Fast Convolution Compression Network for Coronary Artery Disease Detection Using Auscultation Signal

Heart sound auscultation combined with deep learning is a common method for coronary artery disease (CAD) detection. However, current studies predominantly focus on improving network accuracy while neglecting the lightweight structure, especially the runtime. To address the limitations, we propose a fast convolution compression network (FCCN) for automated CAD severity classification. Our experimental dataset comprises 150 clinical heart sound recordings with varying degrees of coronary stenosis, including 80 samples from severe CAD cases and 70 from nonsevere cases. The large tied one-shot aggregation convolution (LTOAC) module is proposed in FCCN, which utilizes shared convolutional filters and concise feature aggregation to improve feature utilization efficiency. FCCN integrates feature extraction and pattern recognition through an end-to-end framework without excessive speed latency and parameter costs. Experiment is performed on the dataset and demonstrates FCCN’s performance, achieving an accuracy of 85.82%, sensitivity of 85.3%, and specificity of 86.26% with 1.9 million parameters. The system balances model complexity with classification performance through parameter-efficient design. Our study based on clinical practice, provides an effective and fast method for CAD detection.