Multiscale Phonocardiogram Signal Analysis with Attention-Based Deep Feature Learning for Automated Cardiovascular Disease Prediction

Abstract

Cardiovascular Disease (CVD) screening based on phonocardiogram (PCG) signals has gained growing attention recently, and deep-learning techniques have been shown to provide a promising opportunity to achieve robust cardiovascular disease classification; however, the clear recognition of the diagnostically relevant information contained in the acoustic patterns requires addressing various temporal scales and the recordings are often influenced by noise and inter-subject variability. This study presents a multi-scale attention-based deep feature learning network to automatically detect normal and abnormal cardiovascular conditions by analyzing PCG recordings. The model development was performed by a retrospective computational design, which consisted of 3,163 recordings of PCGs obtained from 942 subjects, and external evaluation was performed with 3,126 independent recordings. The PCG signals were standardized, filtered, segmented and represented in short, intermediate, and long temporal scales. The parallel convolutional branches tapped on the hierarchical acoustic feature, the attention modules focused on the components with a strong diagnostic value, and the multiscale feature fusion and binary classification modules performed multiscale fusion and binary classification based on extracted hierarchical acoustic features. The proposed framework was validated with the conventional 1D-CNN, CNN-LSTM, Residual CNN, and multiscale CNN and ablated analysis was conducted to evaluate the individual effect of multiscale representation and attention learning. The accuracy, sensitivity, specificity, F1 score, and AUROC values of the proposed model were 94.3%, 92.8%, 95.8%, 94.1%, and 0.972 for the internal test set respectively. Independent evaluation showed an accuracy of 91.7%, and AUROC of 0.951, suggesting that the results were relatively stable across heterogeneous PCG recordings. Ablation analysis showed that both the single-scale learning and the full multiscale attention network improved as the number of scales increased. The results suggest an approach of analysing the PCG at multiple time scales combined with attention-driven feature refinement could help advanced automatic PCG interpretation, aid in computer-assisted cardiovascular screening, digital stethoscope applications, and telemedicine. Further clinical validation and evaluation of multiclass for the disease are recommended for routine clinical use.