基于孪生网络的心电信号智能诊断模型

Intelligent Diagnosis Model for ECG Signals Based on Siamese Networks

摘要: 在构建医学心电信号分类模型时，现有数据库中正样本和负样本存在类别不平衡的问题，即正常心电数据往往多于异常心电数据，从而导致所构建的模型会产生一定的偏差，影响模型最终的分类结果。利用现有12导联心电诊断数据集构建心电信号分类模型，结合对比学习的方法，在传统网络基础上利用孪生网络结构对模型进行预训练，每条心电信号分别与其对应的正例和反例进行匹配得到样本对，并将孪生网络提取的判别特征用于下游任务进行分类。实验表明，基于对比学习预训练范式能够更好地提取心电信号特征，对于克服心电数据不平衡问题有显著效果，该方法可有效地利用了大量正样本，因此基于孪生网络结构的预训练模型所提取的特征判别性能优于传统特征提取模型。测试集上的准确率达到了96.55%，假阳性率6.91%，漏诊率仅2%。

Abstract: In the field of medical ECG signal classification, the issue of class imbalance arises when the existing database contains a higher number of normal ECG data compared to abnormal ECG data. This imbalance can lead to biased models and affect the classification results. To address this problem, we propose a method combining contrastive learning with a siamese network structure to pretrain a model for ECG signal classification. In this study, we utilized a dataset of 12-lead ECG diagnostic data to construct our model. The siamese network structure was employed to pretrain the model, where each ECG signal was matched with its corresponding positive and negative samples to form pairs. The discriminative features extracted by the siamese network were then utilized for downstream classification tasks. Experimental results demonstrate that the pretraining paradigm based on contrastive learning effectively extracts discriminative features from ECG signals and significantly overcomes the issue of data imbalance. This method successfully utilizes a large number of positive samples, resulting in superior discriminative performance of the feature extraction model compared to traditional methods. The accuracy of the test set reached 96.55%, with a false positive rate of 6.91% and a false negative rate of only 2%. Overall, our study demonstrates that utilizing a sia-mese network structure for pretraining, combined with contrastive learning, improves the feature extraction process for ECG signal classification. This approach effectively addresses the issue of data imbalance, resulting in highly accurate classification results.