The objective of this study is to develop an automated clinically relevant diagnostic framework for a group of cardiovascular diseases, namely, valvular heart diseases (VHDs). Though for clinical screening, manual auscultation using a stethoscope is the current standard, it is losing its significance due to the demand for years of skill and experience from the clinician. While the physiological location for capturing heart sound signal, called phonocardiogram (PCG), is standardized, the measurement and classification of PCG into different VHDs need automation to redeem its importance. Recent research on the applicability of machine learning (ML) and deep learning (DL) in PCG-based classification of VHDs produced high accuracy, but often loses the clinical explanation behind them, and hence, often not endorsed by the healthcare fraternity. The proposed study attempts to bridge this gap with two types of clinically explainable automated frameworks: 1) one utilizes clinically relevant PCG features in statistical rule-based models (RBM) as well as ML-based decision-tree (DT) frameworks; 2) the other uses time-frequency representation (TFR) with 2-D convolution neural network (CNN) and signal of PCG cycle with 1-D CNN, reveals similar clinically explainable intelligence. As a case study, mitral valve prolapse (MVP) − a specific VHD categorized into three severity subclasses of early systolic click (ESC), ESC with murmur (ESCM), and mid-SC with murmur (MSCM). The signature events (S1, S2, SC, and murmur) in PCG are segmented using a novel Teager–Kaiser energy operator and continuous wavelet transform (TKEO-CWT) algorithm. Clinically relevant intra- and inter-event features like duration, intensity, and pitch (dominant frequency) in RBM and DT produced 100% and 98.6% accuracies, respectively. Interestingly, gradient class activation (Grad-CAM) of 1-D and 2-D CNN with PCG signal and scalogram as input also converges to the same clinical explanations. Thus, the proposed clinically explainable frameworks demonstrate excellent potential in measurements and classifications in future clinical diagnostic products.
Read full abstract