基于联邦学习的跨被试癫痫发作检测方法

Cross-subject epileptic seizure detection method based on federated learning

提出一种基于联邦学习的跨被试癫痫发作检测方法,以解决由于数据类型不平衡和癫痫患者之间数据分布差异显著导致的深度检测模型训练数据不足和泛化性能低的问题.建立癫痫发作检测的联邦学习框架,聚合多个训练参与方的脑电图数据;设计多尺度时间卷积网络作为客户端局部模型,通过客户端局部模型的训练和参数聚合协作训练全局模型;为避免联邦训练过程中参数量过大,用量化压缩技术提高模型的传输效率.在CHB-MIT头皮脑电图数据中评估联邦学习全局模型的跨被试癫痫发作检测性能,取得平均71.21%的灵敏度和83.99%的准确率.结果表明,联邦学习在不交换各客户端隐私数据的前提下,能够融合局部模型参数生成独立于患者个体的公共检测模型,为跨被试癫痫发作检测提供有效方法.

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.