Preoperative model predicts functional outcomes after glioma resection using clinical and MRI variables

Background and Objectives Preoperative prediction of functional outcomes in contrast-enhancing glioma could support surgical decision-making and patient counseling, yet most existing models incorporate histopathological or postoperative variables unavailable before surgery. Our objectives were to develop a preoperative-only prediction model for one-year functional status and evaluate the added value of MRI-based tumor characteristics beyond clinical predictors. Methods We conducted a retrospective cohort study of consecutive adults (18 years old and older) undergoing first resection of supratentorial contrast-enhancing glioma (WHO grade 2 and above, histopathologically confirmed postoperatively) at a single center, with one-year follow-up. The primary outcome was functional status classified as mortality (Karnofsky Performance Score (KPS) = 0), functional dependence (KPS 10-60), or functional independence (KPS = 70 and above). In addition to clinical variables (age, sex, preoperative KPS, preoperative seizures), a deep learning tool was used to extract structural MRI-based tumor characteristics as predictors. A machine-learning model was developed and conformal prediction was applied to stratify patients by prediction confidence level. Results 552 patients were included (median age: 60 years, range: 18-84; median contrast-enhancing volume: 24 mL, IQR: 10-43; median preoperative KPS: 80, range: 30-100; retrospectively confirmed 88% glioblastoma). Most MRI-based predictors did not improve performance as the best-performing model included three predictors: age at diagnosis, contrast-enhancing volume, preoperative KPS. Bootstrapped areas under the curves were 0.77 (95% confidence interval 0.70-0.84) for mortality, 0.64 (0.52-0.77) for functional dependence, and 0.71 (0.63-0.79) for functional independence. F1 scores per class were 0.65, 0.24, 0.65, respectively. Conformal prediction provided reliable predictions for 18% patients, moderate uncertainty for 57%, and identified 25% with genuinely unpredictable outcomes. Discussion Our preoperative machine-learning model predicted one-year functional status in contrast-enhancing glioma with functional independence being the most reliably classified outcome (ROC-AUC = 0.77, F1 score = 0.65) and functional dependence the most challenging to predict (ROC-AUC = 0.64, F1 score = 0.24). A small set of three preoperative predictors drove model performance, supporting generalizability to broader patient populations. Our open-source model enables individualized risk stratification and may help clinicians identify patients with uncertain prognoses warranting more intensive preoperative counseling or follow-up planning.