基于单通道PPG信号的房颤检测深度学习模型

Deep Learning Model for Atrial Fibrillation Detection Based on Single-Channel PPG Signal

摘要: 心血管疾病早期诊断具有重要意义，PPG信号结合深度学习技术为早期筛查提供了新途径。文章介绍了一种基于单通道光电容积脉搏波(PPG)信号的心血管疾病分类深度学习模型——ResGNet。ResGNet模型由数据预输入层、并行特征提取层和分类决策层三个核心模块构成。在特征提取阶段，采用了改进型ResNet与双向门控循环单元(BiGRU)的并行架构，分别用于捕捉PPG信号的空间和时间特征。通过挤压–激励(SE)注意力机制增强关键特征表示，并使用多层感知机(MLP)进行非线性映射，最终采用softmax函数输出分类结果。实验表明，在MIMIC III、MIMIC PERform AF及Arrhythmia Detection三个数据集上，ResGNet模型的准确率分别达到99.34%、98.91%和96.51%，显示出卓越的分类性能。特别是在复杂的心律失常分类中，相较于经典深度学习模型实现了更高的精确度和灵敏度。

Abstract: Early diagnosis of cardiovascular diseases is of great significance, and PPG signals combined with deep learning technology provide a new way for early screening. This paper introduces a deep learning model for cardiovascular disease classification based on single-channel photoplethysmography (PPG) signals—ResGNet. The ResGNet model consists of three core modules: the data pre-input layer, the parallel feature extraction layer, and the classification decision layer. In the feature extraction stage, the parallel architecture of improved ResNet and Bidirectional Gated Recurrent Unit (BiGRU) was used to capture the spatial and temporal features of PPG signals, respectively. The key feature representation is enhanced by the extrusion-excitation (SE) attention mechanism, the multilayer perceptron (MLP) is used for nonlinear mapping, and finally, the softmax function is used to output the classification results. Experiments show that the accuracy of the ResGNet model reaches 99.34%, 98.91%, and 96.51% on the three datasets of MIMIC III, MIMIC PERform AF, and Arrhythmia Detection, respectively, showing excellent classification performance. Especially in the classification of complex arrhythmias, it has shown higher accuracy and sensitivity than classical deep learning models.