Health Economic Impact Estimation for AI-Assisted Diagnostic Triage: A Simulation Framework Using MIMIC-IV-Derived Parameters

A Monte Carlo health economic simulation evaluating the cost-effectiveness of an AI-assisted diagnostic triage tool for epilepsy, compared to standard NHS care pathways.

Overview

This project implements a probabilistic cost-effectiveness analysis (CEA) using a decision-tree microsimulation framework. It models two clinical pathways:

• Standard care: Current NHS diagnostic pathway for patients presenting with seizures
• AI-assisted triage: An AI tool that risk-stratifies patients at first presentation, fast-tracking high-risk patients to neurology

The simulation integrates three data sources:

1. MIMIC-IV (PhysioNet) -- empirical distributions for ED length-of-stay, hospital length-of-stay, readmission rates, and EEG ordering
2. Clinical literature -- diagnostic accuracy parameters, time-to-diagnosis, QALY utility weights
3. NHS Reference Costs (2023/24) -- unit costs for A&E visits, neurology consultations, EEG, and inpatient days

Key Results

Based on 10,000 Monte Carlo iterations using real MIMIC-IV data (N=10,050 seizure admissions):

Metric	Value
Mean incremental cost	-£237 per patient (AI saves money)
Mean incremental QALYs	+0.1455 per patient
Mean ICER	-£1,743/QALY (dominant)
P(cost-effective | WTP=£20k)	100.0%
P(cost-effective | WTP=£30k)	100.0%

The AI intervention dominates standard care in 100% of iterations -- it is simultaneously cheaper and more effective.

Repository Structure

health-economic-simulation-ai-triage/
├── health_economic_simulation_ai_triage.ipynb  # Main simulation notebook
├── data/
│   ├── mimic_extracted/              # MIMIC-IV CSVs (not included -- see instructions)
│   │   ├── DATA_README.md
│   │   └── extraction_queries.sql   # BigQuery SQL used for MIMIC-IV cohort extraction
│   ├── literature/                   # Reference PDFs for parameter calibration
│   └── nhs_reference/               # NHS Reference Cost data (not included -- too large)
│       └── DATA_README.md
├── _compile_check.py                 # Syntax checker for notebook code cells
├── _smoke_test.py                    # Standalone smoke test of core simulation logic
├── _run_test.py                      # Functional test runner (executes notebook cells)
├── requirements.txt                  # Python dependencies
├── LICENSE                           # MIT License
└── README.md

Getting Started

Prerequisites

• Python 3.9+
• Jupyter Notebook or Google Colab

Installation

git clone https://github.com/riyashet-hds/health-economic-simulation-ai-triage.git
cd health-economic-simulation-ai-triage
pip install -r requirements.txt

Running the Simulation

Option 1: Google Colab (recommended)

• Upload health_economic_simulation_ai_triage.ipynb to Google Colab
• Place MIMIC-IV CSVs in data/mimic_extracted/ (or let the notebook use synthetic fallback data)
• Run all cells

Option 2: Local Jupyter

jupyter notebook health_economic_simulation_ai_triage.ipynb

Data Setup

The notebook auto-detects data availability:

• With MIMIC-IV data: Place exported CSVs (admissions.csv, diagnoses_icd.csv, edstays.csv, patients.csv, procedures_icd.csv) in data/mimic_extracted/. The exact BigQuery queries used to build the cohort and export these CSVs are in data/mimic_extracted/extraction_queries.sql; see data/mimic_extracted/DATA_README.md for full extraction instructions.
• Without MIMIC-IV data: The notebook automatically generates synthetic data calibrated from published literature values. The simulation runs identically in either case.

Notebook Sections

1. MIMIC-IV Data Loading -- Load real or synthetic patient data
2. Distribution Fitting -- Fit lognormal distributions to ED/hospital LOS
3. Clinical Parameters -- Define costs, utilities, diagnostic accuracy
4. Standard Care Pathway -- Microsimulation of current NHS pathway
5. AI-Assisted Pathway -- Microsimulation with Bayesian belief-updating
6. Bayesian Triage Module -- AI tool sensitivity/specificity and belief updating
7. Monte Carlo CEA -- 10,000-iteration probabilistic cost-effectiveness analysis
8. Sensitivity Analysis -- Deterministic one-way sensitivity analysis with tornado diagram
9. Discussion -- Interpretation of results, limitations, and next steps

Figures Generated

• Figure 1: Distribution fits (ED LOS, Hospital LOS) -- empirical vs. fitted lognormal
• Figure 2: Cost-effectiveness plane (incremental cost vs. incremental QALYs)
• Figure 3: Cost-effectiveness acceptability curve (CEAC)
• Figure 4: ICER distribution histogram
• Figure 5: Tornado diagram (one-way sensitivity analysis)

Testing

# Check notebook syntax (no execution)
python _compile_check.py

# Run functional smoke test
python _run_test.py

Key References

• NICE NG217 (2022). Epilepsies in children, young people and adults
• Mulhern et al. Development of a QALY measure for epilepsy: NEWQOL-6D
• Mulhern et al. Valuations of epilepsy-specific health states
• Tittensor et al. Updating beliefs using a diagnostic decision-support aid in a nurse-led first-seizure clinic
• Lopes et al. (2019). Revealing epilepsy type using computational analysis of interictal EEG
• Schmidt et al. (2016). A computational biomarker of idiopathic generalised epilepsy from resting state EEG
• Bonnon et al. Comprehensive cost-effectiveness analysis of resective epilepsy surgery in an NHS setting
• Incidence and prevalence of epilepsy in the United Kingdom 2013-2018
• Impact of diagnostic delay on seizure outcome in newly diagnosed focal epilepsy (2020)

See data/literature/ for full PDFs.

License

This project is licensed under the MIT License -- see LICENSE for details.

Acknowledgements

• MIMIC-IV database via PhysioNet
• NHS National Cost Collection 2023/24