Time-frequency-domain deep representation learning for detection of heart valve diseases using PCG recordings for IoT-based smart healthcare applications

Abstract

Heart valve diseases (HVDs) are pathological conditions that occur due to a defect in any one of the valves of the heart, causing heart failure and sudden cardiac death if not diagnosed in the early stage. Cardiac imaging, using techniques such as echocardiography, is used in clinical settings to diagnose HVDs. Cardiac auscultation is a low-cost and readily available diagnostic framework for detecting HVDs. The automated analysis of the heart sounds or phonocardiography (PCG) data obtained after cardiac auscultation is performed using artificial intelligence (AI) techniques to detect HVDs for smart healthcare applications. The development of novel AI-based techniques for automated detection of HVDs is essential for Internet of Things (IoT)-based innovative healthcare applications and improving patient care. This chapter proposes a novel automated approach deployed on a cloud-based framework to detect HVDs using PCG signals. The synchrosqueezed short-time Fourier transform (SSTFT) is employed to evaluate the PCG signal's time-frequency-domain (TFD) representation. A deep representation learning (DRL)-based model implemented using a frozen MobileNetV2-based transfer learning block followed by global average pooling (GAP) and dense and drop-out layers is used to detect different HVD classes using the TFD representation of PCG signals. The suggested approach is evaluated using PCG signals from two publicly available databases. For database 1, the five class-based HVD detection tasks, such as normal (N) versus aortic stenosis (AS) versus mitral stenosis (MS) versus mitral regurgitation (MR) versus mitral valve prolapse (MVP), are formulated using the SSTFT-based TFD representation and DRL. Similarly, for database 2, the normal versus abnormal two-class classification task is performed. The results demonstrate that the proposed approach obtained accuracy values of 98% and 90.60% using PCG signals from database 1 and database 2, respectively. The suggested approach is compared with existing HVD detection methods using PCG signals from both databases. The TFD-based DRL model is deployed on the cloud-based framework for IoT-enabled HVD detection using PCG recordings. © 2024 by Elsevier Inc. All rights reserved, including those for text and data mining, AI training, and similar technologies.