Seizure Detection And Prediction Based On Deep Learning

Epilepsy is a common neurological disorder that affects about 1%of the world’s population.Once patients have an attack,they will suffer from loss of consciousness,generalized convulsions,and mental abnormalities,which seriously threaten the life and health safety of patients.The pathogenesis of epilepsy is very complex and difficult to treat,and commonly used treatments include medication,surgery,and neurological stimulation and modulation,but about one-third of patients with intractable epilepsy are still not effectively treated.Currently,in neurology,EEG analysis and epilepsy detection are generally done by specialized physicians based on visual observation of clinical experience,a task that is very time-consuming and laborintensive and does not guarantee a stable accuracy rate.Therefore,exploring the development of an automatic epilepsy detection system based on EEG signals can not only reduce the workload of physicians,but also help improve the accuracy of epilepsy diagnosis.In addition,for patients who have been diagnosed with epilepsy disease,related studies have found that seizures can be predicted in advance,so seizure prediction can be performed by EEG signals,and early warning signals can be given before a patient has a seizure to inform the patient or healthcare professionals to take appropriate measures,which can reduce the harm caused by seizures to patients.In this paper,an epilepsy detection and prediction model based on deep learning is proposed.In order to improve the stability and accuracy of the model,this study combine manual feature extraction techniques and automatic feature extraction techniques of neural networks,not only manually designed to extract classical feature information,such as max cross-correlation values and phase locking synchrony values,but also use a residual network based on dilated convolution to mine the deep-level feature information of epileptic EEG,and then use the attention mechanism to filter the feature information and further increase the interclass.After that,this study input the feature information to the bidirectional recurrent neural networks(bidirectional LSTM and bidirectional GRU)for classification and recognition,and finally the output results are smoothed by "k-of-n" analytical smoothing of the output results,which effectively reduces the false positive rate.In this paper,we tested this model using the open-source scalp EEG dataset CHB-MIT,and obtained 98.07%accuracy and 93.33%sensitivity in the epilepsy detection task,and obtained 96.25%accuracy and 93.92%sensitivity in the epilepsy prediction task.The experimental results show that the research in this paper helps to promote the further application of deep learning models in the field of epilepsy detection and prediction,and effectively promote the development of seizure detection and prediction technology based on EEG signals,which has broad research prospects.