10286-CO-3 Development of a Practical Prognostic Scoring Model Using Preoperative MRI in Glioblastoma

Abstract Background Numerous radiomics-based prognostic models have been proposed for glioblastoma, yet their clinical application remains limited due to insufficient external validation and the complexity of the underlying algorithms. This study aimed to develop a simplified and interpretable scoring model and to propose a practical approach toward clinical translation. Methods We retrospectively analyzed 107 patients from our institution (OU) and 105 patients from The Cancer Genome Atlas (TCGA), all with newly diagnosed IDH-wildtype glioblastoma and available preoperative MRI (T1WI, T2WI, and Gd-T1WI). A total of 489 radiomic features were extracted from the preoperative images. Univariate Cox regression analysis was performed in each cohort to identify prognostic features (p < 0.05). Shared prognostic features were binarized using the median as a cutoff, and 1 point was assigned for values associated with poor prognosis. The total score was used to stratify patients. Model performance for survival stratification was evaluated using both the log-rank and Gehan–Wilcoxon tests to account for early deaths. Results A total of 27 prognostic features were commonly identified across both cohorts, primarily representing tumor sphericity and textural homogeneity. Based on the cumulative score (0–27 points), patients were stratified into low (0–9), intermediate (10–18), and high-risk (19–27) groups. Median overall survival was 21.0/19.6/13.3 months (OU), and 17.6/14.7/7.4 months (TCGA), for low, intermediate, and high-risk groups, respectively. The log-rank and Gehan–Wilcoxon tests yielded p-values of 0.01/0.03 (OU) and 0.05/0.002 (TCGA), respectively, highlighting the model’s utility in stratifying early mortality risk. Conclusion This scoring model offers a simple and interpretable framework for glioblastoma prognostication based on routinely acquired preoperative MRI. It demonstrated consistent performance across independent cohorts and showed high sensitivity to early death, supporting its potential role in personalized treatment planning and risk stratification in clinical trials.