A Machine Learning Framework for Crisis Standards of Care: Stochastic Modelling of Pathogen Incursion and Triage Optimization for the 2026 FIFA World Cup

Date: June 2026Concept: Generalizable Mass-Gathering Triage Framework & Policy Simulator 🚨 Interactive Policy Simulator: Access the live Executive Decision Support System directly via your browser: Launch Web Dashboard Project Overview This repository hosts a generalizable, end-to-end computational pipeline designed to evaluate the healthcare system and macroeconomic risks of a hypothetical 'Black Swan' event. Utilizing the 2026 FIFA World Cup in Mexico as a primary mass-gathering case study, the framework simulates a simultaneous multi-pathogen incursion of Ebola Bundibugyo (BDBV)—modeled during an active WHO Public Health Emergency of International Concern (PHEIC)—and pandemic Influenza H5N1 during a Southern Hemisphere winter surge vector. Utilizing a 5-stage methodological framework, the pipeline establishes a Pathogen Incursion Probability Engine via a Poisson arrival process, synthesizes operational clinical cohorts, trains an ensemble machine learning meta-learner optimized for asymmetric misclassification costs, models discrete-time hospital capacity saturation under Crisis Standards of Care (CSC), and projects long-term societal and macroeconomic health deficits using Global Burden of Disease (GBD) and WHO-CHOICE parameters. Generalizability & Adaptability While parameterised herein for H5N1 and BDBV dynamics, the core contribution of this pipeline is its modularity. The deterministic and stochastic engines are entirely pathogen-agnostic. In the event of a novel "Disease X" emergence or any future global Mass-Gathering Event facing an active PHEIC, epidemiologists can rapidly redeploy this framework by updating the parameter matrices (config.R) with new latency distributions and biomarker covariances. The ensemble meta-learner will autonomously re-train and establish optimal decision boundaries within minutes. Repository Structure The project is organized as a modular, self-contained R project: ├── msc-triage-ebola-influenza-fifawc26-mexico.Rproj # Central RProject Anchor ├── renv.lock # Reproducible Environment Lockfile ├── data_dictionary.md # Detailed Feature Metadata ├── README.md # Project Master Documentation ├── assets/ # Repository Badges & Static Image Assets ├── docs/ # Pre-compiled Manuscript PDF & Supp. Materials ├── data/ │ ├── dataset_ebola_vs_influenza_fifawc26_triage_mexico.csv # Ground-Truth Cohort │ └── dataset_operational_triage.csv # Operational Dataset (With Noise) ├── src/ │ ├── config.R # Global Configuration & Parameter Constant Inversion │ ├── 00-pipeline-orchestrator.R # Master Execution Controller & Error Handling Engine │ ├── 01-clinical-simulation.R # Parametric Stochastic Cohort Generation (N=165,500) │ ├── 02-exploratory-data-analysis.R # Clinical Profiling & Multicollinearity Diagnostics (VIF) │ ├── 03-predictive-modelling.R # Stacked Generalization Ensemble (Ridge-RF-XGBoost) │ ├── 04-micro-impact.R # Discrete-Time Resource Saturation Simulation │ └── 05-macro-impact.R # Bounded SIR Community Spillover & Macroeconomics ├── reports/ │ ├── figures/ # Publication-Ready Visualizations (Absolute Savings, etc.) │ ├── models/ # Serialized Binary Model Stacks (.rds) │ └── results/ # Epidemiological & WHO-CHOICE Policy Sensitivity Matrices └── epitriage-dashboard/ ├── app.R # Interactive Policy Shiny Dashboard (UI/Server) ├── data/ # Synchronized Dashboard Core Datasets & Living Visual Objects └── www/ # High-Resolution NATIVE Plots For Executive Views Methodological Framework 1. Stochastic Cohort Generation (src/01-clinical-simulation.R) Generates a balanced clinical population (N = 165,500) calibrated against historical WHO 2007 BDBV outbreak dynamics and pandemic influenza markers. This stage operates strictly as a parametric stochastic simulation. Patient profiles incorporate demographics, right-skewed latency rates derived via Gamma distributions, log-normal multivariate hematological features (capturing severe thrombocytopenia and leukopenia), and conditional length-of-stay (LOS) metrics dependent on terminal survival outcome vectors. 2. Exploratory Data Analysis & Diagnostics (src/02-exploratory-data-analysis.R) Calculates epidemiological characteristics stratified by etiological agent. Implements Variance Inflation Factor (VIF) diagnostic checks through an auxiliary multivariable generalized linear model (GLM) to validate predictive feature independence (enforcing a strict health-sector threshold of VIF < 2.5) prior to machine learning training loops. 