2270 - Evaluating AI Models with MRI Radiomics and Deep Learning Features vs. Clinical Variables for Predicting Extended Survival in Newly Diagnosed Glioblastoma

Purpose/Objective(s): Extended survival (=36 months) occurs in a small subset of glioblastoma (GBM) patients, yet robust predictive models are lacking. Radiomic and deep learning features from preoperative MRI have shown promise in GBM prognostication, but their added value beyond clinical variables for predicting extended survival is unclear. We compared clinical, radiomic, and deep learning imaging features for extended survival prediction in newly diagnosed GBM patients.

Materials/Methods: This is a single institution retrospective study. Three feature sets were extracted: (1) 18 clinical variables including demographics, MGMT methylation, EGFR amplification, extent of resection, and segmentation-derived tumor volumes; (2) 1,070 radiomic features (CaPTk) from four tumor subregions; and (3) VGG-19 ImageNet transfer learning features from T1-contrast and FLAIR sequences. Seven models spanning traditional ML, gradient boosting, and tabular foundation models (TabPFN, LimiX) were evaluated using 5×5 repeated stratified cross-validation (80/20 per fold) with cost-sensitive learning. Feature selection used Elastic Net regularization. Performance was assessed by AUC-ROC, Brier score, and Nadeau-Bengio corrected t-tests with Bonferroni correction. SHAP analysis identified key predictors.

Results: A total of 209 adult patients with newly diagnosed GBM were included. 38 were considered extended survival (=36 months), 171 were short-term (<36 months). The 36-month cutoff represents more than double the median GBM survival under standard treatment. Clinical-only models achieved the highest discrimination in our cohort. LimiX with 18 clinical features achieved AUC 0.744 (SD 0.066, 95% CI 0.717–0.769), followed by Elastic Net LR (0.720, SD 0.087) and TabPFN (0.718, SD 0.071). Traditional radiomics ensemble achieved AUC 0.578 (SD 0.102), significantly lower than clinical LimiX (p=0.033). Radiomics-only AUC was 0.578 (SD 0.102) with traditional ML but 0.717 (SD 0.111) with LimiX, approaching but not surpassing clinical-only models. Combined clinical-imaging models did not significantly improve over clinical-only performance (all p>0.05). LimiX showed better calibration versus Elastic Net (Brier 0.131 vs 0.195, p<0.001). Top SHAP predictors were age at diagnosis (24.7%), EGFR amplification (19.0%), contrast-enhancing volume (17.3%), and WHO grade (11.2%).

Conclusion: In the current study, radiomics and deep learning imaging features did not significantly improve extended survival prediction beyond clinical features alone. The tabular foundation model LimiX showed the highest discrimination and calibration among models tested. While clinical-molecular variables showed strong predictive value, the lack of radiomics improvement may reflect limitations in cohort size, imaging heterogeneity, or feature extraction. Multi-institutional validation is warranted.