Validation of a machine learning algorithm for the prediction of first anti-seizure medication response in focal epilepsy: a multicenter cross-sectional study

INTRODUCTION For people with epilepsy, the EEG is a pivotal neurophysiological technique for confirming the diagnosis and guiding clinical management. However, no definite EEG prognostic biomarkers for anti-seizure medications' (ASMs) efficacy have been defined. AIMS The purpose of this study is to use a machine-learning (ML) approach to determine the predictive power of conventional scalp EEG for seizure freedom in a population of newly diagnosed focal epilepsy patients after a first ASM initiation. We hypothesize that quantitative EEG features can predict clinical outcome (seizure freedom) before ASM initiation. METHODS We examined 203 patients with newly-diagnosed focal epilepsy from six epilepsy centers across Italy between March 2018 and July 2022. We dichotomized clinical outcomes into seizure-free (SF) and non-seizure-free (NSF) after one year of ASM initiation. We built a cross-validated data-driven model based on the ML model: Partial Least Squares regression. We then performed a K-fold cross-validation procedure to assess the generalization of our model in predicting clinical outcome. RESULTS A total of 76 features were extracted from the conventional 19-channel EEG recordings. The most common ASMs employed in our cohort were Levetiracetam (LEV) (61%) and Lamotrigine (LTG) (11.8%). The ML model was able to predict seizure freedom with an area under the curve (AUC) of 0.62. The ML model used on patients treated with LEV achieved a prognostic prediction with an AUC of 0.77. CONCLUSION This study provides an ML algorithm for predicting the clinical response to ASMs in people with epilepsy. Future studies may benefit from the pipeline proposed in this study for the development of a clinical decision-making tool for data-driven, individualized ASM choices for people with epilepsy.