Assessment for coronary heart disease severity by STFTs from synchronized PCG and ECG

Electrocardiogram (ECG) and phonocardiogram (PCG) signals reflect the electrical and mechanical activities of the heart, respectively. Noninvasive assessment of the severity of coronary heart disease (CHD) using ECG and PCG signals not only reduces the burden on physicians and frees up medical resources, but also serves as an affordable means of assessment and reduces the financial burden on patients. The purpose of this study was to provide a dual-input lightweight model based on the ShuffleNet architecture to assess CHD of different severity. The distinction between patients with Unconfirmed CHD with chest pain, patients with mild CHD, and patients with severe CHD was achieved without relying on the manual extraction of waveform features of ECG and PCG. ECG and PCG signals were prepossessed and transformed into short-time Fourier Transform spectrograms (STFTs). The proposed model in this study was trained in a five-fold cross-validation and the performance was compared with six common deep learning (DL) models. The proposed model in this study demonstrated an accuracy of 94.27%, precision of 94.37%, sensitivity of 94.22%, and F1 score of 94.25%. Validation of its validity and reliability in the task of classifying CHD at different levels of severity. The model proposed in this study exhibited lower sensitivity and variability compared to other classical models, indicating its strong robustness.