Predicting the therapeutic response to valproic acid in childhood absence epilepsy through electroencephalogram analysis using machine learning.

Childhood absence epilepsy (CAE) is a common type of idiopathic generalized epilepsy, manifesting as daily multiple absence seizures. Although seizures in most patients can be adequately controlled with first-line antiseizure medication (ASM), approximately 25 % of patients respond poorly to first-line ASM. In addition, an accurate method for predicting first-line medication responsiveness is lacking. We used the quantitative electroencephalogram (QEEG) features of patients with CAE along with machine learning to predict the therapeutic effects of valproic acid in this population. We enrolled 25 patients with CAE from multiple medical centers. Twelve patients who required additional medication for seizure control or who were shifted to another ASM and 13 patients who achieved seizure freedom with valproic acid within 6 months served as the nonresponder and responder groups. Using machine learning, we analyzed the interictal background EEG data without epileptiform discharge before ASM. The following features were analyzed: EEG frequency bands, Hjorth parameters, detrended fluctuation analysis, Higuchi fractal dimension, Lempel-Ziv complexity (LZC), Petrosian fractal dimension, and sample entropy (SE). We applied leave-one-out cross-validation with support vector machine, K-nearest neighbor (KNN), random forest, decision tree, Ada boost, and extreme gradient boosting, and we tested the performance of these models. The responders had significantly higher alpha band power and lower delta band power than the nonresponders. The Hjorth mobility, LZC, and SE values in the temporal, parietal, and occipital lobes were higher in the responders than in the nonresponders. Hjorth complexity was higher in the nonresponders than in the responders in almost all the brain regions, except for the leads FP1 and FP2. Using KNN classification with theta band power in the temporal lobe yielded optimal performance, with sensitivity of 92.31 %, specificity of 76.92 %, accuracy of 84.62 %, and area under the curve of 88.46 %.We used various EEG features along with machine learning to accurately predict whether patients with CAE would respond to valproic acid. Our method could provide valuable assistance for pediatric neurologists in selecting suitable ASM.