Hospital Triage Scheduling Simulation

ISEF Category: Engineering Technology: Statics and Dynamics

Subcategory: Industrial Engineering-Processing  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Emergency departments do not run like a neat line at a school cafeteria. One delayed decision can ripple through every patient after it. If you can model that queue well, you can measure real minutes saved, lost, or reallocated. That makes this project both practical and very data-rich.

What Is It?

This project studies how a hospital-like line works when one server, like a doctor or triage station, can help only one patient at a time. Patients do not all behave the same. Some leave before being seen, some need help fast, and some take longer once treatment starts. That waiting-and-leaving behavior is called reneging.

You will compare different rules for deciding who gets served next. FIFO means first in, first out. Acuity-weighted scheduling gives priority to sicker patients. A machine learning model can also predict length of stay, or LOS, which means how long a patient will likely occupy care once started. Think of it like choosing checkout order, except the stakes are human health and the metric is total patient-minutes saved.

Why This Is a Good Topic

This is a strong science fair topic because you can test real scheduling rules with public health data and clear output numbers. The question is narrow, measurable, and tied to a real problem that hospitals care about. You can learn queueing theory, simulation, and model comparison without inventing a new medicine or building a physical device.

Research Questions

• How does FIFO compare with acuity-weighted scheduling in total patient-minutes saved?
• What is the effect of using predicted LOS instead of triage order on average waiting time?
• Does adding reneging behavior change which scheduling policy performs best?
• To what extent does patient acuity improve queue performance when arrival rates rise?
• Which policy reduces the number of patients who leave before being seen?
• How does prediction error in ML-based LOS estimates change the policy ranking?

Basic Materials

• Laptop or desktop computer with at least 8 GB RAM.
• Spreadsheet software for initial data inspection.
• Python installed with pandas, NumPy, SciPy, and simpy.
• Access to the MIMIC-IV database and required training or credentials.
• Text editor or notebook for documenting assumptions.
• Digital timer or calendar for tracking project milestones.

Advanced Materials

• Laptop or workstation with 16 GB or more RAM.
• Python with pandas, NumPy, SciPy, scikit-learn, simpy, and statsmodels.
• Access to the MIMIC-IV database and data use agreement compliance.
• Jupyter Notebook for simulation experiments and plots.
• Git for version control.
• Optional cloud notebook or university server for running many simulation trials.

Software & Tools

• Python: Builds the simulation, cleans the data, and calculates queue metrics.
• Jupyter Notebook: Lets you test ideas, document assumptions, and show results clearly.
• pandas: Organizes triage records, wait times, and patient-level variables.
• simpy: Runs discrete-event queue simulations with different scheduling rules.
• scikit-learn: Fits a model that predicts length of stay from patient features.

Experiment Steps

1. Define the queue system you want to simulate, including who counts as a patient, what counts as service, and when reneging happens.
2. Choose one outcome metric first, such as total patient-minutes saved, then decide how you will also track waiting time and abandonment.
3. Build a baseline FIFO model so you have a plain comparison point.
4. Add an acuity-based rule, then define how you will turn triage severity into a priority score.
5. Train or approximate an LOS prediction model, then test how prediction error changes scheduling results.
6. Compare policies across many simulation runs, then test whether the differences stay large under uncertainty.

Common Pitfalls

• Using raw hospital timestamps without cleaning out missing or impossible records, which breaks the queue logic.
• Mixing triage priority with treatment duration, which makes it hard to tell whether a policy works because of ordering or because of shorter stays.
• Ignoring reneging, which can make a policy look better than it would in a real waiting room.
• Training the LOS model on the same patients used to evaluate the policy, which inflates performance.
• Reporting one simulation run instead of many runs, which hides random variation.

What Makes This Competitive

A stronger version of this project goes beyond showing that one policy has a lower average wait. You would test policy performance across different arrival patterns, different reneging rules, and different levels of prediction error. You could also use confidence intervals, paired tests, or bootstrapped comparisons instead of a single chart. That kind of analysis shows that you understand both the model and the real system behind it.

Project Variations

• Compare the same scheduling policies on pediatric, adult, and emergency-only subsets of the data.
• Replace LOS prediction with a simpler rule based on triage score, then test whether the extra model complexity actually helps.
• Measure fairness metrics, such as whether one policy consistently delays lower-acuity patients more than others.

Learn More

• MIMIC-IV Documentation: Official database guide from PhysioNet, useful for understanding the data structure and access steps.
• PhysioNet: Free repository that hosts MIMIC-IV and related critical care datasets.
• MIT OpenCourseWare, Introduction to Operations Research: Free lecture materials on queues, optimization, and simulation.
• NIH PubMed: Search review articles on emergency department crowding, queueing theory, and patient flow.
• SimPy Documentation: Free guide to building discrete-event simulations in Python.
• US Census and CDC Data Resources: Helpful for background on hospital demand, though not specific to this model.