🏥 NHS Triage Simulation Framework

Reducing Patient Wait Times in NHS Triage Using a Mixture-of-Agents Simulation Framework

Author: Raunak Burrows Supervisor: Dr. Xiatian Zhu — Senior Lecturer, People-Centred AI, University of Surrey Affiliation: MSc Artificial Intelligence (2025), University of Surrey

---

📖 Abstract

This repository implements the research project “Reducing Patient Wait Times in NHS Triage Using a Mixture-of-Agents Simulation Framework.” The work addresses a persistent gap in NHS emergency department (ED) operations — post-triage patient routing — by designing a Mixture-of-Agents (MoA) system that dynamically coordinates multiple specialist AI agents to allocate resources efficiently.

The framework achieved:

• 58.2 % reduction in average doctor wait time (204.3 → 85.3 minutes)
• 40.5 % reduction in MRI wait time and 60.8 % in ultrasound wait time
• 19.3 % improvement in resource utilisation
• Bias reduction: demographic parity gap improved from 15.9 % → 1.2 %
• Clinical safety maintained at 94.6 % appropriateness

All experiments were validated via 30 independent simulation runs (95 % CI) and expert review on 237 synthetic cases.

---

🎯 Problem Context

• Only 76.4 % of NHS A&E patients are seen within 4 hours (Q4 2023-24).
• Over 35 000 patients experience > 12-hour delays.
• The Manchester Triage System (MTS) — used in > 90 % of UK EDs — focuses solely on initial classification.
• Post-triage routing remains manual, sequential, and resource-inefficient.

The proposed MoA framework builds on MTS and fuzzy-MTS literature to enable data-driven, context-aware routing using collaborative AI agents.

---

🧠 Methodology Overview

Comparative Framework

1. Rule-Based (Baseline) — traditional MTS flow: sequential physician-first protocol.
2. Single-Agent — GPT-OSS-20B model (with Mistral-7B tested for NHS deployment) generating routing recommendations.
3. Multi-Agent (MoA) — neurology, cardiology, and abdominal agents coordinated via LangGraph aggregator.

Technical Stack

• Simulation Core: SimPy 4 Discrete-Event Simulation (M/M/c queues, Poisson arrivals λ = 12 patients/h)
• Synthetic Data: 10 000+ FHIR R4 (UK Core) compliant patients via modified Synthea workflows
• Routing Logic: LangGraph multi-agent collaboration layer with Ollama runtime for local LLM execution
• Bias Framework: intersectional demographic analysis (adapted from Lee et al., 2024)

---

⚙️ Installation

git clone https://github.com/burrows99/nhs-triage-simulation.git
cd nhs-triage-simulation
pip install -r requirements.txt

---

🚀 Usage

Quick Run

python main_simulation.py

Full Experiment Suite

python scripts/run_experiments.py --all

Bias Analysis

python analysis/bias_evaluation.py --intersectional

Performance Reports

python analysis/performance_metrics.py --visualize

---

📁 Repository Structure

nhs-triage-simulation/
├── main_simulation.py
├── src/
│   ├── simulation/         # SimPy environment & patient generator
│   ├── routing/            # Rule-based, Single-agent, MoA strategies
│   ├── agents/             # Specialist and aggregator agents
│   └── utils/              # FHIR parser, metrics calculator
├── analysis/               # Bias & performance analysis scripts
├── config/                 # Simulation and demographic parameters
├── data/                   # Synthetic + validation datasets
├── results/                # Figures and reports
└── docs/                   # Methodology, API, deployment guide

---

📊 Key Results

Metric	Rule-Based	MoA	Δ Improvement
Doctor Wait	204.3 min	85.3 min	-58.2 %
MRI Wait	42.2	25.1	-40.5 %
Ultrasound Wait	158.7	62.2	-60.8 %
Bed Wait	166.2	45.9	-72.4 %
Resource Use	76 %	91 %	+19.3 %

Clinical Accuracy (by Demographic)

Group	Rule	Single	MoA
White Male	84.3 %	87.6 %	94.2 %
Black Male	78.2 %	76.5 %	92.4 %
Asian Female	79.1 %	78.6 %	95.2 %

Bias Reduction: Demographic parity gap ↓ from 15.9 % to 1.2 %. Clinical appropriateness: 94.6 % maintained.

---

🔐 Deployment Considerations

• GDPR Compliance: All LLM components run within NHS-controlled infrastructure; no external data transmission.
• FHIR Integration: Fully compliant with HL7 FHIR R4 (UK Core).
• Audit Trails: Decision logging for clinical governance.
• Computational Cost: ≈ £0.04 per consultation (run-time validated on local EKS cluster).

---

🧪 Validation Protocol

1. Statistical: 30 independent simulation runs (95 % CI).
2. Clinical: 237 expert-reviewed synthetic cases across specialties.
3. Bias: Intersectional analysis (gender × ethnicity).
4. Safety: Confidence threshold monitoring and audit traceability.

---

📚 Academic Reference

Primary Citation:

Burrows, R. (2025). Reducing Patient Wait Times in NHS Triage Using a Mixture-of-Agents Simulation Framework. MSc Dissertation, University of Surrey.

BibTeX:

@mastersthesis{burrows2025nhs,
  title={Reducing Patient Wait Times in NHS Triage Using a Mixture-of-Agents Simulation Framework},
  author={Burrows, Raunak},
  school={University of Surrey},
  year={2025},
  type={MSc Dissertation},
  note={Available at: https://github.com/burrows99/nhs-triage-simulation}
}

---

🧩 Contribution Guide

We welcome contributions to extend the MoA framework:

• New specialist agents (pediatric, orthopedic, etc.)
• Enhanced bias mitigation strategies
• Integration with real FHIR clinical data
• Visualization and performance optimisation

git checkout -b feature/improvement
git commit -am "Add improvement"
git push origin feature/improvement

---

📞 Contact

Raunak Burrows 📧 [email protected] 🔗 LinkedIn 🐙 GitHub @burrows99

Supervisor: Dr. Xiatian Zhu — University of Surrey

---

🏛️ Acknowledgments

• University of Surrey — academic supervision and resources
• NHS England — operational data and policy context
• SimPy Community — open-source simulation framework
• Synthea Team — FHIR synthetic data generation
• Manchester Triage Group — standardised triage protocols

---

🔍 Keywords / Topics

healthcare-ai simulation nhs mixture-of-agents langgraph llm fhir simpy synthetic-data discrete-event-simulation ai-bias triage healthtech python university-of-surrey msc-dissertation