BACKGROUND

Predicting long-term outcomes in newly diagnosed epilepsy remains limited by reliance on clinical features and visual EEG interpretation. Machine learning enhances this potential by identifying complex patterns in EEG data, as demonstrated in studies on predicting surgical outcomes and seizure initiation. However, its application to predicting ASM response in newly diagnosed epilepsy has been underexplored. This study aimed to develop a machine learning model to predict ASM response in newly diagnosed epilepsy patients, with the goal of improving personalized treatment strategies and early identification of drug resistance.

METHODS

This retrospective cohort study included adult patients with a new epilepsy diagnosis who underwent EEG prior to ASM initiation. Patients with structural brain lesions on MRI were excluded. Seizure control was assessed two years after starting ASM treatment, with responders defined as those achieving at least one year of seizure freedom during the second year. We applied three prediction approaches: one using only clinical variables, another using only EEG features, and a third integrating both clinical and EEG data. For each approach, Logistic Regression, Extreme Gradient Boosting (XGBoost) and Random Forest models were implemented to assess predictive performance and suitability for ASM response prediction. Model performance was evaluated at both epoch and patient levels, with patient-level predictions generated by averaging class probabilities across epochs.

RESULTS

This study included 94 patients with newly diagnosed epilepsy who received EEG before starting ASM treatment, of whom 77 (81.9%) achieved seizure freedom. Machine learning models using clinical features showed moderate predictive performance, with the XGBoost model achieving the highest AUROC of 0.69. For EEG features, patient-level predictions improved model performance, with the Random Forest model achieving an AUROC of 0.68. The combined clinical-EEG model significantly enhanced accuracy, with Random Forest model achieving the best performance (AUROC: 0.81). Among EEG features, power spectral density (PSD) in the beta and gamma bands, along with sample entropy, were identified as the most predictive of treatment response.

CONCLUSION

Quantitative EEG analysis using machine learning shows significant potential in predicting the long-term prognosis of newly diagnosed epilepsy, even in patients without structural brain lesions or visually abnormal background EEGs. By integrating clinical variables with quantitative EEG features, these machine learning models demonstrate potential to support individualized treatment planning and the early identification of drug resistance. However, further validation in larger and diverse populations is needed before clinical implementation.