Machine learning models trained on adult patients are increasingly used in clinical settings that include children. When applied to pediatric patients, particularly younger age groups whose anatomy and physiology differ substantially from adults, these models often fail systematically, producing confident but incorrect predictions without any indication that something has gone wrong. Retraining with labeled pediatric data is frequently infeasible due to data scarcity. This project develops out-of-distribution-aware uncertainty quantification frameworks that function as a practical safety layer on top of existing deployed models, explicitly increasing uncertainty for inputs that deviate from the training distribution. The framework produces uncertainty estimates that are better spatially aligned with prediction errors for out-of-distribution pediatric subgroups, and translates directly into more reliable downstream clinical tasks. By making silent failures visible, this research contributes to safer and more equitable AI deployment across diverse pediatric populations.