Automatic Detection Of High-Frequency Oscillations Signal Based On Deep Learning And Research On Lesion Localization Of Hippocampal Sclerotic Epilepsy

Epilepsy is a chronic neurological disease.About 9 million people are suffering from epilepsy in China.Long-term and repeated seizures led to huge physical and mental damage to patients,and also brought a huge burden to their families and social medical care.About 20%of all epilepsy patients have drug-resistant epilepsy.For these patients,only invasive methods such as surgical resection or thermal coagulation can be used for clinical treatment.Therefore,accurate localization of the epileptogenic zone is the key to effective preoperative evaluation.Many studies have shown that intracranial high frequency oscillations(HFOs)signals are closely related to epileptic seizures,and thus can be used as a new biomarker to guide the localization of epileptogenic zone.Although expert visual inspection remains the gold standard for HFOs analysis,manual analysis of HFOs is time-consuming and highly subjective.In addition,the existing HFOs automatic detection algorithms have limitations such as single analysis perspective and insufficient signal characterization capabilities;the existing epileptogenic zone localization schemes also have shortcomings such as insufficient signal feature mining for each channel.Therefore,in order to solve these problems,this paper proposes a multianalysis perspective method based on a deep feature fusion model for the common pathological type of drug-resistant epilepsy-Mesial Temporal Lobe Epilepsy-Hippocampal Sclerosis(MTLE-HS),which realizes the complementary advantages of different modal data for signal pattern characterization,and on this basis,a MTLE-HS localization method based on a multi-feature fusion model has been proposed,which realizes the feature mining of all frequency band for each channel.In the study,the clinical data were used to verify the effectiveness of the proposed scheme,where the value and application potential of the research work were confirmed.The main research work is summarized below:First,preprocessing and initial detection were performed on the raw SEEG signals of 5 MTLE-HS patients,then the HFOs candidate event set can be screened out.Then,combined with time-frequency transformation,experts were invited to complete the visual analysis,where a private highfrequency oscillation dataset was formed,with a total of 15,000 standard HFOs records.Then,an automatic detection algorithm for HFOs based on an Endto-end Bi-branch Fusion Neural Network(EBFNet)is designed.Construct a hybrid network based on ResNetlD and long short-term memory(LSTM)to realize the mining of spatio-temporal features of bandpass signals;then construct a ResNet2D convolution neural network based on Spatial Attention(SA)and Channel Attention(CA)to realize the effective mining of key information of time-frequency images;construct a deep feature fusion module to realize the automatic classification of HFOs.In the study,several experiments have been designed to verify the effectiveness of the proposed method.In intra-patient validation and cross-patient validation,the average sensitivity was 94.62%and 92.00%,and the average specificity was 92.70%and 88.26%,respectively.The effectiveness of the two branch fusion strategy and the optimization strategy for each single branch are confirmed by ablation experiments and many other experiments.In addition,the agreement between the EBFNet and the clinical gold standard is discussed by calculating the kappa coefficient.Finally,the localization method of MTLE-HS was studied.The EZ localization method based on the distribution of HFOs in the channel and the characteristic deviation method was applied,and the clinical cases were analyzed,which confirmed the reliability of HFO as a biomarker.However,it is limited by the performance bottleneck of the automatic detector,so the sensitivity is low.Therefore,to solve this problem,a multi-feature fusion model based on the extreme gradient boosting algorithm(Multi-Feature Fusion-XGBoost,MF-XGB)is designed.The original signal timefrequency domain features,δ,θ,α,β,γ and other frequency band power features and HFO-related features are combined into a new feature space,and the XGB decision model is trained to realize the discrimination of channel epileptogenicity.In the study,several experiments have been designed to verify the effectiveness of the proposed method.In intrapatient validation and cross-patient validation,the average sensitivity was 89.47%and 68.18%,and the average specificity was 94.83%and 94.47%,respectively.In addition,several categories of features of high importance for the determination of epileptogenicity are discussed for the combined features.Finally,a new epileptogenicity index(EI)is defined,and the epileptogenicity of each case was analyzed and discussed,which further confirmed the significance and practical value of the work in this paper.