Transfer learning based Chronical Heart Failure Detection from Heart Sounds

Abstract

Chronic Heart Failure (CHF) is a major health condition, and early detection is crucial for effective management. In this study, we propose a method for CHF detection using heart sounds, utilizing four different predictive algorithms: (1) a machine learning (ML) approach with Random Forest, (2) a deep learning (DL) model based on a custom Convolutional Neural Network (CNN), (3) a hybrid ML-DL approach combining both techniques, and (4) a transfer learning-based approach using a pretrained VGGish model for feature extraction. The PhysioNet Heart Sound Dataset is used, and the dataset undergoes pre-processing, including noise reduction and feature extraction. The models are then trained and evaluated on the dataset, with the final system predicting CHF from unseen heart sound recordings. Our approach demonstrates a robust framework for leveraging both classical ML and advanced DL techniques for the early detection of CHF, potentially improving patient outcomes through timely intervention.