Deep Convolutional Neural Networks to Automatically Classify Human Heart Diseases

Automatic and reliable classification of human heart sounds is essential for self-monitoring heart conditions to improve life quality and public health. However, technology for self-monitoring of heart conditions is expensive, unavailable, and thus inaccessible to many people, especially in low-income regions. This study therefore presents a highly accurate automatic heart rate classification system using deep convolutional neural networks (DCNNs) conveniently implemented as a smartphone application that provide users with timely heart disease warnings. Two datasets were employed, and heart sound features were extracted using the Log-Mel spectrogram and the Mel-frequency cepstral coefficients (MFCCs). Five deep learning (DL) models were developed and combined through ensemble strategies that fused the predicted probabilities. When comparing the DL models using Log-Mel and the training technique two, the study found that Model 1c outperformed the others, attaining a 98.38% F1 score on the binary dataset. Meanwhile, Model 1b excelled with a 98.46% F1 score on the multi-class dataset. The improvements were statistically significant. The data augmentation method exhibited varying effects on model performance. Moreover, distinction in model predictions was observed when differentiating between normal and abnormal heart conditions using the Log-Mel feature and a Shapley Additive Explainability (SHAP) approach. The ensemble techniques were particularly successful, with the mean strategy in Ensemble C achieving a 98.62% F1 score for the binary dataset, surpassing existing methods, and reaching a 98.73% F1 score for the multi-class dataset.