Research On The Algorithm Of Spike Detection In BECT Epileptic Syndrome

Benign epilepsy with centrotemporal spikes(BECT)is a highly prevalent epileptic syndrome in childhood epilepsy.Its primary characteristic is the occurrence of numerous spikes in the central-temporal region of the brain during interseizure periods.The frequency of spikes is correlated with the degree of wakefulness in the patients.The frequency of spikes in nonrapid eye movement sleep stage plays an important role in clinical diagnosis and medication evaluation of BECT patients.Therefore,the design of a fast and accurate spike detection algorithm for BECT syndrome can assist clinical diagnosis and treatment,and it is also the key to the subsequent automated analysis of BECT patients’ conditions.Currently,the mainstream spike detection algorithms are mostly based on preprocessing and feature extraction,followed by advanced classification algorithms for spike detection.However,these algorithms rarely consider the impact of individual differences on detection,and the representation ability of single-channel EEG signals is poor,leading to poor spike detection results on EEG data of BECT syndrome patients.To address the aforementioned issues,this paper conducted the following research efforts:1.A single-channel spike detection method for BECT syndrome based on deep learning and canonical correlation analysis is proposed.This method extracts features using deep transfer learning based on the EEG spectrogram of a Dense Net deep network,as well as traditional hand-crafted features based on a smooth nonlinear energy operator filter.The traditional handcrafted features are selected using a variance filtering algorithm,and then combined with the deep transfer learning features using canonical correlation analysis(CCA)for feature fusion.Finally,the random forests classification algorithm was used to realize spike detection.The performance of the proposed method was evaluated in the EEG data of 15 patients with BECT syndrome from the Children’s Hospital Zhejiang University School of Medicine(CHZU).The experimental results show that the detection performance of the proposed spike detection algorithm can obtain 98.72% sensitivity,84.69% precision and 91.17% F1 score on average,which is overall better than several state-of-the-art spike detection methods.2.A spike detection method for BECT syndrome based on multi-channel weighted fusion and feature fusion is proposed.Multi-channel EEG data has some problems such as high complexity and difficulty in implementation.Therefore,an improved spike detection algorithm based on weighted fusion of multi-channel EEG data was proposed.The multichannel data were selected by feature threshold and multi-channel data fusion to form candidate synthetic single-channel spike samples.On this basis,the detection results of stacked Bidirectional Long Short-Term Memory(Bi-LSTM)used in the original method were analyzed.Four effective hand-crafted features were extracted from the candidate samples,and the hand-crafted features were fused with the depth features of stacked Bi-LSTM using the CCA algorithm.Finally,the features were input into the random forests classifier to realize the multi-channel spike detection of EEG.The proposed method achieved 97.75% sensitivity,98.61% precision and 98.17% F1 score in the CHZU spikes data set,which increased the1.57% sensitivity,2.78% precision and 2.2% F1 score compared with the original method.The average model training time for each patient was only increased by 3.97 seconds.At the same time,the 95.49% sensitivity,95.18% precision and 95.33% F1 score are achieved in the mixed data set.Compared with the original method,the sensitivity,precision and F1 score of the proposed method are increased by 1.44%,0.95% and 1.19%.