Childhood Rolandic Epilepsy Detection Based On Video-EEG

Childhood Rolandic Epilepsy,one of the most harmful types of epilepsy,is the main cause of sudden unexpected death in epilepsy(SUDEP).It is important to detect the seizures of Childhood Rolandic Epilepsy in time,so that the children can receive timely care and treat-ment.However,doctors have no time to mark the long-term video EEG of the child due to limited medical resources.Therefore,the use of machine learning and deep learning tech-nology to detect the seizures of Childhood Rolandic Epilepsy has great application value and development prospects.At present,electroencephalogram(EEG)is regarded as the main-stream method of epilepsy detection,and many algorithms based on EEG extraction features have been developed.However,these methods still have some common problems:1)EEG collection is difficult,and it is easy to cause damage to the head of the child,and also restrict the action of the child;2)Only the EEG data was used to extract features,and the synchro-nized video data of the clinical Video-EEG project was not used rationally;3)The problem of data imbalance between seizure and interictal is not considered.To address these issues,this paper mainly focuses on the detection of seizures based on video and EEG single-modal and multi-modal signals.The main contributions of this paper are as follows:(1)A Childhood Rolandic Epileptic seizures detection algorithm based on multimodal signals,which consists of a fusion feature of Mel-Frequency Cepstral Coefficients(MFCC)and Linear Predictive Cepstral Coefficients(LPCC)to characterize EEG,and Word Frequency Histogram(HWF)to characterize synchronized video,is proposed.This algorithm splits EEG and synchronized video data,extracts features separately,and then performs feature fusion to train machine learning classifiers.It supports single data or multi-modal data input for seizure detection.For EEG,a typical non-stationary signal,the MFCC and LPCC feature are extracted to characterize the characteristics of EEG in the frequency domain;for synchronous video,a bag of words model will be constructed based on the Histogram of Oriented Gradient(HOG)and the Histograms of Oriented Optical Flow(HOF)extracted from the Spatio-temporal in-terest points(STIP)to obtain the HWF feature.The feature vectors of EEG and video are fused and used to train the classifier to obtain a seizure detection model based on Video-EEG multi-modality.At the same time,the SMOTE+Tomek Links method is used to deal with the problem of data imbalance.Finally,the effectiveness of the seizure detection algorithm is verified through comparative experiments.(2)An improved method of HWF video feature extraction based on the STIP screening strategy of YOLOv3 network combined with the combination of multiple feature descriptors is proposed.The trained YOLOv3 neural network is used to predict the patient position bound-ing box(BBox)in each frame,filter the STIP outside the prediction BBox,and reduce the in-terference of other characters or environmental factors in the video.A method by adding Uni-form Local Binary Pattern(ULBP)and Motion Boundary Histogram(MBH)feature descrip-tors,and selecting the best feature descriptor combination HOG+HOF+ULBP+MBH through feature analysis is proposed to further improve STIP’s ability of characterizing videos.Exper-imental results show that the effectiveness of the improved video feature extraction method is greatly improved,making the non-contact Childhood Rolandic Epileptic seizures detection program based on video feature HWF more feasible.(3)A detection system for Childhood Rolandic Epileptic seizures based on EEG and video multi-modal signals is developed.Taking the characterization of the input data EEG and synchronized video as the breakthrough point,the functions of data preprocessing,fea-ture extraction,fusion and classification are completed.The system supports the step-by-step import of video EEG data and graphic display of input signals,as well as single-modal seizure detection of input video or EEG data and multi-modal seizure detection of input EEG and syn-chronized video.