Prediction of personalized antiseizure medications response based on clinical signatures in epilepsy

Despite the increasing number of available antiseizure medications (ASMs), optimal medical therapy is still a process of trial and error. We aimed to predict the responsiveness of various ASMs based on initial tests using artificial intelligence. The study consisted of 2586 patients fulfilling the following criteria: (1) first visit to the epileptologists from 2008 to 2017, (2) a diagnosis of epilepsy, and 3) ≥ three years of follow-up duration. The clinical characteristics, ASM history, seizure frequency, laboratory, EEG, and MRI results, were collected. Machine algorithms were utilized to predict the responsiveness of specific regimens. Valproate showed the highest area under curve (AUC), 0.636. The AUCs of levetiracetam, oxcarbazepine, and lamotrigine were 0.614, 0.633, and 0.674. The AUCs of common dual regimens were 0.543 for levetiracetam + oxcarbazepine, 0.454 for levetiracetam + valproate, and 0.583 for levetiracetam + lamotrigine. Levetiracetam + carbamazepine showed the highest AUC, 0.686. In Shapley Additive exPlanations analysis, seizure type significantly impacted prediction performance for valproate responsiveness, and onset age and disease duration for lamotrigine. The prediction performances for the response based on initial data differ according to ASMs. An enormous dataset from a multicenter would improve the prediction power of ASM responsiveness in the future.