Cardiovascular disease detection based on deep learning and multi-modal data fusion

Electrocardiogram (ECG) and phonocardiogram (PCG) are widely used for early prevention and diagnosis of cardiovascular diseases (CVDs) because they accurately reflect the state of the heart from different perspectives and can be conveniently collected in a non-invasive manner. However, there are few studies using both ECG and PCG for CVD detection, and extracting discriminative features without losing useful information is challenging. In this study, we propose a dual-scale deep residual network (DDR-Net) to automatically extract the features from raw PCG and ECG signals respectively. A dual-scale feature aggregation module is used to integrate low-level features at different scales. We employ SVM-RFECV to select important features and use SVM for the final classification. The proposed method was evaluated on the "training-a"set of 2016 PhysioNet/CinC Challenge database. The experimental results show that the performance of our method is better than that of methods using only ECG or PCG as well as existing multi-modal studies, yielding an accuracy of 91.6% and an AUC value of 0.962. Feature importance of ECG and PCG for CVD detection is analyzed.