Five predictors of meningioma surgery: an AI analysis.
Abstract
Background
Meningioma surgery aims to preserve neurological function, usually measured by the Karnofsky Performance Scale (KPS), but predicting postoperative KPS change remains challenging.
Methods
We retrospectively analyzed 397 meningioma resections from 14 Italian centers. Preoperative clinical and MRI features (age, tumor size, eloquent-area location, vessel and cranial-nerve involvement, prior surgery, baseline KPS, etc.) were used to train classifiers (logistic GAM, weighted logistic regression, linear/RBF SVM, random forest) with cross-validation. We compared performance to the Staartjes et al. GAM_MCS model (predicting KPS drop ≥10).
Results
All models achieved only moderate discrimination for new deficits: AUC ≈0.65-0.70 (Staartjes2 GAM_MCS 0.692; logistic regression 0.694). Predicting major changes was more accurate: for KPS drops ≥20 or ≥30 points, AUC ≈0.76-0.77, and for any postoperative improvement AUC ≈0.81, whereas 10-point declines were hardest (AUC ≈0.70). For example, weighted logistic regression yielded AUCs of 0.697, 0.765, 0.762, and 0.811 for predicting KPS drops ≥10, ≥20, ≥30, and any improvement, respectively. The most important predictors were anatomical complexity factors: eloquent-area involvement, major vessel manipulation, and tumor size (consistent with the "Big Five" risk factors).
Conclusions
Preoperative anatomical complexity remains the dominant predictor of postoperative functional outcome. Logistic-based ML classifiers had moderate overall accuracy but better performance for extreme outcomes (major decline or clear improvement). Using a higher threshold (e.g. ≥20-point KPS drop) improved discrimination and aligns with clinically meaningful outcome definitions. These findings support emphasizing high-impact anatomical features and outcome thresholds to improve prognostic models.