ayabz22/sepsis-mortality-prediction

Analysis code for a prospective cohort study on host-response biomarkers and mortality risk stratification in pediatric sepsis in Tanzania. Submitted to Nature, 2026. ## Host-Response Biomarkers Improve Risk Stratification in Children with Sepsis in Tanzania: A Prospective Cohort Study This repository contains the analysis code I wrote for a multi-site prospective cohort study on pediatric sepsis mortality prediction in Tanzania, led by PI Teresa Kortz (UCSF) and Hendry Sawe (MUHAS/MNH). I led the statistical analysis and model development, data cleaning, biomarker screening, feature selection, model tuning, and validation, and contributed to manuscript drafting. The study is a large collaborative effort across UCSF, Muhimbili National Hospital, and Muhimbili University of Health and Allied Sciences. ## Project overview I investigated the relationship between clinical severity scores, host-response biomarkers, and in-hospital mortality in children with sepsis. The goal was to determine whether biomarkers such as sTREM-1, IL-8, and procalcitonin improve mortality prediction beyond standard clinical-sign-based scores (qSOFA, LODS, SICK), and whether a model restricted to locally feasible point-of-care biomarkers could retain that performance for use in resource-constrained settings. ## Analysis workflow **Baseline characteristics** - `table1_and_table2.R` — Baseline clinical characteristics and biomarker summaries by survival status. **Main combined model (LODS + sTREM-1 + IL-8)** - `clinicalscores_biomarkers_table.R` — Cross-validated AUC for each biomarker alone and combined with clinical scores. - `feature_selection.R` — Random Forest and XGBoost biomarker importance ranking. - `Biomarker_combinations.R` — Selects the best biomarker combination and evaluates it on the test set (AUC, ΔAUC, NRI). - `machine_learning_comparison.R` — Compares logistic regression, LASSO, random forest, and GBM across candidate models. - `calibration_main_model.R` — Brier score, calibration slope/intercept, and calibration plots. - `triage_main_model.R` — Classification performance and likelihood ratios across five predicted-probability threshold …