Recognition Of High Frequency Oscillation Of Electroencephalogram And Deep Prediction Model Of Epileptic Seizure

Epilepsy is a very common chronic systemic neurological disease,surgical treatment of epilepsy can help drug-resistant patients to get rid of the pain and physical damage during seizures,but inaccurate localization of the epileptogenic focus limits the effectiveness of surgical treatment.At present,the localization of epileptogenic focus mainly relies on the visual detection of high-frequency oscillation(HFO)components in electroencephalography.However,due to the hidden manifestations of HFO,it is difficult to accurately detect epileptogenic focus and the efficiency is extremely low.At the same time,uncured patients can use seizure prediction to avoid safety hazards caused by sudden seizures.However,currently limited by the influence of prior knowledge feature extraction and the accuracy of preictal stage labels,it is difficult to achieve accurate real-time seizure prediction.To solve the above problems,this paper uses experimental data of epilepsy to build deep learning models to realize automatic HFO recognition in EEG and end-to-end seizure prediction,so as to improve the localization efficiency and the realization ability of seizure prediction.Firstly,in this thesis,a 2D convolutional neural network model based on color time-frequency map input is proposed to improve the automatic recognition performance of HFO.The time-frequency image of the EEG signal is divided into frequency bands and the time-frequency image is independently normalized and then fused to construct a color time-frequency image,which enhances the characteristic expressiveness of the high-frequency components of the time-frequency image and effectively separates HFO from the background waves.Identification and validation of the model on the acquired stereoelectroencephalogram(SEEG)of patients showed that it is effective to approximate the epileptogenic zone using the automatically identified HFO-generating zone,thus improving the efficiency of clinical localization.Secondly,a distributed ensemble model is designed to realize epileptic seizure prediction without prior feature extraction and preictal stage labels.The distributed models are constructed by drawing on distributed networks of brain,the convolutional auto-encoding network is used to perform dimensionality reduction coding on SEEG signal of HFO-generating zone,the convolutional long short-term memory network is used to predict dimensionality reduction coding,and the fully connected network is used to discriminate the seizure of predictive coding.Epileptic seizure prediction is realized by integrating above networks.Finally,based on the above-mentioned deep prediction model of epileptic seizure,a mobile epileptic seizure prediction system based on Raspberry Pi is constructed.Inference calculation of the deep learning model is performed through the ARM-CPU of the Raspberry Pi,and the model is appropriately quantified before deployment,so as to improve the inference speed of the model on the ARM chip under the premise of ensuring a relatively balanced calculation speed and performance.In addition,a wireless real-time receiver module is built under the seizure prediction system to provide a port to connect to the acquisition device and receive EEG data wirelessly,and an interactive interface display module is built for display of waveforms and prediction results and user operation.The HFO automatic identification method and epileptic seizure prediction system proposed in this paper provide new research ideas for high-frequency oscillation signals and EEG prediction,and help to promote the implementation of mobile terminal deployment of EEG-related deep models.