Seizure risk prediction using machine learning following glioma resection surgery in seizure-naïve patients

Despite the ongoing controversy around the prophylactic use of antiseizure medications (ASMs) in seizure-naïve patients undergoing brain tumor surgery, this practice has persisted for years. This study aims to develop and validate a machine-learning framework for stratifying postoperative seizure risk.

Demographic, tumor topographic, surgery-related details, and biomarkers were collected from a retrospective study involving patients undergoing glioma resection. The dataset was split in a stratified manner into an 80/20 ratio for training and testing purposes. Machine learning (ML) models, including random forest (RF), XGBoost, gradient boosting decision tree (GBDT), multi-layer perceptron (MLP), bootstrap-aggregation ensemble classifier with decision tree classifier (Bagging), and logistic regression (LR), were developed and evaluated. The SHAP method was applied to interpret the attribution values of the top features.

Among the 786 eligible patients, with a median age of 42.0 years (interquartile range [IQR] = 25.3-54.0), 154 (19.6%) experienced postoperative seizures. The multi-layer perceptron model demonstrated the best predictive performance, incorporating demographic, topographic, surgery-related, and biomarker variables (Test: AUC: 0.74, Accuracy: 0.70, Sensitivity: 0.56, Specificity: 0.73). Notably, a simpler model relying solely on demographic and topographic features also yielded comparable performance.

This study underscores the effectiveness of the multi-layer perceptron model, which integrates demographic, topographic, surgery-related, and biomarker variables. This integration successfully developed a personalized prediction model for postoperative seizure risk. Such a model holds the potential to aid physicians in optimizing postoperative management, particularly in guiding decisions regarding the duration and discontinuation of prophylactic antiseizure medications.