A Method Of Updating ECG Classification Algorithm Based On Class Incremental Learning

Objective: Continuous heart rate monitoring and real-time heart rate detection are the main concerns of contemporary healthcare.The emergence of deep learning systems has improved the application scale and accuracy of ECG automatic diagnosis.Due to the advancement of medical technology,the understanding of diseases has gradually deepened,the classification of diseases and causes has become more refined,and the analysis of disease types has become more comprehensive.The training of existing electrocardiogram signal classification networks is based on simultaneously training datasets with all disease types.When new categories are added to the dataset,the network needs to retrain all the data,which affects the efficiency of algorithm updates.Therefore,it is necessary to efficiently update the classification categories.Method: This research designs an algorithm based on class incremental learning to update the ECG classification algorithm,uses Res Net network as the classification network,uses class incremental learning method to add training set data for training class by class,and uses the nearest neighbor method to remember the data closest to the feature vector as an example of the trained data.In the training process of increasing number of classes,the sample set and new class data form a new training set.Distillation loss function and classification loss function are used to guide the network training process to reduce the catastrophic forgetting in subsequent training.Results: This research uses the algorithm for incremental learning on two different databases.From the results,this scheme surpasses fine-tuning algorithm and Lw F algorithm,proving that this algorithm can effectively reduce the impact of catastrophic forgetting.In addition,this research further tests the network performance under different network depths and different size of memory space.The results of confusion matrix show that the classification accuracy of this method for each category is relatively close.Conclusion: Based on the experimental results,the ECG signal classification algorithm based on class incremental learning proposed in the research can update the classification category,effectively reduce the impact of catastrophic forgetting,and is superior to the commonly used fine-tuning algorithm and Lw F algorithm.At the same time,it shows relatively average accuracy in classification of various types,and has the prospect of application in larger data sets.