廖桂鑫 甘力
（广东工业大学，广东 广州 510006）
摘要：针对网络轻量化后，模型复杂度降低带来的检测性能不足等问题，提出一种基于自适应卷积的心电图（ECG）心律分类方法。首先，采用轻量级的卷积神经网络模型框架构建双分支结构，主分支提取ECG的波形特征，子分支提取ECG样本与正常心律的差异信息；然后，通过自适应卷积的方法，将ECG样本与正常心律的差异信息融入到主分支中，提高模型的检测性能；最后，在公开的数据集上进行实验，F1分数、准确率、召回率分别为93.58%、95.53%和91.70%，相较于未加入ECG样本与正常心律的差异信息的网络有明显提升，验证了该方法的有效性。
关键词：心电图；心律分类；轻量级卷积神经网络；自适应卷积；双分支结构
中图分类号：TP391 文献标志码：A 文章编号：1674-2605(2023)05-0011-06
DOI：10.3969/j.issn.1674-2605.2023.05.011
ECG Rhythm Classification Method Based on Adaptive Convolution
LIAO Guixin GAN Li
(Guangdong University of Technology, Guangzhou 510006, China)
Abstract: A heart rhythm classification method for electrocardiogram (ECG) based on adaptive convolution is proposed to address the issues of insufficient detection performance caused by reduced model complexity after network lightweighting. Firstly, a lightweight convolutional neural network model framework is used to construct a dual branch structure. The main branch extracts the waveform features of ECG, while the sub branches extract the difference information between ECG samples and normal heart rhythm; Then, by using adaptive convolution method, the difference information between ECG samples and normal heart rhythm is integrated into the main branch to improve the detection performance of the model; Finally, experiments were conducted on publicly available datasets, and the F1 score, accuracy, and recall rates were 93.58%, 95.53%, and 91.70%, respectively. This showed a significant improvement compared to the network that did not include the difference information between ECG samples and normal heart rhythm, verifying the effectiveness of this method.
Keywords: electrocardiogram; heart rhythm classification; lightweight convolutional networks; adaptive convolution; double branch structure
