Chronic Heart Failure Diagnosis Through Heart Sounds Processing Using Advanced Deep Learning Models

Abstract

Chronic heart failure (CHF) remains a major global health challenge, with rising incidence and significant mortality. Despite advances in healthcare technologies, accurate early detection of CHF is still a complex task due to the high-dimensional nature of clinical data. In this study, we propose an innovative hybrid approach for CHF detection, combining classic Machine Learning (ML) and advanced Deep Learning (DL) techniques. Our method integrates expert feature-based ML with DL models trained on spectro-temporal representations of heart sound signals (PCG). Leveraging data from both publicly available datasets and a newly curated CHF-specific dataset, we demonstrate that our approach significantly improves detection accuracy and efficiency. The proposed framework includes lightweight convolutional neural networks (CNN), hybrid CNN-autoencoder models, and parallel architectures for feature fusion. The system achieves an accuracy of 92.9%, with a robust ability to distinguish between healthy individuals and CHF patients. Additionally, our model effectively classifies different CHF phases, such as decompensated and recompensated states, with a 93.2% accuracy. This novel approach offers promising potential for early diagnosis and real-time CHF monitoring, paving the way for personalized and home-based healthcare solutions aimed at reducing hospitalizations and improving patient outcomes.