Comparative Analysis of Feature Selection Based on Metaheuristic Methods for Human Heart Sounds Classification Using PCG Signal

Abstract

Cardiovascular disease is a critical threat to human health, as most death cases are due to heart disease. Although several doctors employ stethoscopes to auscultate heart sounds to detect abnormalities, the accuracy of the approach is considerably dependent upon the experience and skills of the physician. Consequently, optimal methods are required to analyse and classify heart sounds with Phonocardiogram (PCG) signal-based machine learning methods. The current study formulated a binary classification model by subjecting PCG signals to hyper-filtering with low-pass and cosine filters. Subsequently, numerous features are extracted with the Wavelet Scattering Transform (WST) method. During the feature selection stage, several metaheuristic methods, including Harris Hawks Optimisation (HHO), Dragonfly Algorithm (DA), Grey Wolf Optimiser (GWO), Salp Swarm Algorithm (SSA), and Whale Optimisation Algorithm (WOA), are employed to compare the attributes separately and determine the ideal characteristics for improved classification accuracy. Finally, the selected features were applied as input for the Bidirectional Long Short-Term Memory (Bi-LSTM) algorithm, simplifying the classification process for distinguishing normal and abnormal heart sounds. The present study assessed three PCG datasets: PhysioNet 2016, Yaseen Khan 2018, and PhysioNet 2022, documenting 94.85%, 100%, and 66.87% accuracy rates with 127-SSA, 168-HHO, and 163-HHO, respectively. Based on the results of the PhysioNet 2016 and 2022 datasets, the proposed method with hyperparameters demonstrated superior performance to those with default parameters in categorising normal and abnormal heart sounds appropriately.