Machine Learning Analysis of Routine EEG Accurately Predicts Anti-Seizure Medication Response

Despite the availability of more than 20 anti-seizure medications (ASMs), approximately half of patients with newly diagnosed epilepsy fail their first drug trial. Unfortunately, clinicians lack objective tools or consensus guidelines to match individual patients with the most effective therapy, frequently leading to years of uncontrolled seizures. Here, we developed machine learning models to utilize a single, baseline routine resting-state scalp EEG to forecast ASM efficacy. EEGs and treatment outcomes were drawn from 280 participants with new-onset focal epilepsy in the prospective, multicenter Human Epilepsy Project. Recordings were acquired within four months either before ASM initiation (unmedicated EEG) or after treatment onset (medicated EEG). For each recording we computed band-limited static and dynamic functional-connectivity and entropy-based matrices in consecutive time windows. We trained and tested classifiers in a nested fashion to predict future seizure freedom. Separate classifiers were trained to (i) predict levetiracetam response from unmedicated EEGs (22 responders, 32 non-responders) and from EEGs recorded on an ASM (53 responders, 31 non-responders); (ii) predict lamotrigine response from unmedicated EEGs (12 responders, 21 non-responders); and (iii) distinguish participants who ultimately proved refractory to all ASMs from unmedicated EEG (67 responders, 16 refractory) and EEGs recorded while on an ASM (34 responders, 80 refractory). Two model architectures were tested for each classifier. Performance, evaluated with nested leave-one-out cross-validation, was robust across at least one model architecture for each classifier: area under the ROC curve (AUC) 0.88 and balanced accuracy 0.85 for unmedicated levetiracetam, AUC 0.82 and balanced accuracy 0.77 for medicated levetiracetam, AUC 0.79 and balanced accuracy 0.80 for unmedicated lamotrigine, AUC 0.92 with balanced accuracy 0.87 for unmedicated refractory, and AUC 0.82 with balanced accuracy 0.75 for the medicated refractory model. These findings indicate that routine EEG harbors machine-learning-detectable signatures predictive of specific ASM efficacy, laying groundwork for precision-medicine tools that could shorten the costly trial-and-error period in epilepsy treatment.