The Research Of Remote Online Pre-ictal Prediction For Epilepsy Under Strong-noise EEG Based On Multi-source Information Fusion With CHMM

Epilepsy is a chronic disease with disorder of nervous system, which can lead to a loss of consciousness and muscle twitch. At present, about6millions of patients have epilepsy in the global scope, while about6～9millions of patients with epilepsy in China. If seizure can be effectively predicted, patients and doctors can make timely and effective protective measures, to reduce the harm to patients and improve the quality of their life. Therefore, the research of prediction of epilepsy under strong-noise EEG means great significant.The study on pre-ictal prediction of epilepsy are mostly in the algorithm simulation and static modeling phase. At present, the relevant report of online pre-ictal prediction of epilepsy, which can be used in clinical trials, is much less. Therefore, the research of online pre-ictal prediction of epilepsy is indispensable.The main purpose of this paper is to model pre-ictal EEG under strong-noise dynamically and build remote on-line pre-ictal prediction system for a single server but multiple clients system based on automatic identification of epileptic seizures. The aim is also to improve prediction accuracy and prediction efficiency of the system and reduce missed prediction rate and error prediction rate. Continuous Hidden Markov (CHMM) is used to model strong-noise dynamic pre-ictal EEG, multi-source information fusion of linear and nonlinear is employed to reconstruct the input of CHMM. Experiments were doing under interactive training mode, feature modeling mode and remote on-line pre-ictal prediction mode. Experimental results of interactive training model show that:When N varies from2to5, it only takes at most467ms to train a CHMM model for a segment of EEG signal with5s sample points. Experimental results of feature modeling indicate:The fusion of Hurst Exponent, Approximate Entropy and CO complexity performs best for distinguishing pre-ictal state from ictal state and normal state. Pre-ictal state can be predicted with the accuracy of85.74%±2.43averagely. Specification and sensitivity accuracy of the model are found as69%and89%respectively. Furthermore, efficiency of this model can reach the level of millisecond, it takes almost193ms to recognize an unknown EEG segment. The experimental results of remote on-line model demonstrate that the prediction accuracy is100%, which means pre-ictal state can be recognized in every EEG segment. Prediction time is23.71±5.85s, which means epileptic seizure will occur in23.71±5.85s,and operating efficiency of the system is17.79±11.99s. Meanwhile, missed prediction rate and error prediction rate of this model are all zero. Therefore, the parameters of the system can achieve ideal level.This remote online pre-ictal prediction system is composed of interactive training mode, off-line analysis model and remote on-line pre-ictal prediction model, and it enjoys good prediction accuracy, prediction time and operating efficiency. Furthermore, the parameters of the system are adjustable, so it has strong adaptability. The system has wide application prospect in the diagnosis of epilepsy, remote detection, brain science and cognitive domain.