基于多模态特征融合与集成学习的心音信号分类研究

Heart Sound Signal Classification Based on Multi-Modal Feature Fusion and Ensemble Learning

摘要: 针对心音分类中存在的数据不平衡、特征提取不足与模型泛化能力有限等问题，本文提出了一种基于多模态特征融合与高级集成学习的心音异常检测方法。首先，采用Challenge2016数据集，通过带通滤波与归一化预处理，并提出异常类别导向的数据增强策略以解决样本不平衡问题。其次，综合提取时域统计、频域谱质心及小波域能量熵等多维特征，构建全方位特征空间。在模型构建上，设计了CNN-BiLSTM深度神经网络以捕捉局部与长时依赖特征，并结合支持向量机(SVM)的稳定性优势。最后，提出包含动态加权、置信度加权及元学习等六种集成策略。实验结果表明，元学习集成策略性能最佳，准确率达98.89%，F1-score为0.9872。该方法在保证高精度的同时也具备良好的鲁棒性，为心脏瓣膜疾病的智能诊断提供了有效方案。

Abstract: To address the challenges of data imbalance, insufficient feature extraction, and limited model generalization in heart sound classification, this paper proposes a heart sound anomaly detection method based on multi-modal feature fusion and advanced ensemble learning. Firstly, using the Challenge2016 dataset, data preprocessing including band-pass filtering and normalization is applied, and an anomaly-oriented data augmentation strategy is proposed to mitigate sample imbalance. Secondly, multi-dimensional features, including time-domain statistics, frequency-domain spectral centroid, and wavelet-domain energy entropy, are extracted to construct a comprehensive feature space. In terms of model construction, a CNN-BiLSTM deep neural network is designed to capture local and long-term dependency features, combined with the stability of Support Vector Machine (SVM). Finally, six ensemble strategies, including dynamic weighting, confidence weighting, and meta-learning, are proposed. Experimental results show that the meta-learning ensemble strategy performs the best, with an accuracy of 98.89% and an F1-score of 0.9872. This method ensures high accuracy while maintaining good robustness, providing an effective solution for intelligent diagnosis of valvular heart diseases.