Research On Mechanism And Control Of Absence Seizure Based On Machine Learning

Absence epilepsy is a kind of idiopathic generalized epilepsy.Its main characteristics are spontaneity,transience,repeatability of absence seizure.Clinically,the generalized spike-wave discharges can be recorded simultaneously in both hemispheres of the brain,accompanied with brief and sudden lapse of consciousness.So far,it is not clear about the mechanism of absence epilepsy.The absence seizure is ictal,and there exists differences among epileptic even among seizures of one patient.As a result,it is difficult to treat absence epilepsy.In our study,we will apply the machine learning methods into the EEG signal and neural mass model of absence seizure to explore the multiple mechanism of seizure onset,predict and design adaptive brain stimulation to inhibit spike-and-wave discharge in absence seizure.Firstly,we propose a statistical framework based on the decision tree and random forest to analysis dynamics of a neural mass model of absence seizure in its highdimensional parameter space,and quantify the importance of parameters for the transition between different dynamics in this model.The results show that the onset of seizure may be the result of multiple factors,including the input of cortical cells and synaptic strength among them,especially the input of pyramidal neurons.These results are in favor of the opinion that the cortex is more important than thalamus for the generation of spike-wave dischages.Secondly,we present a reinforcement learning algorithm that optimizes stimulation strategy for controlling absence senzure with minimum stimulation energy.We apply our method to a neural mass model which simulates inter-ictal and ictal EEG data.With a specialized reward function and state-space discretization fixed,we test the effectiveness of the Actor-Critic reinforcement learing algorithm.The results show that this algorithm is able to identify parameters that dectect and control seizures quickly.Additionally,the Actor-Critic algorithm has an advantage that it is adaptive so that the stimulation signal can change over time.Compared with a proposed optimal control method for absence seizure abatement,our method has much applicability and comsume less stimulation energy.Lastly,we propose a method combined the visibility graph method and convolutional neural network to predict epileptic seizures based on the EEG signal.In the preprocess,we introduce a modified visibility graph method to transform the EEG signal to images.And then a convolutional neural network with inception module is build to recognize the interictal EEG and preictal EEG.The proposed approach achieves sensitivity of80% and a false prediction rate of 0.21/h on the test EEG dataset,and it is also statistically better than an unspecific random predictor for most of the patients in this dataset.compared to some proposed approaches for predict seizures,our method is able to refine the best features from images automatically instead of artificial features extraction so that it is more efficient,and has better generalization capacity.