HVDNet: An Interpretable Deep Learning Framework for Heart Valve Disease Classification Using Tri-Axial Seismocardiogram Signals

Abstract

Effective screening for heart valve disease (HVD) is crucial for impeding its progression. However, current approaches lack transparency in classifying diverse HVDs. The seismocardiogram (SCG) signal provides comprehensive insights into cardiac activities across three axes, offering valuable information for detecting valvular abnormalities. To leverage this potential and address the aforementioned challenges, we propose HVDNet, an interpretable deep-learning framework for HVD classification using tri-axial SCG signals. The architecture integrates three modules: stacked 1-D convolutional neural networks with skip connections (sCNNs) to learn hierarchical features associated with morphological variations in SCG at different scales, long short-term memory (LSTM) layers to capture temporal variations within the feature maps, and self-attention (SA) layer to emphasize clinically relevant attributes. Evaluation on publicly available SCG databases demonstrate high accuracies: 99.35% on the validation set and 98.98% on the test set for HVD without co-existing diseases, and 99.21% on the validation set and 98.89% on the test set for aortic stenosis (AS) co-existing with other HVDs. Through an ablation study of different model variants, we found that integrating information from each axis component of the SCG signal yields optimal performance. Moreover, closely examining the learned attention weights reveals how the model emphasizes clinically relevant SCG attributes that characterize HVD. With its inherent transparency and superior performance compared to existing methods, the proposed model can become a reliable diagnostic tool for HVD, potentially improving patient care and treatment efficacy.