Development and validation of a multimodal data-based scale for predicting epilepsy surgery outcomes

Seizure outcomes after drug-resistant epilepsy surgery are hard to predict, and existing models lack rigorous validation. We developed and internally validated a multimodal scale predicting seizure freedom.

Retrospectively, 157 of 816 screened patients with complete data and ≥ 1 year follow-up were analyzed (116 seizure-free [SF], 41 non-seizure-free [NSF]). Predictors came from clinical variables, MRI, FDG-PET, video-EEG, and MEG. From 56 variables, eight predictors were fixed by screening, clinical judgment, four machine-learning algorithms, and minimum-AIC logistic regression, then re-estimated in all 157 patients by Firth's bias-corrected logistic regression without repeat selection. Internal validation used 2000 bootstrap resamples. A cutoff maximizing 0.4 ×sensitivity + 0.6 ×specificity was selected once in the full cohort and applied unchanged to the ≥ 24 months subgroup.

The model retained disease duration, age at surgery, four EEG patterns, temporal MRI lesion, and unilateral unifocal PET hypometabolism. Development-phase AUCs (0.864 training, 0.810 test) were descriptive. Firth re-estimation gave an apparent AUC of 0.771 and an optimism-corrected AUC of 0.724 (Brier 0.156/0.176). At the 79.64-point cutoff, sensitivity was 61.2% and specificity 85.4%; all 95% confidence intervals included the null. Among 101 patients followed ≥ 24 months, the locked model gave AUC 0.770 (95% CI 0.664-0.865), Brier 0.158, sensitivity 53.9%, specificity 96.0%.

Multimodal data integration yielded a fully specified, reproducible eight-variable scale with stable discrimination in longer-term follow-up, supporting exploratory use for preoperative risk stratification and counseling alongside presurgical evaluation. Prospective multicenter validation is the next step toward clinical implementation.