基于深度学习的多标签不均衡心电信号分类

Multi-Label Imbalanced ECG Signal Classification Based on Deep Learning

摘要: 心电图(ECG)自动分类技术已经成为医学诊断的热门研究领域。在临床上，同一个患者通常伴随多种心律失常。为了能够精准分类出多种心律失常，本文提出了一种涵盖多种注意力机制的一维卷积神经网络模型。相比传统的机器学习或者深度学习方法提取特征的能力有限，引入的注意力机制能够引导模型学习心电信号中不同通道和空间位置之间的差异信息的同时结合全局稀疏自注意力，提高分类仿真模型对与心律失常有关特征的关注程度；其次，通过重采样技术以及在损失函数的设计上融入了标签相关性策略来应对多标签带来的类别不均衡问题。在两个公开的多标签心电数据集CPSC2018和PTB-XL上评估该网络的性能，macro-F1分别达到84.13%和73.6%。实验结果表明，该模型具有良好的分类性能。

Abstract: The technology of automatic Electrocardiogram (ECG) classification has emerged as a hot area of research in medical diagnostics. In clinical practice, a single patient often exhibits multiple types of cardiac arrhythmias. To accurately classify these diverse arrhythmias, this paper introduces a one-dimensional Convolutional Neural Network model that incorporates multiple attention mechanisms. Compared with the limited ability of traditional machine learning or deep learning methods to extract features, the introduced attention mechanism can guide the model to learn the difference information between different channels and spatial locations in ECG signals, and combine with global sparse self-attention to improve the classification simulation model’s attention to arrhythmia-related features. We address the class imbalance problem brought about by multi-labels through resampling techniques and the integration of label correlation strategies into the design of the loss function. The performance of this network is evaluated on two public multi-label ECG datasets, CPSC2018 and PTB-XL, with macro-F1 reaching 84.13% and 73.6%, respectively. Experimental results indicate that the model demonstrates excellent classification performance.