A cross-subject seizure detection method based on federated learning was proposed to address the issues of insufficient training data and low generalization performance of deep detection models,resulting from imbalanced data types and significant differences in data distribution among epilepsy patients.A federated learning framework was established for epileptic seizure detection to aggregate elec-troencephalogram data from multiple training participants.A multi-scale time convolutional network was designed as the client local model,and the global model trained collaboratively through the training and parameter aggregation of the client local model.To avoid excessive parameters during federated training,quantization compression technology was introduced to improve the transmission efficiency of the model.The performance of the federated global learning model for cross-subject seizure detection was evaluated on the CHB-MIT scalp electroencephalogram dataset,and an average of 71.21%sensitivity and 83.99%accuracy achieved.The results showed that federated learning could fuse local model parameters to generate a public detection model being independent of individual patients without exchanging the private data of each client,which can provide an effective method for cross-subject epileptic seizure detection.
维普
万方数据
