Automated detection of heart valve disorders with time-frequency and deep features on PCG signals. (September 2022)

• Record Type: Journal Article
• Authors: Arslan, Özkan
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Heart valve diseases (HVDs) -- Phonocardiography (PCG) -- Time-frequency analysis -- Deep features -- Multilayer extreme learning machine; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103929
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Detection of four type heart valve disorders based on PCG signals has been realized. The performances of DWT, WPT, PWPT, and EMD methods were analyzed. The performances of deep features obtained by VGG16, ResNet50, MobileNetV2, and ML-ELM were investigated. RFE algorithm was used as features selection method. The highest performance (accuracy of 99.4%, MCC and G-mean of 99.3%) was obtained using PWPT + EMD features and RF classifier. Abstract: Heart valve diseases (HVDs) can cause cardiac arrhythmias, heart attacks, and sudden cardiac death if not diagnosed early. Therefore, the detection of HVDs is critical in order to avoid heart-related mortality. The focus of this research is to establish an efficient computer-aided diagnosis approach that detects HVDs using phonocardiogram (PCG) signals. The proposed approach uses traditional time–frequency and deep features with machine learning models. The time–frequency features are extracted from non-linear measurements using discrete wavelet transform (DWT), wavelet packet transform (WPT), perceptual wavelet packet transform (PWPT) and empirical mode decomposition (EMD) methods. Deep features are extracted from VGG16, ResNet50 and MobileNetV2 pre-trained CNN models, and multilayer extreme learning machine (ML-ELM) model using scalogram images of PCG signals. Recursive feature elimination (RFE) algorithm is applied to all features and the most distinctive features are selected. Experimental results show that the PWPT + EMD features selected by RFE and the random forest (RF) classification model achieve the highest performance with accuracy of 99.4%, Matthews correlation coefficient (MCC) and G-mean of 99.3%. In another proposed approach, ML-ELM deep features selected by RFE algorithm and RF classification model provide accuracy and G-mean of 98.9%, and MCC values of 98.6%. It was observed that the time–frequency features have outperformed compared to deep features for the detection of HVDs. The proposed approach is compared with the existing studies and it has obtained higher performance values ​​than the approaches using the same database. The proposed approach can be considered as an easily integrated system on the embedded platform.