IMG-02. Longitudinal clinicoradiomic modeling of overall survival in diffuse midline glioma patients using serial MRI scans

Abstract Background Diffuse midline gliomas (DMGs), including diffuse intrinsic pontine gliomas (DIPGs), are highly aggressive and fatal pediatric brain tumors. Accurate prediction of patient prognosis and treatment response at different time points during the treatment course may enhance clinical decision-making and facilitate the timely administration of alternative therapeutic approaches, thereby potentially improving patient outcomes. Longitudinal MRI-based assessments offer valuable insights into treatment response and mortality risk. However, studies investigating the application of longitudinal MRI in evaluating treatment response in DMGs remain limited. This study, through leveraging a multi-institutional dataset, explores the utility of longitudinal MRI-derived radiomic features, combined with clinical variables, for predicting overall survival (OS) in DMG patients. Methods Multiparametric MRI and clinical data were retrospectively analyzed for 32 pediatric DMG patients, including 20 participants from the PNOC003 and PNOC007 clinical trials and 12 patients from the Children’s Hospital of Philadelphia. Data included pre-treatment baseline and subsequent post-treatment follow-up imaging time points (at least three), resulting in a total of 158 imaging sessions. Radiomic features were extracted from pre- and post-contrast T1-weighted, T2-weighted, and T2-FLAIR images. Clinical features including age, sex and treatment methods were also included. A univariate Cox proportional hazards (Cox-PH) regression model was applied for feature selection, followed by a multivariate time-varying Cox-PH model with least absolute shrinkage and selection operator (LASSO) regularization using 5-fold cross-validation to predict OS based on time-varying radiomic features. Results Nine features, including texture features from T2-weighted and post-contrast T1-weighted images, along with treatment method were selected. The final predictive model achieved a concordance index (c-index) of 0.78 (95% confidence interval: 0.7–0.88). Conclusion This study highlights the potential of a longitudinal clinicoradiomic predictive model, which integrates non-invasive insights from serial MRI to capture tumor changes over the disease course, to estimate OS in pediatric DMG patients and inform individualized clinical management.