Machine Learning–Enhanced Prognostic Modeling in Elderly Glioblastoma IDH-wildtype: A Multidimensional Single-Center Cohort Study

Background: Glioblastoma IDH-wildtype (GBM IDHwt) in elderly patients presents challenges due to biological heterogeneity and underrepresentation in clinical trials. Despite rising incidence, prognostication remains inadequate, with treatment decisions based on subjective criteria. Objective: To determine clinical, radiological, surgical, and molecular determinants of survival in elderly GBM IDHwt patients and explore prognostic utility of machine learning (ML) models using clinical and pre-treatment data. Methods: We analyzed 155 patients aged ≥70 years with confirmed GBM IDHwt who underwent neurosurgery at a tertiary care institution. We examined variables related to clinical presentation, imaging, surgery, and molecular markers using multivariate regression and Histogram Gradient Boosting Regression (HGBR) ML models. Two ML models were developed: one incorporating full dataset variables, and another focusing on preoperative features. Results: Median overall survival (OS) was 11.3 months for patients undergoing resection and 3.7 months for biopsy. Independent predictors of prolonged OS included gross total resection (GTR), MGMT promoter methylation, non-acute symptom onset, and concomitant radiotherapy with temozolomide (RT + TMZ). ML models confirmed RT + TMZ and GTR as strongest predictors, while Karnofsky Performance Status (KPS) showed negative importance. BMI emerged as impactful, and the pretreatment model emphasized BMI and cognitive decline. Conclusion: This study confirmed prognostic relevance of GTR, MGMT methylation, and RT + TMZ combination. Baseline KPS and age did not demonstrate independent prognostic value, while BMI and cognitive decline were potential preoperative predictors. Our findings advocate a multidimensional, data-driven approach to preoperative risk stratification in elderly glioblastoma patients, which may facilitate individualized treatment strategies.