基于时序心脏模型样本均衡方法的心律失常分类

心律失常自动分类作为计算机在临床上的重要应用,可以有效辅助心血管疾病的诊断,但实验中样本不均衡问题严重影响分类精度。目前用于解决样本不均衡问题的主流方法为对抗神经网络,但存在训练不稳定和模式崩溃等问题,且仅依靠数据进行学习,缺乏一定的生理意义。因此提出基于时序心脏模型的样本均衡方法生成心电数据,在2018年中国生理信号挑战赛提供的12导联数据集上进行实验,采用深度残差网络作为分类模型分别对每个导联进行训练,通过极端梯度提升算法实现导联融合。经过样本均衡后,各类F1分数均有提升,左束支阻滞(LBBB)、ST段降低(STD)、ST段抬升(STE)的改善尤其显著,分别由扩增前的0.706、0.684、0.524提升至0.832、0.809、0.618。为验证本方法的通用性,对PTB数据集进行独立测试,分类准确率达到98.64%。实验结果表明,基于时序心脏模型生成仿真数据能够有效改善实验样本不均衡现象。

Arrhythmia Classification Based on Time-Series Cardiac Model Sample Equalization

As an important clinical application of computer, automatic classification of arrhythmia can effectivelyassist in the diagnosis of cardiovascular diseases. However, the sample imbalance in experiments seriously affects the classification accuracy. At present, the mainstream method to solve the problem of sample imbalance is counter neural network, but it has some problems such as unstable training and mode collapse, and only relies on data for learning, which is lack of certain physiological significance. Therefore, this paper proposed a sample equalization method based on the time-series cardiac model to generate ECG data. The experiment was carried out on the 12 lead dataset provided by China physiological signal challenge (CPSC) in 2018. The deep residual network was used as the classification network to train each lead, and the lead fusion was realized by XGBoost algorithm. After sample equalization, F1 scores of all types were improved, especially in left bundle branch block (LBBB), ST segment depression (STD) and ST segment elevation (STE), which was increased from 0.706, 0.684 and 0.524 to 0.832, 0.809 and 0.618, respectively. In order to verify the universality of this method, PTB dataset was tested independently, and the classification accuracy reached 97.42%. The experimental results showed that the generation of simulation data based on the sequential heart model effectively improved the imbalance of experimental samples.