet Transforms-based Feature Extraction Of Vibroarthrographic Signal For Knee Osteoarthritis Detection And Stage Determination.

Abstract

Knee Osteoarthritis (KOA) is a disease that causes the knee joint cartilages to thin and the surfaces of the joint to become rougher, resulting in bone-on-bone contact, inflammation, and the formation of bone spurs, leading to pain, stiffness, so that the knee not able to move as the expected. KOA can affect any age range however; it is more common in the growing age of the population and women above 50 years. Osteoarthritis is the most common cause of years lived with disability (YLDs) globally. These conditions have a profound impact on people’s ability to live and function. KOA can be diagnosed using various imaging and signal-based techniques such as X-ray, Computed Tomography (CT)-scan, magnetic resonance imaging (MRI), and Vibroarthrographic signal. This research aimed to detect and determine the stage of knee osteoarthritis using features extracted by empirical wavelet transform from vibroarthrographic signal. The developed pipeline involved designing an electronic stethoscope using SG Electret microphone and single-headed stethoscope to optimize joint sound detection in knee osteoarthritis. Acoustic signals were recorded from both healthy (with no symptoms) and knee osteoarthritic subjects with age limit above 45 years with symptoms of pain, stiffness, crepitus, and functional limitation. The recorded signals were preprocessed using a fourth-order low-pass filter to remove noise and artifacts. Signal decomposition was performed using empirical wavelet transform (EWT) to extract audio features. Various features were extracted from the EWT coefficients, including kurtosis, skewness, and entropy. Long short term memory deep learning and different machine learning models, including support vector machine, KNN, and ensemble classifiers, were trained and tested using the extracted features. An ensemble machine learning model with an accuracy of 90.00% was selected for knee osteoarthritis diagnosis. The findings of this thesis demonstrate that the proposed ensemble machine learning method achieved a high-level performance in the classification task, as evidenced by the precision of 96.34%, recall of 97.53%, and f1_score of 96.93%. The sensitivity and specificity metrics 97.53% and 92.86%, respectively, also indicated the effectiveness of the method in correctly identifying positive and negative instances. The proposed method delivered a considerable result for the classification of knee osteoarthritis as stage 0(healthy) up to stage 4(severe). Vibroarthrographic signal-based diagnosis is preferable because of its portability, ease of use, reliability, and inexpensiveness. Moreover, as a further III consideration for future endeavors, the diagnosis approach would be more powerful if wireless sensors were used because signals could be collected at different loading styles using wireless sensors.