Uncertainty Gated Residual Learning with Patient Level Bagging for Longitudinal Glioma Progression Prediction

Abstract

Precise prediction of tumor progression between successive post treatment MRI scans could support timely clinical decision making in glioma management. However longitudinal radiomic modeling is still challenged by patient level data leakage and uncertainty in prediction. The study of two stage machine learning pipeline using the MU-Glioma-Post dataset  which has 203 patients; multi-modal T1c, T1n, T2f, T2w MRI with tumor segmentations and clinical records. Non-leaky longitudinal pairs where n = 391, 155 patients were constructed with progression defined as a ≥25% increase in tumor volume at the subsequent timepoint, occurence 38.1%. Custom 3D radiomic, texture and time safe clinical event features fed an anchor ensemble e.g. logistic regression, tree ensembles, XGBoost, CatBoost followed by an uncertainty gated residual correction stage and a monotonic tumor kinetics expert combined via patient level bagging under strict nested, patient grouped cross validation. The pipeline achieved a nested out-of-fold AUC of 0.728 (95% CI: 0.678–0.777), exceeded a manuscript reference benchmark of 0.713. Balanced accuracy (0.691 vs. 0.672), F1-score (0.638 vs. 0.607) and sensitivity (0.738 vs. 0.658) improved while specificity decreased modestly (0.645 vs. 0.686). Uncertainty gated residual correction combined with patient level bagging improves discrimination and sensitivity for longitudinal glioma progression prediction over baseline benchmarks, though specificity trade-offs requires further calibration before clinical translation.