Epilepsy Prediction Technology Based On Two Dimensional Image Information

Epilepsy is caused by neuronal lesions in the brain and affects tens of millions of people worldwide.However,the classification of epileptic seizures is not very accurate at present,so more network models are needed to make predictions.If a model that can predict epileptic seizures using electroencephalogram(EEG)signals can be developed,it will have important implications for patients with epilepsy.This paper firstly summarizes the epileptic seizure prediction methods characterized by extracting image information at home and abroad,and also introduces the commonly used models for epileptic seizure prediction in detail.Taking time as the breakthrough point,the EEG analysis methods are summarized,and finally the focus of this paper is summarized.The relevant knowledge of the application is introduced in detail.The main results of this paper are as follows:(1)For long-range multi-channel EEG datasets,an attention-based residual network epilepsy prediction method is proposed.First,the Short Time Fourier Transform(STFT)algorithm was used to convert the 60-s EEG of each electrode in the Boston Children’s Hospital Epilepsy EEG dataset(CHB-MIT)into a two-dimensional color time-frequency image,and finally input to Classification in Residual Networks with Attention Mechanism.In view of the large number of electrodes in the data set,the channel attention mechanism in the attention mechanism is introduced into the electrode dimension,and different weights are assigned to the electrodes to capture the dependence and importance between different electrodes.A spatial attention mechanism is also introduced to perform adaptive feature refinement from the frequency band and channel dimensions.Residual networks can alleviate the problems of gradient disappearance and gradient explosion,and automatically learn features through convolution and pooling operations.In the experimental part,multiple comparative experiments were designed to evaluate the performance of the proposed model,and the final average accuracy,sensitivity and specificity were 92.07%,92% and 92.40%,respectively.(2)Aiming at the problem of severe class imbalance in EEG datasets,a seizure prediction method based on multi-band correlation and CNN-Conv LSTM hybrid network is proposed.Compared to interictal,there is more correlation between paired electrodes near seizures,so EEG multiband electrode correlations are extracted after noise reduction using the Ensemble Empirical Mode Decomposition(EEMD)algorithm characteristics to distinguish the state of the brain.Then,a hybrid network is used to simultaneously extract the temporal and spatial features of the EEG,and the EEG is studied and analyzed from multiple dimensions.In view of the imbalance of data set categories,which will lead to problems in the performance evaluation of the model,the loss function is set to Focal loss to reduce the weight of categories with more samples,so that the model pays more attention to difficult-to-classify samples during training,thereby improving the sensitivity and accuracy of prediction.In the experimental part,comparative experiments are designed to evaluate the performance of the proposed model.The average accuracy,average sensitivity,average specificity and average AUC are 90.68%,89.84%,91.26% and 0.908,respectively.