Construction Of Prediction Model And Extraction Of Risk Factors For Refractory Epilepsy Based On Machine Learning Technology

Objective: To understand the basic demographic characteristics,clinical features,medication use,imaging and video electroencephalogram(EEG)of epilepsy patients in Northeast China,a more mature support vector machine classifier in machine learning was used to construct a refractory epilepsy(RE)prediction model and extract potential risk factors associated with RE.Methods: Patients who attended the Epilepsy Center of the First Hospital of Jilin University and diagnosed with epilepsy from January 2016 to December 2018 were included and followed up for at least 2 years,with the final follow-up in December 2020.Their basic demographic characteristics,clinical characteristics,medication use,imaging and EEG data were fully collected,and the statistically different features were assigned and preprocessed after statistical analysis of the data,and a support vector machine classifier was used to construct a RE prediction model based on the KW algorithm with a ten-fold cross-test method,and a binary logistic analysis was performed on the features under the best accuracy prediction model,to extract risk factors for the eventual development of RE in patients with epilepsy.The study subjects were divided into RE group and non-RE group according to the follow-up results.Results:(1)A total of 366 patients were collected,146 patients who eventually developed RE and 220 patients with non-RE.76 variables were included in the univariate analysis,and variables with statistically significant differences between the RE and non-RE groups included: age at onset(Z=-5.747,P<0.01),interval between first and second onset(χ~2=44.051,P<0.01),history of febrile convulsions(χ~2=5.732,P=0.017),history of abnormal birth(χ~2=4.665,P=0.031),history of abnormal growth and development(χ~2=4.608,P=0.032),children(χ~2=5.746,P=0.017),work status(χ~2=84.474,P<0.01),clustered seizures(χ~2=24.699,P<0.01),and status epilepticus(χ~2=14.329,P<0.01),aura symptoms(χ~2=4.87,P=0.027),hazy state after seizure(χ~2=9.938,P=0.002),drowsiness after seizure(χ~2=13.1,P<0.01),asymptomatic after seizure(χ~2=9.892,P=0.002),treatment effect after first visit(Z=7.24,P<0.01),structural(etiology)(χ~2=4.467,P=0.035),etiology unknown(χ~2=4.475,P=0.034),temporal lobe epilepsy(χ~2=4.2,P=0.04),bilateral discharges(χ~2=4.767,P=0.029),number of discharges in the central region(Z=-3.509,P< 0.01),hippocampal sclerosis(χ~2=6.574,P=0.01).(2)The prediction model was successfully constructed with a maximum model accuracy of 78.99% and model sensitivity and specificity of 73.09 % and 82.74%,respectively.(3)The combination of features under the best prediction model accuracy included: first and second onset interval,medication effect after the first visit,age at onset,clustered seizures,bilateral discharges,hazy state after seizures,temporal lobe epilepsy,status epilepticus,work status,and hippocampal sclerosis,and binary logistic regression revealed that the risk factors for RE included:(i)interval between first and second onset(OR=0.205,95% CI: 0.121-0.349);(ii)age at onset(OR=0.964,95% CI: 0.949-0.98);(iii)clustered seizures(OR=2.282,95% CI: 1.251-4.163);(iv)status epilepticus(OR= 2.341,95% CI: 1.175-4.663);effect of first treatment medication(seizure reduction <50%: OR=13.252,95% CI:4.807-36.538;seizure reduction 50%-75%: OR=4.811,CI:1.737-13.325);temporal lobe epilepsy(OR= 1.968,95% CI: 1.146-3.379).Conclusions:(1)A support vector machine classifier based on the KW algorithm was used to analyze 366 epilepsy patients who first visited our epilepsy center,and a prediction model for refractory epilepsy was successfully constructed with an accuracy of 78.99%,sensitivity and specificity of 73.09 % and 82.74%,respectively,with superior classification performance.(2)The risk factors for RE,including the interval between first and second onset,age at onset,clustered seizures,status epilepticus,effectiveness of first treatment,and temporal lobe epilepsy,were extracted by machine learning methods,providing a theoretical basis for future clinicians to screen high-risk patients and make the best treatment plan.