Research On Detection Algorithm For Epileptic Interictal Spikes Based On Deep Learning

In the field of the analysis about EEG signals,the methods based on manual detection have been unable to meet the practical needs increasingly.With the vigorous development of machine-learning and deep-learning algorithms,the combination of artificial intelligence and EEG signal analysis is becoming deeper and deeper.In the aspect of epilepsy,spikes are a kind of epileptiform patterns.Therefore,accurate interictal（non-seizure）spike detection is of great significance for the diagnosis and prediction of epileptic seizures.However,some existing predictive algorithms for spike detection have a few limitations.Therefore,this paper propose two kind of methods for epileptic interictal spike detection based on deep learning.Relevant specific studies are as follows:Firstly,this paper proposes a spike detection method based on the Time-Frequency and Waveform-Change features fusion deep networks（TFWC networks）.In order to achieve a better detection effect,this method combines the discrimination process used by the experts in artificial recognition of spike.Short Time Fourier Transform（STFT）,a kind of time-frequency feature is extracted to describe the changes of EEG waveforms in time and frequency domain macroscopically.Data-point Difference（DD）is extracted to describe the difference between the peak position and the flat position.Finally,the two features are fused.The method is tested on the spike data set provided by Xuanwu Hospital.And the experimental results show that the accuracy of TFWC networks in the spike detection task reaches 91.29%.Secondly,according to TFWC networks,this paper proposes a spike detection method based on Time-Frequency Residual and More-order-difference Waveform-Change features fusion deep networks（TFRMWC networks）.In the analysis of time-frequency features,because of the excellent performance of Residual Networks in processing grid data,the convolution blocks are replaced by residual blocks.In the analysis of waveform-change features,second-order difference and third order difference are added as the supplement to the waveform-change features.Finally,the two features are fused.The method is tested on the spike data set provided by Xuanwu Hospital.And the experimental results show that TFRMWC networks have better effect in the spike detection task than TFWC networks,and the accuracy reaches 91.78%.Finally,the spike recognition system is developed by integrating each step of spike recognition method.The system has three functions: feature extraction and visualization,model training and waveform recognition.In the part of feature extraction and visualization,feature categories include STFT and DD.In the part of model training,the recognition algorithm is TFWC Networks.In the part of waveform recognition,spike can be recognized in the EEG signal segments that meet the requirements of sampling frequency and time length.