Research On Seizure Detection Method Based On Abnormal Waves Of Scalp Eeg Signals

Epilepsy is one of the nervous system disease caused by abnormal discharges of brain neurons,which has the characteristics of paroxysmal and repetitive.The scalp electroencephalogram(s EEG)in epileptic seizure state exhibits significant abnormalities,mainly manifested as specific waveforms such as spikes,sharp waves,spike slow waves,sharp slow waves,or rhythmic bursts.At present,the s EEG recognition and analysis work completed by clinical doctors and EEG technicians based on experience and visual detection still has various limitations,such as high subjectivity and high time consumption.Therefore,it is extremely important to use computer-aided technology to detect seizures.Epileptic spikes in sEEG are important and widely accepted biomarkers for epilepsy diagnosis,and their waveforms are similar to non-epileptic brain activity or artifacts,making them difficult to detect.Therefore,this thesis focuses on seizure detection and conducts research from two aspects: epileptic spikes detection and abnormal wave feature fusion.The main research content of this thesis is as follows:(1)To address the limitations of manual visual detection of s EEG,an EEG epileptic spike detection method based on sliding windows and traditional machine learning is proposed.The method takes into account the temporal continuity and information integrity of epileptic spikes,uses a sliding window with overlap rate to segment the s EEG signal,and validates the effectiveness of the method on the dataset using three models: K-nearest neighbor,support vector machine and random forest,which have low computational overhead and strong interpretability.The experimental results show that the method proposed in this thesis can achieve automatic detection of s EEG epileptic spikes and achieve the purpose of distinguishing epileptic spikes from non-epileptic brain activity or artifacts.(2)According to the spike waveform characteristics and individual differences among patients,an EEG epileptic spike detection method based on adaptive template matching and onedimensional convolutional neural networks is proposed.This thesis uses K-means clustering to optimise the template shape based on generic template matching,thereby producing more accurate matching results in the specific template matching stage.Due to the phenomenon of a segment of data being reused multiple times during sliding window segmentation in research content(1),in order to avoid this problem and ensure the integrity of the spike waveform,the s EEG signals are segmented based on the location of the candidate spike peaks based on template matching,and the segmented segments are fed into a 1D convolutional neural network as input.The deep learning model proposed in this thesis has the advantages of simple structure and a few trainable parameters.The experimental results show that the proposed method improves the performance of epileptic spike detection.(3)To address the problems of limited existing data samples,insufficient feature richness and insufficient validity,a seizure detection method based on feature fusion and Transformer encoder is proposed.To enrich the sample features,this thesis adds spike correlation features and spike features as the complement to the seizure features on top of the traditional time domain and frequency domain features extracted from s EEG.In this regard,the waveforms of abnormal s EEG discharges are carved using smoothed non-linear energy highlighting,combined with adaptive thresholding to delineate spikes and calculate spike correlation features,and spike wave features are constructed for the epilepsy dataset using the method in research content(2).To improve the impact of different channels of s EEG on the detection results,the obtained fused features are fed into a Transformer network with a learnable feature vector,and the channel embedding method enables the Transformer to automatically assign weights to each channel,thus optimising the model performance.The experimental results show that the method can improve seizure detection performance compared to the baseline method.