KDD 2022 | epileptic wave prediction based on hierarchical graph diffusion learning

subject ：BrainNet： Using hierarchical graph diffusion learning in SEEG Epileptic waves are predicted in the data （BrainNet: Epileptic Wave Detection from SEEG with Hierarchical Graph Diffusion Learning）

 

author ： Chenjunru （ Zhejiang University ）, Yang Yang （ Zhejiang University ）, Yu Tao （ Zhejiang University ）, Fanyingying （ Zhejiang University ）, Moxiaolong （ Noel Medical ）, Yang Ji （ Emory University ）

 

meeting ：28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

 

Epilepsy is one of the most common serious neurological diseases , It is characterized by abnormal neurophysiological activities , Characterized by seizures or behavioral abnormalities . By the end of 2021 year , There are more than 6,500 Ten thousand people with epilepsy , About one-third of epilepsy patients are resistant to drugs . let me put it another way , Drugs cannot effectively control the condition of these patients , Surgical removal of brain regions involved in seizures is considered the only effective treatment . To assess the location of the seizure area (SOZ) Or so-called epileptic foci , And guide epilepsy surgery , It is necessary to record the discharge activity of the patient's brain . There are usually two types of electrophysiological monitoring methods ： Cortical EEG （EEG） And stereoscopic EEG （SEEG）. The former is non-invasive , The latter is intrusive （ That is, electrodes need to be inserted into the brain ）, Therefore, it also contains more deep three-dimensional information . for example , When SOZ Located in the deep structure of the brain （ For example, hippocampus or insular body ） in , Or when the laterality of seizures is unknown , Noninvasive detection will not be able to locate the exact seizure focus , here ,SEEG Methods are necessary .

 

In order to promote the development of epilepsy treatment , In this paper, the research team from the third class hospital , A collection of data based on real patient records SEEG Data sets , The data set contains high-frequency multi-channel data of multiple epileptic patients SEEG The signal （ On average, each patient has a continuous 53 Hours 、 The size is 77GB The record of ）. Based on this dataset , The task of automatic prediction of epileptic waves is further proposed and studied in this paper . This paper aims to 1 As an example , Further introduce the details of this paper . chart 1（a） The top of the shows a human brain , An electrode with three channels is inserted in the upper left corner of the human brain . After inserting electrode , The doctor can collect and monitor the patient's SEEG data ; Pictured 1（a） At the bottom of the , Actually SEEG Data can be considered as continuous multi-channel time series . get SEEG After the data , The goal of this paper is to construct an automatic data-driven method to determine the time and location of epileptic waves （ In the figure 1（a） Marked with a yellow rectangle ）.