基于卷积神经网络的多标签心电信号分类

Multi-Label ECG Signal Classification Based on Convolutional Neural Network

摘要: 针对当前主流心电图多标签分类方法在特征提取能力不足及缺乏有效标签关系建模的挑战，本研究提出了一种多标签心电图卷积神经网络分类模型。该模型包含三个关键模块：多尺度卷积模块、残差注意力模块和标签增强模块。具体而言，多尺度卷积模块能够在不同尺度上有效捕捉心电信号的空间特征，增强对不同心电图信号特征的捕捉能力。残差注意力模块结合并行注意机制，提取心电图信号中时间和空间信息，增强心电图信号中的潜在特征捕捉能力。标签增强模块通过建模标签间的相关性，进一步优化对心律失常类别的预测能力。在两个心电图数据集上进行实验，所提出模型的F1分数分别达到83.2%和73.0%。验证了所提出模型在分类性能和泛化能力上的显著优势。

Abstract: Aiming at the challenges of current mainstream ECG multi-label classification methods in terms of insufficient feature extraction capability and lack of effective label relationship modelling, this study proposes a multi-label ECG convolutional neural network classification model. The model contains three key modules: a multiscale convolution module, a residual attention module, and a label enhancement module. Specifically, the multi-scale convolution module can effectively capture the spatial features of ECG signals at different scales and enhance the ability to capture different ECG signal features. The residual attention module combines the parallel attention mechanism to extract temporal and spatial information in ECG signals and enhance the ability to capture potential features in ECG signals. The label enhancement module further optimizes the prediction ability of arrhythmia categories by modelling the correlation between labels. Experiments are conducted on two ECG datasets and the F1 score of the proposed model reaches 83.2% and 73.0%, respectively. The significant advantages of the proposed model in classification performance and generalization ability are verified.