Methods For Epileptic Seizure Detection Based On Deep Learning

Electroencephalography(EEG)is a commonly used clinical approach for the diagnosis of epilepsy which is a life-threatening neurological disorder.The analysis of EEG recordings is still mainly carried out by experienced neurologists which can be time-consuming and not possible in real-time applications.To make the detection of seizures faster and more accurate,and to reduce the burden of neurologists,attention is increasingly paid to automated seizure detection from EEG signals based on deep learning.However,the success of deep learning is built upon annotated datasets of large scale.With the limitation of privacy and resources,publicly available datasets are small and few,using which deep learning methods cannot give full play to their advantages and may encounter severe overfitting.Though there are many methods and paradigms targeting the lack of data in computer vision and natural language processing,the special complicity of EEG signal limited the simple and direct application of them.Besides,inconsistency of data acquisition parameters including sampling rate and combination of electrodes across datasets,subjects,and sessions is also a critical problem.On one hand,to make adequate use of annotated samples containing seizures,which is limited in quantity,a supervised deep metric learning method is proposed to address the requirement for huge amount of annotated data which is common for conventional deep learning methods.Single-channel embedding module based on convolutional neural networks architecture is designed,and a multichannel embedding module which is improved by utilizing attention mechanism for extraction and fusion of both intra-channel and inter-channel features is also proposed.The model takes pairs of samples as input.After the deep feature extraction using embedding modules,metric learning is performed in the latent embedding space.A stage-wise training strategy is adopted to get an end-to-end classification model.Experiment shows that the proposed method performs well on both the small single-channel dataset and the more realistic multichannel dataset.On the other hand,aiming at the inconsistent combination of electrodes in EEG data and in order to decouple the model structure from certain combination of electrodes,seizure detection based on convolutional attention and electrode positional encoding is studied.First,by improving the channel-wise convolutional layer that is used to focus on the extraction of intra-channel feature,channel-independent convolutional layer is proposed via sharing weights among all channels irrespective to the electrodes.Second,convolutional attention mechanism and Convoformer module are proposed based on the Transformer networks for inter-channel feature extraction and fusion.The use of the Convoformer module also avoided the loss of temporal information which can happen with the direct use of the Transformer networks.Finally,electrode positional encoding module is also proposed to reintroduces the location information of electrodes without binding to the model structure.Experiments on public datasets verified the performance of the proposed method.Comparison among cases of random loss of channels is also conducted,and the tolerance to inconsistent combination of electrodes is tested.