Abstracts

Rationale: Epilepsy affects approximately 50 million people worldwide and is characterized by recurrent, unpredictable seizures with diverse etiologies. Antiseizure medications (ASMs) remain the mainstay of therapy, and eslicarbazepine acetate (ESL) is widely used for partial-onset seizures. However, treatment responses vary, and reliable prediction tools are lacking. This study aimed to identify clinical predictors of ESL response and develop supervised machine learning (ML) models to support individualized treatment decisions.

Methods: We retrospectively analyzed 165 patients with epilepsy treated with ESL at Ajou University Hospital (Jan 2022–Dec 2024), followed for ≥12 months. Collected variables included ESL dose, seizure frequency, seizure type, etiology, concurrent ASM use, adverse reactions, adherence, age, and sex. Treatment response was defined as ≥50% seizure reduction at 12 months, classifying patients as good responders (GR) or poor responders (PR). Five ML algorithms—Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Random Forest, and Logistic Regression—were trained on baseline data (80:20 split). Performance was assessed via accuracy, sensitivity, specificity, and F1 score.

Results: Among 165 patients (84 males, 67 females; mean age 48.6 ± 17.0 years), 102 (61.8%) achieved a good response. GR patients had higher mean ESL doses (1,083 ± 275 mg vs. 943 ± 258 mg; p = 0.004) and lower baseline seizure frequency (3.1 ± 4.5 vs. 8.7 ± 10.2/month; p < 0.001) than PR. Focal to bilateral tonic-clonic seizures were more frequent in GR (38.2% vs. 19.7%; p = 0.018), whereas perinatal and infectious etiologies were more common in PR (14.8% vs. 3.9%; p = 0.021). XGBoost and SVM showed the best predictive performance (accuracy 0.81, sensitivity 0.91, specificity 0.50, F1 0.84). KNN reached sensitivity 1.00 but specificity 0.18. SHAP analysis identified baseline seizure frequency, ESL dose, and seizure type as the top predictors.