Optimized Feature Selection Framework for Efficient Heart Sound Classification

Global data from the WHO indicates that cardiovascular illnesses are the world's biggest cause of mortality. The complexity of the heart's sound data and its dynamic features across time and frequency domains make the classification tasks difficult. Phonocardiograms (PCG) are very useful for assessing heart health and taking the required actions for early heart disease detection and prevention. In this paper, signal processing algorithms extract spatial and frequency features from PCG recordings and fuse them to create a generic feature vector. Further, optimized feature selection through genetic algorithm is done to reduce the dimensions and achieve the most relevant feature set. Finally, several machine learning algorithms are being examined to categorize heart sounds for both binary and multiclass classifications. The proposed framework attains performance accuracy of 91.22% for multiclass (3-class) and 93.92% for binary classification using CatBoost with GA feature selection. The performance is compared with many recent related methods to corroborate the proficiency of the proposed model over them.