Epileptic Seizure Prediction Using Deep Learning

Epilepsy is a neurological disorder caused by abnormal neuronal discharges in the brain,and the transient occurrences of unexpected seizures significantly affect the life quality of patients.Previous studies have shown that subtle changes may exist in brain patterns before the seizure onset.Combining signal processing and pattern recognition techniques to analyze electroencephalography(EEG)signals can provide an early warning of seizures.The EEG signal of epilepsy patients is a nonlinear and non-smooth time series,which has four characteristics: the signal has strong randomness and time-variability;the signal contains a large amount of artifact and noise;the sample distribution varies greatly among patients;the multi-channel signal contains spatio-temporal dependencies.These four features derive four types of problems to be investigated in the area of seizure prediction: automatic feature extraction,artifact/noise removal,inter-patient domain adaptation,and spatio-temporal representation.The rapid development of deep learning techniques in recent years has brought breakthroughs in the field of seizure prediction.Inspired by the encouraging success,this dissertation focuses on the aforementioned problems to provide a theoretical basis for developing implantable/wearable seizure warning devices.The main research contents and contributions are as follows.(1)In terms of automatic feature extraction,conventional methods usually use fixed parameters and patterns to learn representations from complex and variable EEG data.However,this constant strategy cannot adapt to the dynamic environment of the sample space.This dissertation proposes a seizure prediction method based on the Fourier neural network(FNN),which can extract time-frequency domain features adaptively.The FNN can automatically update the node parameters in the feature extraction module by backpropagation,so that the time-frequency domain representations can be learned stably from the changing data pattern.The model further introduces spectral power ratio and convolutional neural network(CNN)to construct a multi-feature fusion serial framework,which can enhance the model performance by increasing feature diversity.This method shows good generalization ability on the highly nonlinear EEG sample space.It achieves high sensitivity of 93.59% and 92.11% on the intracranial/scalp public EEG datasets,respectively.The precision is improved by about 9% compared with conventional methods.Experimental results indicate that the proposed model is suitable for application scenarios with high prediction accuracy requirements.(2)In terms of artifact/noise removal,conventional methods usually rely on the human visual to identify artifact components after decomposing EEG signals.Most automatic denoising methods can process only one type of noise instead of multi-source noises in the clinic.In this dissertation,a seizure prediction model robust to multisource noise is proposed based on the architecture of denoising adversarial autoencoders(DAAE).This method adds multiple noises to EEG signals with the corruption module of DAAE.The autoencoder is trained to reconstruct the original data over the noiseenhanced samples.The network learns the noise reduction ability during the process.The model also builds an adversarial sub-network to project the hidden code of DAAE into a known distribution.The encoding-decoding structure is optimized using convolutional and deconvolutional layers to accelerate the convergence.The method is tested using a dataset contaminated with electrooculogram,electromyography,and hardware noises.A prediction sensitivity of >84% is achieved.The experimental results illustrate that this model has good practicality and robustness.It is suitable for application scenarios with severe noise interference.(3)In terms of inter-patient domain adaptation,conventional methods usually perform the training and testing process independently on the same subject(domain)to evade the domain gap between patients.This strategy requires a large number of EEG samples from the target patient.However,it is difficult to perform long-range(> 24h)sampling for each subject in practice,so training with other patients’ data is required.This dissertation proposes an epileptic EEG classification model based on the Riemannian manifold,which can be domain adapted among variant sample distributions.The model enhances the differentiability of EEG samples by projecting them into Riemannian space with covariance matrices.The cross-domain features are described explicitly using the parallel transport property.The inter-domain distance measure is constructed with the Riemannian mean.The model can learn the universal distribution from different sample spaces by minimizing the inter-domain distance.In addition,the model builds a multi-constraint network based on the architecture of adversarial autoencoders.The distribution alignment,inter-domain distance minimization,and classifier error are used to jointly optimize the latent layer of the autoencoder,which aims to learn the general representations among domains.The test environment simulates a clinical data acquisition situation by using other patients’ data for training.The method achieves a prediction sensitivity of 80.36%,reducing the data size restriction while maintaining an above-par precision.It is suitable for application scenarios where the amount of target patient data is insufficient.(4)In terms of spatio-temporal representation,conventional methods usually do not analyze EEG signals of each channel in grades.However,only several encephalic regions will show hyper-and desynchronized abnormal discharges before upcoming seizures,and these channels should be monitored particularly.In addition,these methods only focus on the spatial sparsity description of electrodes but ignore the long-term and short-term information in the temporal dimension.This dissertation proposes a seizure prediction model based on the convolution block attention module(CBAM)and bidirectional-Long short term memory(Bi-LSTM)networks,which uses high-resolution spatio-temporal features to capture disordered discharges of neurons.This model introduces an attention mechanism using the CBAM spatial domain module.The convolution kernel moves across channels on a spatial subunit,outputting spatial features in the receptive field.The spatial feature stream from CBAM is then transmitted to the Bi-LSTM to analyze its temporal features.Each LSTM unit learns both long-term and short-term information.The model is tested using stereo-electroencephalography data,which is widely used in the clinic.The proposed method achieves a high sensitivity of 95.83%,indicating that it has significant advantages in processing multi-channel EEG data.The method effectively improves the spatio-temporal resolution and is suitable for application scenarios using multi-channel EEG signals.