Hybrid Model Approach for Heart Murmur Detection in Phonocardiographic Signals

Abstract

Abstract The conventional method for diagnosing heart murmurs is cardiac auscultation performed by a trained professional. Which is prone to human error, subjectivity, and inter-observer variability. However, the utilization of computational methods for detection and interpretation presents various associated challenges. The goal of this study is to improve the performance of computational methods for detecting the presence or absence of heart murmurs in phonocardiograms. This study explores two techniques from the acoustic anomaly detection domain for heart murmur detection. The first approach is to use a reconstruction technique based technique to model the distribution of normal data and detect anomalies based on the reconstruction error. Upon experimentation, it is identified that both normal and abnormal recordings share a significant amount of similar patterns. Thus, using a threshold based on reconstruction error would not yield significant results for murmur detection. The second approach is to use a classification-based method using a Temporal Convolutional Network (TCN) based model architecture. This approach was able to achieve a weighted accuracy score of 79.07% and F-Score of 66.22%, exceeding the performance of the latest methods on the PhysioNet 2022 dataset. ii