Advanced heart disease classification based on multi-channel heart sound coupling features.

Conventional heart sound classification methods often rely on single-channel, one-dimensional feature extraction, which inadequately captures pathological relationships across different auscultation zones, thereby limiting the accuracy of heart disease detection. To address this issue, a novel classification framework based on multi-channel heart sound coupling feature extraction is proposed to enhance heart disease identification. This approach begins with denoising preprocessing applied to four-channel heart sound signals and a single-channel electrocardiogram. These five-channel signals are systematically paired to extract five types of coupling features, resulting in 130 distinct features per multi-channel sample. The ReliefF algorithm is then used to evaluate feature importance, retaining the top 20% of features to construct a coupling feature set. A convolutional neural network is employed to classify normal and abnormal heart sounds. When applied to clinical congenital heart disease datasets, the proposed method achieved a classification accuracy of 95.6%, while on the PhysioNet heart sound challenge dataset, it reached an accuracy of 98.3%. Experimental results demonstrate that compared to single-channel, one-dimensional features, multi-channel coupling features more effectively capture pathological characteristics in heart sound signals, significantly improving the accuracy of heart disease classification and addressing challenges in the refined categorization of cardiac conditions.