3. Stacked Ensemble Meta-Learner (src/03-predictive-modelling.R) Injects 10% to 15% instrument variance and recall field bias into physiological data to mimic an acute healthcare surge environment. Combines heterogeneous base learners—Regularized Logistic Regression (Ridge Bounds), Random Forests (Non-linear interactions), and Extreme Gradient Boosting (XGBoost Error Correction)—via Penalized Meta-Model Stacking. The final classifier optimizes the Precision-Recall Area Under Curve (PR-AUC) and configures an F-beta=2 decision threshold to heavily penalize fatal False Negatives (missing Ebola isolation) over managed False Positives. 4. Micro-Operational Impact Simulation (src/04-micro-impact.R) Executes a 60-day discrete-time stochastic resource simulation. Incorporates dynamic sigmoid optimization boundaries to adjust clinical log-odds triage constraints relative to real-time hospital occupancy rates across a baseline threshold of 350 emergency-retrofitted AIIR units. Evaluates operational infrastructure elasticity over a multivariable policy parameter landscape via functional mapping. 5. Macroeconomic Health Economics (src/05-macro-impact.R) Translates hospital overflow thresholds into community transmission matrices through a bounded SIR spillover simulation across a susceptible population envelope (N = 5.0M). Quantifies long-term multi-sectoral societal health deficits isolating premature mortality Years of Life Lost (YLL) and Disability-Adjusted Life Years (DALYs). Introduces a multi-scenario macroeconomic burden analysis applying WHO-CHOICE Value of Statistical Life (VSL) ranges (1x, 3x, and 5x GDP per capita). Interactive Shiny Dashboard (epitriage-dashboard/app.R) The deterministic and stochastic outputs have been encapsulated into a production-grade interactive dashboard. To bridge the gap between computational modelling and actionable public health policy, this dashboard allows policymakers to: Perform real-time multivariable sensitivity analyses. Manipulate epidemiological reproduction numbers (Rt). Dynamically adjust hospital capacity constraints (AIIR beds). Instantly evaluate the resulting WHO-CHOICE macroeconomic burden projections. Access the live application without requiring a local R environment: EpiTriage 2026 Dashboard The architecture comprises six specialized analytical views: Executive Summary: Real-time KPI tracking showcasing Poisson arrival risks, maximum lives saved, and absolute economic burden averted. Interactive Policy Simulator: Active user sliders enabling real-time scenario modelling by manipulating susceptible demographic sizes, effective reproduction numbers (Rt), systemic collapse mortality metrics, and infrastructure capacity caps. Operational Triage (Micro): Analyzes hospital infrastructure saturation dynamics and visualizes the asymmetric cost matrix of the triage algorithm. Macroeconomic Impact: Details the WHO-CHOICE burden sensitivity across GDP and VSL variants. Clinical & ML Diagnostics: Presents the synthetic cohort's EDA, biomarker distributions, and the PR-AUC evaluation of the stacked meta-learner. Raw Data Explorer: Allows dynamic filtering and export of the generated datasets, efficacy tables, and sensitivity matrices. Conclusions & Limitations Key Findings: The computational pipeline demonstrates that waiting for definitive syndromic clustering during an MGE is a mathematically catastrophic policy. By integrating a Stacked Generalization meta-learner optimized through an asymmetric Fβ=2 cost function, the algorithm prioritizes clinical sensitivity over precision. Explicitly tolerating managed false positives to guarantee zero false negatives is the only viable strategy to protect healthcare elasticity under CSC. Implementing algorithmic triage averts systemic collapse and effectively prevents multi-trillion dollar ($> 3.97 trillion USD) economic paralysis. Limitations: The model is constrained by its reliance on a synthetic operational sample frame. While tightly calibrated to historical WHO reference parameters, the haematological covariance matrices abstract the complete complexity of real-world multi-morbidity profiles. Furthermore, the algorithm's discriminative power relies on the existence of physiological or chronological divergence between concurrent pathogens. Installation & Reproducibility This project utilizes the renv package management system to guarantee local environment and package version lock-in consistency across systems. Hardware Prerequisites & Performance Metrics The script stack utilizes parallel processing (doParallel) to accelerate hyperparameter grid sweeps. Performance metrics and cross-validation loops have been natively compiled and optimized locally on Apple Silicon architecture hardware running macOS, ensuring exceptionally fast computation times during the ensemble compilation and simulation phases. While the dashboard is available via web deployment, the full pipeline can be reproduced locally. Execution