P10.15.B ENHANCING GLIOBLASTOMA TREATMENT STRATEGIES: A MACHINE LEARNING DECISION TREE MODEL FOR PREDICTING PROGRESSION FREE SURVIVAL USING INTEGRATED SURGICAL, VOLUMETRIC, AND MOLECULAR DATA

Abstract

Abstract BACKGROUND Predicting progression-free survival (PFS) in glioblastoma (GB) is crucial for post-surgery treatment decisions. Detecting post-operative volumes via pre-radiotherapy (preRT) MRI, combined with clinical and molecular data, improves prognosis. A decision tree model using these features provides valuable insights into PFS when aligned with the new RANO resection categories. MATERIAL AND METHODS We analyzed retrospective data from 205 GB treated by STUPP protocol across three French centers (2009-2021) as part of the AIDREAM project. Data included clinical and anatomopathological information. Advanced automated segmentation of tumor subregions on preRT MRI was used, subsequently revalidated by experts, and a detailed brain atlas to refine lesion site. A survival analysis was performed using Kaplan-Meier with log-rank tests to compare survival distributions according to clinical data, surgical extension and the 4 RANO volume subcategories. A conventional univariate and multivariate Cox proportional HR analysis based on clinical/anatomopathological data and residual non-contrast-enhanced (nonCE) and contrast-enhanced (CE) volumes on pre-RT MRI was also conducted. The SurvivalTree algorithm, chosen for its capability to handle survival data and missing values, was applied in two experimental setups: one using only clinical data and another using combined clinical and volumetric data. Variables included age, sex, KPS, surgical extension, MGMT, tumor site and nonCE, CE, tumor bed and postoperative changes volumes. In our 5-fold cross-validation process, we individually assessed the importance of each feature in every fold through permutations. This approach enabled us to evaluate the consistency and dependability of feature importance across various data subsets. We then calculated the mean importance of each feature over all folds. RESULTS Median PFS was 8 months (1-62) and median OS was 18 months (2- 93). Based on the log-rank test, it was shown that extensive surgical resection and lower RANO categories improved PFS, with significant p-values. MGMT status correlated strongly with longer PFS (p < 0.001). Univariate analysis highlighted significant predictors including nonCE (p=0.002) and CE volume (p=0.017), while multivariate analysis confirmed only the significance of nonCE volume (HR 1.011, 95% CI 1.005-1.018, p < 0.001). The decision tree model’s performance increased with volumetric data, from a mean C-index of 0.546 ± 0.057 to 0.576 ± 0.066, identifying CE and nonCE volumes and MGMT status as key predictive factors. CONCLUSION Integrating surgical, volumetric, and molecular data into a decision tree model could effectively enhance the prediction of PFS in glioblastoma. Expanding the study to the entire AIDREAM cohort of 733 patients will validate the robustness of the model and its clinical utility for personalized treatment strategies.

Ähnliche Arbeiten