Predicting seizure-free outcomes in people with treatment-resistant epilepsy: A machine learning approach

Epilepsy surgery is an important intervention for treatment-resistant epilepsy, butthe ability to predict long-term seizure freedom post-surgery has yet to be achieved. Machine learning (ML) models could improve outcome prediction by analyzing the effects of resections of brain regionswith respect to International League Against Epilepsy (ILAE) scores of post-surgical seizure freedom. This study developed and validated ML models to predict seizure-free years based on brain resections.

We analyzed 443drug-resistant epilepsy (DRE) patients who underwent surgery. Intersected volumetric data from resected neurological structures with post-surgical seizure outcomes was utilized. We developed and tested supervised ML models using the resected brain volumes and 5-year outcomes. First, regression models (Random Forest, Gradient Boosting, XGBoost) treated the 5-year ILAE score (1-5) as an ordinal outcome to explore whether seizure-free duration could be approximated from resection patterns. Second, classification models (logistic regression, support vector machines, and an ensemble classifier) predicted 5-year seizure freedom (ILAE 1-2 vs ILAE 3-5). Model performance was evaluated using Mean Squared Error (MSE) and R² for regression, and accuracy, sensitivity, specificity, and ROC-AUC for classification. SHAP and LIME interpretability techniques were applied toanalyze feature importance.

Random Forest yielded themost consistent predictions (MSE = 1.82, R² = -0.57), but predictive performance waned over time. Theclassification model was able to identify seizure-free patients, distinguishing them from non-seizure-free patients with an accuracy of 58.5%.

The crucial regions dictating seizure outcomes included the fusiform gyrus, superiortemporal cortex, and insula. In future research, such predictive models should incorporate multi-modal biomarkersand external validation to improve prediction accuracy.