Feature Fusion And Collaborative Decision Method For Epilepsy Diagnosis

Epilepsy is the non-communicable chronic neurological disease.The unpredictability of seizures disrupts the patient’s life.Additionally,it can cause various complications,some of which can be life-threatening in severe cases.Diagnosis of epilepsy based on EEG is the most commonly used clinical diagnostic tool for this disorder.However,reading an EEG relies primarily on a visual examination by an experienced physician,which is time-consuming and laborious.Therefore,it is particularly important to study automatic epilepsy diagnostic models for assisting physicians in reading EEG.With the development of technology,many models for epilepsy diagnosis have emerged,although there are still some challenges.First,raw EEG data suffer from incomplete information,and the fact that spatio-temporal information can interfere with each other.The second is the problem of insufficient marker data and sample imbalance for seizures.The third is how the explored extracts the valid features of comprehensiveness.The fourth is the challenge of heterogeneity of data and diversity of classification settings.To address this,two new methods based on feature fusion and collaborative decision making are proposed in this paper.They are used for seizure detection and seizure subtype classification tasks,respectively.A proposal for a prototype online service platform was further developed subsequently.The main three works included in this paper are as follows:(1)To address the issue of incomplete information of the original data and the discriminatory ability of features.We propose a multi-scale deep feature fusion and multi-view collaborative epilepsy detection algorithm in this paper.First,three shallow view features are extracted and deep features are obtained by deep learning.Then,a multi-scale fusion network is constructed to learn multiple global fusion view features,including a generalized global fusion view and three locally enhanced global fusion view.Further,a multi-view TSK fuzzy system is trained on the basis of the obtained three deep views and four global fusion views.The final epilepsy detection model with high generalizability and interpretability was constructed.The model was applied to the international open-source Boston Children’s Hospital epilepsy EEG dataset for performance evaluation,and it performs well in all three test performances of accuracy,sensitivity and specificity.Compared with other novel methods,the detection algorithm proposed in this paper has better overall performance.(2)To address the issues of the feature validity,the heterogeneity of data,and the diversity of classification settings.We propose a subtype classification algorithm with autocorrelation feature fusion and fuzzy classification.The algorithm consists of three parts: deep feature processing,autocorrelation fusion and fuzzy classification.Deep feature processing includes a series of operations such as channel compression,spread spectrum,etc.Regarding the autocorrelation fusion process,it first relies on multiheaded attention to extract autocorrelation information from both the time and frequency domains.Next,two autocorrelation features are combined to create the timefrequency autocorrelation feature.Finally,the multi-view information of the timefrequency view is learned to obtain the time-frequency fusion feature.Fuzzy classification is performed using the classical TSK fuzzy system as the foundational model.We conducted experiments using the Temple University epilepsy EEG dataset and the Huazhong University of Science and Technology children’s dataset.The results showed good performance in terms of both F1-Score and accuracy.(3)Based on the above two algorithms,we developed and designed a prototype system of an intelligent epilepsy diagnosis platform.The front-end of the platform is implemented by Vue and the back-end is based on Spring Boot and Flask.This platform is designed to serve both clinicians and patients and is supervised by administrators for information and permissions.The physician first confirms whether the seizure is a seizure and marks the seizure period with the seizure detection model,and then uses the automatic seizure subtype detection model to specify the patient’s seizure category.This allows physicians to more efficiently have sufficient pathological information and make symptomatic diagnoses.Patients can check their condition remotely with the help of the platform.At the same time,they can also communicate with doctors conveniently and timely through the online chat function to achieve the purpose of facilitating remote consultations between doctors and patients.The platform not only improves the diagnosis efficiency of doctors,but also reduces the consultation cost of patients.It enhances doctor-patient communication and information sharing.