A Study Of Federated Learning Algorithms In ECG Signal Classification Scenarios

Cardiovascular disease is one of the highest mortality rates in the world,and ECG signals are an effective means of diagnosing cardiovascular disease in the clinic.Using deep learning algorithm models to identify and classify ECG signals can share the pressure of medical workers and relieve the strained medical resources.The ECG signal classification model based on deep learning algorithm needs to collect a large amount of ECG signal data for training,however,ECG signal data contains patient privacy,and it is difficult to share ECG signal data between hospitals and medical institutions.Therefore,this paper introduces the federated learning algorithm into the training process of ECG signal classification model to realize the safe sharing of data.Meanwhile,the traditional federated learning algorithm is improved to alleviate the impact of non-independent identically distributed data of client on the model performance.The main work and research of this paper are as follows:(1)To address the problem of difficulty in sharing ECG signal data among different hospitals or medical institutions,this paper introduces a federated learning algorithm instead of the traditional centralized model training,so that the data of each participating hospital or medical institution can be stored locally to reduce the risk of ECG signal data leakage,The model parameter aggregation strategy in Federated Averaging Algorithm(FedAvg)is improved to determine the model parameter weights according to the client data timeliness.The experimental results show that the ECG signal classification model based on federated learning has good performance while protecting data security,while the improved model parameter aggregation strategy further improves the classification accuracy of the model.(2)To address the problem that(hospital,medical structure)non-independent identically distributed data of client can affect the performance of the federated learning algorithm,the client-side model training process is optimized,and a correction term is added to the objective function of the client-side training model to limit the client-side model update and prevent the client-side local model from deviating from the global model,and the size of the correction term is dynamically adjusted according to the timeliness of the client-side data.The experimental results show that the optimized client-side training process has better performance than the traditional federated learning algorithm.(3)To further mitigate the impact of non-independent identically distributed data of client on model performance,the idea of contrast learning is introduced into the federated learning algorithm,and the contrast federated learning algorithm is proposed.Contrast loss is added to the client-side model training process,and the magnitude of contrast loss is calculated based on the difference between the client-side local model parameters and the server global model parameters to prevent the update of the local model from deviating from the global model.The experimental results show that the performance of the contrast federated learning algorithm is better compared with other federated learning algorithms,and the model classification accuracy is higher than that of the single-point training model and close to the accuracy of the centralized training algorithm.In summary,the federated learning algorithm reduces the risk of privacy leakage during data sharing and achieves secure data sharing,the accuracy of the improved federated learning algorithm is higher than that of the traditional federated learning algorithm,and the model algorithm performance is close to that of the centrally trained model.