An Explainable Deep Learning Framework for Heart Disease Classification from Multi-Modal Clinical Data

Abstract

Heart disease is one of the leading causes of death; therefore, it becomes extremely important to devise accurate diagnostic procedures to classify it correctly. The work proposed in this study presents an explainable deep learning approach for the classification of heart disease using multimodal biosignals like ECG (Electrocardiography) and PCG (Phonocardiography). At first, the input ECG and PCG signals are extracted from the public domain database and undergo a denoising process using the Structure-Oriented Space Varying Median Filter (SOSVMF) to retain the significant features of the biosignals. The denoised ECG and PCG are then fed to the Variational Autoencoder (VAE) for extracting the relevant latent features. The latent features of both modalities are merged using the proposed Multiscale Feature Fusion Module (MFFM), and the fused features are classified using the ResNet to obtain normal and abnormal classes of heartbeats. Further, Bayesian LIME and SHAP are used to provide physicians with insight regarding the signal regions contributing to the final classification decision. The simulation results reveal that the proposed method achieves 99.21% classification accuracy and 98.87% F1-Score on the benchmark database. These results are comparatively better than the state-of-the-art methods and provide an interpretable and reliable framework for classifying normal and abnormal classes of heartbeats in a multimodal setting.