Open Collections

Abstract

Background: Prediction modelling has become increasingly prominent in global health as a potential tool to support evidence-based decision-making and optimize resource allocation, particularly in low- and middle-income countries (LMICs) where pediatric mortality remains high. Despite growing interest, major gaps persist across both the continuums of pediatric care and model development, including limited evidence for older children, insufficient external validation across settings, and challenges translating models into clinical practice. Addressing these gaps requires approaches that balance methodological rigor with feasibility and contextual relevance within resource-poor health systems. Methods: This thesis employed a multi-phase approach spanning predictor identification, external validation, and implementation planning. A modified Delphi study engaged multidisciplinary experts with experience in LMIC pediatric care to identify feasible and clinically relevant predictors of post-discharge mortality in children older than five years. An independent external validation study evaluated the performance and generalizability of 21 published pediatric in-hospital mortality prediction models using a large cohort of Ugandan children, applying multiple performance metrics including discrimination, calibration, and clinical utility. Finally, a structured implementation plan was developed to guide the translation of validated models into clinical practice, informed by contextual factors relevant to the setting. Results: The modified Delphi process identified 57 candidate predictors for post-discharge mortality, addressing a critical evidence gap for an understudied population. During external validation, logistic regression models generally outperformed simple scoring systems and more complex machine learning approached, with parsimonious models containing fewer variables demonstrating comparable or superior performance to more complex alternatives. The implementation plan outlined practical strategies to integrate prediction tools into existing workflows while minimizing burden on health systems and proposed frameworks to evaluate implementation success. Conclusion: This thesis demonstrates that effective pediatric mortality prediction in LMICs requires thoughtful considerations around predictor selection, parsimonious and context-specific models supported by robust external validation, and strategic implementation planning. The findings challenge assumptions that greater model complexity yields better performance and emphasize the importance of feasibility and local relevance. Together, this work links predictor identification, model validation, and implementation, offering evidence to advance pediatric mortality prediction models from research toward meaningful clinical impact for children globally.