Dynamic Mode Decomposition-Based Features for Cardiovascular Disease Analysis From Phonocardiogram Signals

Phonocardiogram signal (PCG) plays a vital role in the diagnosis of cardiovascular diseases through its ability to showcase heart sounds in the form of signals. Even though several studies have been put forward for the analysis of PCG, there still remains scope to concisely explain the substantial characteristics of PCG. DMD is a popular and promising data-driven method that is used to find coherent structures from the underlying data. In this paper, a novel approach is introduced using dynamic mode decomposition (DMD) based features to analyze the underlying patterns of PCG. Features related to time domain, frequency domain, statistical and information gain are extracted from the DMD modes for the detection of heart diseases using PCG. The study evaluates several machine learning algorithms and compares them with deep learning approaches using two open-source PCG datasets, representing both binary and multi-class data. The results showed an overall Precision of 93.5% and F1-score of 93.3% for heart sound data. The relevance of each feature is established through permutation importance feature ranking and local interpretable model-agnostic explanations (LIME), an explainable artificial intelligence (XAI) technique. The study delivers efficient classification accuracy across various cardiac conditions, recommending its potential generalizability in applied clinical scenarios.