Data Accessibility Statement

Due to the sensitive nature of the data collected for this study, access to the underlying data is restricted and overseen by the relevant Data Governance Committee. The dataset and detailed access instructions are available via the Harvard Dataverse (https://doi.org/10.7910/DVN/X6FAGX). The code files included in this submission are compatible with the CC0 waiver required by Dryad for publication; however, the input files required to run the code are not included due to data sensitivity constraints. If there are any questions or concerns regarding data or code availability, readers are encouraged to contact the corresponding author of this dataset.

Scripts

The set of scripts (scripts.zip) to analyze the data from this study and generate all the results and figures presented in the manuscript are contained in the scriptsisp file and are described below. The script should be run in the sequence presented.

packages.csv

List of packages and version numbers for all packages needed for reproduction of the analyses in the manuscript.

figure_0.1_install_packages.R

Sets up the project environment using the renv package, ensuring Bioconductor version 3.17 and any specific package versions needed for the analysis.

figure_0.2_helper_functions.R

A set of custom helper functions needed across the analysis includes model training with XGBoost, pathway analysis, model performance assessment, etc.

figure_1.1_dataset_description.R

Generates descriptive statistics the number of patients per sfor ite and their relationship with malnutrition strata. Also generates descriptive statistics for the primary clinical and multiomic feature sets.

figure_1.2_plotting.R

The visualization script for Figure 1. This script loads and processes multi-cohort patient and omics data, then generates bar plots and alluvial diagrams to visualize participant composition, outcomes, malnutrition strata, and dataset size/modularity across cohorts.

figure_2.1_survival_model.R

Executes the predictive workflow for the admission cohort (A0), training individual modality models and an integrated multiomic model while extracting "Gain" scores to rank feature importance. Calculates performance metrics for models trained.

figure_2.1.1_survival_model_site_analysis.R

Evaluates model generalizability through leave-one-site-out cross-testing and uses mixed-effects modeling and meta-regression to identify features with consistent effects across all nine study sites.

figure_2.1.2_survival_model_across_time.R

Slices survival data into Early, Mid, and Late windows (0-3d, 3-14d, >14d) to determine if admission biological signals remain predictive of late-occurring mortality.

figure_2.1.3_survival_model_external_factors.R

Correlates multiomic risk scores with "domain scores" such as Caregiver Characteristics and Access to Healthcare, and calculates Hedges' g for specific clinical signs to see how biological risk aligns with physical symptoms.

figure_2.2_feature_deviations.R

Identifies time-specific biological markers by running Wilcoxon tests and Hedges' g calculations specifically for patients who died within discrete timeframes (0-3d, 3-14d, >14d).

figure_2.3_omic_distances.R

Quantifies biological "dysregulation" by scaling patient data against a community baseline (CP) and calculating the L2 norm (Euclidean distance) for each patient’s omic profile relative to healthy community children.

figure_2.4_pathway_analysis.R

Prepares data for biological interpretation, including proteomic, lipidomic, and metabolomic pathway analysis.

figure_2.4.1_metabolomics_validation.R

Performs a direct comparison between the discovery cohort's metabolomic fold-changes and a secondary study (Wen et al., 2022) to confirm the robustness of metabolic mortality signals.

figure_2.5_minimal_model.R

Uses a bootstrapping approach and XGBoost importance rankings to identify a minimal set of features (10 omic markers plus anthropometry) that retain the predictive power of the full multiomic set.

figure_2.6_plotting.R

The visualization script for Figure 2. This script evaluates and visualizes multi-omic survival prediction models, comparing performance across datasets, sites, time windows, and model types using ROC/PR curves, effect sizes, and pathway enrichment analyses.

figure_3.1_clinical_vs_omics.R

Defines risk subgroups based on the "error" (difference) between integrated multiomic scores and clinical scores and characterizes these subgroups.

figure_3.2_minimal_model.R

Trains specialized predictive models to distinguish between biological and clinical risk categories, identifying the specific biomarkers that signal high risk when clinical exams appear normal.

figure_3.3_omic_discrepancy.R

Computes an "Omic Discrepancy" score (standard deviation across modality scores per patient) to identify individuals with highly discordant biological signals across different systems and characterizes these subgroups.

figure_3.4_plotting.R

The visualization script for Figure 3. This script analyzes clinical and multi-omic risk prediction in an admission cohort by comparing clinical, integrated, and minimal models. It generates figures showing risk score agreement, survival curves, model performance (AUROC/AUPRC), omic discrepancy, and pathway enrichment. It also identifies key clinical and omic features associated with outcomes using statistical testing and heatmap visualizations.

figure_4.1_survival_model_discharge.R

Replicates the integrated modeling and biological distance calculations for the discharge (D0) cohort to predict post-discharge mortality. Calculates performance metrics for models trained.

figure_4.2_clinical_vs_omics_discharge.R

Defines risk subgroups based on the difference between integrated multiomic scores and clinical scores at discharge and characterizes these subgroups.

figure_4.3_omic_discrepancy_discharge.R

Computes an "Omic Discrepancy" score to identify individuals at discharge with highly discordant biological signals across different systems and characterizes these subgroups.

figure_4.4_admission_vs_discharge_model.R

Tests model reciprocity by training an admission model to predict discharge outcomes and vice versa to assess the stability of the mortality signature over time.

figure_4.5_hospitalization_model.R

Focuses on longitudinal change by calculating "Delta" values (D0 - A0) for all multiomic features and testing their association with mortality with predictive modeling.

figure_4.6_plotting.R

The visualization script for Figure 4.

figure_5.1_malnutrition_strata_analysis.R

Investigate if malnutrition-related mortality risk is mediated by the same multiomic signals or if different biological drivers exist for children at different levels of malnutrition.

figure_5.2_age_group_analysis.R

Investigates if age-related Investigatesk is mediated by the same multiomic signals or if different biological drivers exist for children of different ages.

figure_5.3_plotting.R

The visualization script for Figure 5. This script examines how age and malnutrition modify multi-omic survival signals by performing mediation, stratified performance, and discrepancy analyses across admission and discharge cohorts. It visualizes how these characteristics influence model performance, omic distances, and feature effect variability.

figure_6.1_survival_model_validation.R

Validates the trained multiomic and minimal models on the independent validation (V0) cohort.

figure_6.2_clinical_vs_omics_validation.R

Validates the risk subgrouping logic (e.g., Biological Risk Not Clinically Reflected) in the validation cohort.

figure_6.3_plotting.R

The visualization script for Figure 6. This script evaluates and validates proteomic and clinical survival risk models across discovery and validation cohorts, comparing predictive performance, feature consistency, and omic distances. It generates figures assessing model discrimination, calibration, and biological agreement between cohorts.

Requirements

• R version 4.3.1

Human subjects data

All caregivers provided written informed consent for their child to participate in the study and for the data produced to be accessible after de-identification. Caregivers were assured that any report or publications on this study would not use participant’s names or identities. They agreed to information being shared with other researchers in ways that do not reveal individual participants’ identities. Any information that could identify people, such as their names and general locations, were replaced with number codes.