Automated Detection of Heart Valve Diseases Using Stationary Wavelet Transform and Attention-Based Hierarchical LSTM Network

This article presents a phonocardiogram (PCG)-based heart valve disease (HVD) detection for primary healthcare units. In this work, we propose a stationary wavelet transform (SWT) decomposition, followed by an attention-based hierarchical long short-term memory (HLSTM) network for each subband to detect HVDs. Initially, the PCG signal is preprocessed and segmented into PCG segments. Then, each PCG segment is decomposed using SWT. We employ subband-specific HLSTM networks to utilize the temporal and relative temporal information present in each subband of the PCG segment for detecting HVDs. Then, the outputs of each subband-specific HLSTM are fed to their respective intra-subband attention layer for weighted aggregation to obtain the subband-specific representation. Furthermore, the inter-subband attention layer aggregates these subband-specific representation vectors to improve the detection of HVDs. The proposed method is tested and validated using two open-access databases. The Physionet Challenge 2016 database shows an overall sensitivity (OSe) of 97.96%, overall specificity (OSp) of 99.02%, and overall accuracy (OA) of 98.55%, using the proposed method for binary classification. Similarly, the heart sound (HS) murmur database shows impressive classwise performance measures and an OA of 99.47%, using the proposed method for multiclass HVD classification. The proposed method’s impressive performance and generalization can help to detect HVD anomalies during preliminary checkups in healthcare units.