Research And FPGA Implementation Of Epileptic Sei-zure Prediction Based On Deep Learning

Epilepsy,also known as epilepsy,can cause symptoms such as blurred consciousness,mental disorders,and systemic spasms once a patient develops symptoms.Patients diagnosed with epilepsy can predict seizures through EEG signals,provide warning signals before the onset,and patients or medical staff can take appropriate measures to reduce secondary injuries.Currently,researchers have proposed various excellent methods for predicting epileptic seizures,but there are also the following problems: firstly,the public epilepsy dataset information is limited,and most researchers are unable to obtain more effective data for model validation;Secondly,research often uses a single neural network framework.Although the model has a simple structure and strong explanatory power,its learning ability is limited and it cannot train and fit well on large-scale data.In response to the current shortcomings,the main work of this study is as follows:(1)To address the impact of high-frequency noise and other artifacts on the prediction results of the original epileptic EEG signal,a fifth order Butterworth bandpass filter was used for filtering processing.At the same time,considering the differences in EEG operating frequencies among different samples,it was not possible to manually select a specific frequency range for each sample.A feature extractor was designed,which includes two parts: FBCSP feature extraction and LDA feature classification.Using a CNN network that performs well on public datasets for training validation and result analysis on the preprocessed dataset,considering that epileptic EEG is a one-dimensional signal and one-dimensional convolution can retain more local feature information of the data,CNN is improved to obtain a one-dimensional epileptic seizure prediction model ODEP.The experimental results indicate that the ODEP model not only improves prediction performance but also accelerates prediction speed.(2)Due to the fact that the number of datasets far exceeds that of public datasets,the prediction performance of the ODEP model did not meet expectations.Based on this,the ODEP-ATT-LSTM prediction model is proposed.Among them,ODEP-ATT constitutes a feature learning network,effective learning channels and deep features in the time dimension,carries out adaptive feature refinement processing on EEG signals,and then classifies and predicts them in combination with short-term memory network(LSTM),and smoothes the output results with "k-of-n" smoothing.Using the improved Focal Loss function as the loss function,the final average prediction accuracy reached 82.65%,and the model was applied to the CHB-MIT epilepsy EEG database,and compared with other algorithms to verify the effectiveness of the model.(3)When applied in the medical field,it is necessary to consider both the overall processing speed and portability of the system.This study adopts the PYNQ-Z2 platform for the hardware design and implementation of the epilepsy prediction algorithm,and builds a verification platform to comprehensively verify the hardware system.Finally,the resource utilization and speed performance of the hardware circuit are analyzed and evaluated,and experimental verification is conducted,the hardware circuit designed in this article can meet the basic requirements of epilepsy seizure prediction methods.