Open Collections

Akter, Tanjila

University of British Columbia

Childhood mortality remains a significant public health challenge globally, particularly in low- and middle-income countries (LMICs). Despite advances in healthcare, many children under five continue to die following discharge from hospitals, particularly after treatment for severe infections. The transition from hospital to home is a vulnerable period yet often overlooked, with prediction models focusing on in-hospital outcomes. The Smart Discharge approach was developed to address this gap, leveraging predictive algorithms to identify children at high risk of post-discharge mortality. However, current models primarily rely on admission data, overlooking in-hospital disease progression and discharge characteristics that could further improve the accuracy of risk predictions. This thesis seeks to enhance the Smart Discharge model by incorporating both admission and discharge variables into the predictive algorithm. We used an elastic net regression approach incorporating both admission and discharge characteristics to develop an improved post-discharge mortality prediction model for children under 5. We applied 10-fold cross-validation for internal validation and assessed improvements in risk classification with the Net Reclassification Index (NRI). Subgroup analyses were performed to compare predicted mortality across different strata, and missing data were addressed using k-nearest neighbor (KNN) imputation for admission variables and Multiple Imputation by Chained Equations (MICE) for discharge variables. In the under-6-month cohort, the enhanced admission plus discharge model (AUROC=0.82, 95% CI: 0.79 – 0.85) significantly outperformed the admission-only model (AUROC= 0.75, 95% CI: 0.72 -0. 79). The precision-recall AUC was also higher (PR-AUC = 0.37 vs 0.24), indicating better identification of true positives. Calibration plots showed that the combined model was better calibrated at higher predicted probabilities. Similarly, the 6-60-month cohort showed superior performance with the combined model (AUROC = 0.81, PR-AUC = 0.28) versus the admission-only model (AUROC = 0.75, PR-AUC = 0.17). Key discharge variables in the final model for both under-6 and 6-60 months cohorts included oxygen saturation at discharge, discharge feeding status, and discharge status (e.g., unplanned discharge or referral to higher care). The enhanced Smart Discharge models significantly improved predictive performance compared to admission-only models, enabling more accurate identification of children at risk of post-discharge mortality.

Text

2025-03-03

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/90453

Population and Public Health

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/