On the premise of ensuring system accuracy,lightweight models can be deployed on embedded devices or mobile terminals with limited hardware resources.Therefore,a method using modified lightweight residual network is proposed for arrhythmia classification.The method transforms one-dimensional electrocardiogram data into Gramian angular summation field maps which are then taken as the model input,and reduces the number of model parameters by substituting ShuffleNet V2 convolutional units for the traditional convolution inside the ResNet34 basic residual blocks.In addition,the network incorporating efficient channel attention module makes the model focus on important feature regions,thereby improving model accuracy and realizing the automatic arrhythmia classification.The proposed model has an accuracy of 99.78%on MIT-BIH arrhythmia database,and it reduces the number of parameters,FLOPs and MAdd by 95%,91%and 91%,as compared with the traditional ResNet34 model,demonstrating its characteristics of lightweight and high accuracy,and proving the possibility of deployment on mobile devices.