基于多域特征融合与通道注意力的异常心音检测

Anomaly Heart Sound Detection Based on Multi-Domain Feature Fusion and Channel Attention

摘要: 心音信号蕴含丰富的生理与病理信息，是辅助心血管疾病筛查与早期诊断的重要生物特征。传统心音分析方法通常依赖单一特征或手工规则，难以充分刻画时频结构的复杂变化，导致在多噪声环境和不同个体差异下识别准确率受限。为此，本文提出一种基于多域特征融合与通道注意力机制的心音异常检测卷积网络(TTFI-CNN)。该网络综合利用MFCC (Mel频率倒谱系数)、STFT (短时傅里叶变换)与DWT (离散小波)三类互补特征，构建“时域–频域–时频域”的多特征表示，并设计轻量化SE (Squeeze-and-Excitation)通道注意力模块增强特征选择能力。此外，模型还结合EarlyStopping与交叉验证设计，以提升泛化能力。基于2016 PhysioNet/CinC心音挑战赛数据集进行验证，所提出的模型在异常心音检测任务上取得97.8%的准确率。其中，消融实验证明：去除DWT分支会导致准确率下降至96.4%，去除SE注意力则下降至97.09%，充分说明三分支融合与通道注意力在提升分类性能方面的有效性。实验结果表明，本方法能够稳健地捕获多尺度时频特征，对心音异常检测具有良好的应用潜力。

Abstract: Heart sound signals contain rich physiological and pathological information, serving as crucial biomarkers for assisting in the screening and early diagnosis of cardiovascular diseases. Traditional heart sound analysis methods typically rely on single features or manually designed rules, making it difficult to fully characterize the complex variations in time-frequency structures. This often results in limited recognition accuracy in environments with high noise levels or significant individual differences. To address this, this paper proposes a time-frequency feature fusion and channel attention-based convolutional neural network (TTFI-CNN) for abnormal heart sound detection. The network comprehensively utilizes three complementary features—Mel-Frequency Cepstral Coefficients (MFCC), Short-Time Fourier Transform (STFT), and Discrete Wavelet Transform (DWT)—to construct a multi-feature representation spanning the “time-domain, frequency-domain, and time-frequency domain”. Additionally, a lightweight Squeeze-and-Excitation (SE) channel attention module is designed to enhance feature selection capabilities. Furthermore, the model incorporates EarlyStopping and cross-validation strategies to improve generalization. Validated on the 2016 PhysioNet/CinC heart sound challenge dataset, the proposed model achieves an accuracy of 97.8% in the task of abnormal heart sound detection. Ablation studies demonstrate that removing the DWT branch reduces accuracy to 96.4%, while removing the SE attention module lowers it to 97.09%, fully confirming the effectiveness of the three-branch fusion and channel attention in enhancing classification performance. The experimental results indicate that this method can robustly capture multi-scale time-frequency features, demonstrating promising potential for application in abnormal heart sound detection.