SVM-based Prediction Model For Seizure-free On Epileptic Patients With Initial Therapy Of Levetiracetam

Objective: Anti-epileptic drugs(AEDs)are the basic treatment of epilepsy.However,30% to 40% of patients with epilepsy(PWEs)are still suffering from drug refractory epilepsy.Previous studies have proved that the failure of the first antiepileptic drug is one of the risk factors of refractory epilepsy;however,the efficacy even has no significant difference among selected AEDs according to the principle,which makes it difficult to choose a suitable first antiepileptic drug accurately.With the development of machine learning technology,some studies have predicted the prognosis of epilepsy surgery by establishing a support vector machine(SVM)model.However,there are still no studies to predict the efficacy of drugs and guide drug selection.This study collected the clinical and quantitative electroencephalogram(QEEG)characteristics of PWEs before levetiracetam,and established a SVM model to predict the probability of seizure-free(SF),clinical and EEG factors that affect the outcome were also identified,which may guide the AED selection for newly diagnosed patient.Methods: The proposed algorithm consists of features extraction,SVM classification using 5-fold cross-validation,model evaluation.46 newly diagnosed PWEs who once visited the Epilepsy Clinic from Henan Province People's Hospital from 2014 to 2016 and had an initial treatment with levetiracetam were enrolled.These patients were divided into SF group(22 patients)and not seizure-free(NSF)group(24 patients)according to the therapeutic effect after 1 years of follow-up.Eleven clinical factors consisted of MRI findings,family history,seizure circadian rhythm,LEV initiation to the last seizure,comorbidity,duration of epilepsy,temporal lobe epilepsy,age,seizure type,interictal spike,seizure frequency before LEV together with four quantitative EEG features consisted with sample entropy of ?,?,?,? wave before taking levetiracetam were compared and used to build the model.We selected 17 patients from SF group and 19 patients from NSF group randomly as the training set to establish a SVM model with 5-fold cross-validation,which was further used to predict the therapeutic efficacy of the remaining 5 patients from SF group and 5 patients from NSF group(test set).Finally,we used the parameter of accuracy,specificity,sensitivity and AUC to evaluate the model after comparing with the real effect.The algorithm of mean impact value(MIV)was used to rank the relativity between each factor and the outcome.Results: 46 PWEs were enrolled,22 were included in SF group and 24 belonged to the NSF group.Compared to the SF patients,the NSF patients displayed a specific decrease on EEG sample entropy in ? band from F4 channel(p<0.001),? band from Fp2 channel(p=0.008)and F8 channel(p=0.011),? band from C3 channel(p<0.001),but no significant difference in clinical features was found.The SVM model was built and the accuracy of cross-validation is 72.22%.In test set,the prediction accuracy was 90%,the specificity was 90%,the sensitivity was 100%,the positive predictive value was 83.33%,the negative predictive value was 100%,AUC value was 0.96.MIV showed that ? band from Fp2 channel,? band from F4 channel,? band from C3 channel,MRI findings,family history,seizure circadian rhythm,? band from F8 channel,time between initial LEV to the last seizure,comorbidity,duration of epilepsy,temporal lobe epilepsy have an effect on the outcome.While the age,seizure type,interictal spike,and seizure frequency were ineffective.Conclusions: Newly diagnosed PWEs who would achieve SF and NSF after levetiracetam have a difference of quantitative EEG.Moreover,the efficacy of LEV could be predicted by a SVM model in advance,which may guide the LEV selection.