Research On Seizure Warning Method Based On Deep Learning

Epilepsy,one of the most common neurological disorders,is a recurrent seizure disorder caused by abnormal electrical waves in the brain that may lead to loss of consciousness,physical convulsions and other symptoms.According to 2019 statistics from the World Health Organization(WHO),about 50 million people worldwide suffer from epilepsy,and about 80%of these patients live in low-income countries or regions.According to the Chinese Epilepsy Society,there are about 10 million people with epilepsy in China.Research studies by the World Health Organization show that seizures can be controlled by antiepileptic drugs or surgery in about 70%of patients,and the rest are patients with intractable epilepsy.For these patients with refractory epilepsy,if their seizures can be predicted in advance,the adverse consequences of seizures for the epileptic patients can be mitigated or avoided by treatments such as electrical stimulation of the nervous system.Thus,early prediction of seizures can not only reduce the suffering of patients and improve their quality of life,but also reduce the cost of medical care.Electroencephalography(EEG)is an important tool for analyzing and processing brain signals and is widely used in the diagnosis and treatment of epilepsy disorders.Epilepsy prediction technology based on EEG signals is still in the research stage,and although some research progress has been made at the algorithm level,there are still many technical and clinical application challenges.For example,issues such as how to improve the accuracy of prediction and how to apply the prediction technology to clinical practice need further research and exploration.The main task of epilepsy prediction is the identification of pre-ictal.Epilepsy prediction is more difficult and challenging compared to seizure detection.In this paper,an epilepsy prediction method based on Stockwell Transform(ST)and Convolutional Neural Network(CNN)combined with Multi-Head Attention(MHA)is proposed.First,the raw EEG recordings were data expanded to balance the two types of data and reduce the risk of overfitting..Then the EEG recordings were cut into 6-second-long EEG fragments.Secondly,the time-frequency features of EEG segments were extracted using ST,and a combined CNN and MHA model was used for classification.Finally,post-processing techniques were used to improve the classification accuracy and reduce the false positive rate of epilepsy prediction.The system was evaluated using EEG data from 23 epilepsy patients in the CHB-MIT database.The results after K-fold cross-validation showed that the segment-based accuracy was 97.49%,sensitivity was 98.24%,and specificity was 96.70%,the event-based sensitivity was 94.78%,and the false prediction rate was 0.05/h.The segment-based and event-based results showed that the proposed the epilepsy prediction method has some application value.