Simulating Patient Flow Across UC San Diego Health

The Problem

When the Emergency Department Becomes a Waiting Room

Imagine you've just been admitted to the hospital, but there are no inpatient beds available yet. Instead of moving to the right unit to start your care, you wait — in the emergency department — sometimes for many hours. This is called ED boarding, and it's a widespread problem in American hospitals. It leads to longer wait times for everyone in the ER, lower quality of care for admitted patients, and enormous stress on frontline staff.

UCSD Health operates three hospital campuses — La Jolla, Hillcrest, and East Campus — that collectively see hundreds of patients every day. In theory, a patient who can't get a bed at one campus could be transferred to another with available capacity. In practice, health systems have no safe way to experiment: you can't reroute hundreds of real patients just to see if a policy might work better.

Who This Is For

This tool is designed for hospital operations leaders and health system planners at UCSD Health — people responsible for capacity management, patient throughput, and inter-campus coordination. Rather than relying on gut instinct or historical averages, they can use this simulation to pressure-test transfer policies in a risk-free virtual environment before committing to any real change in operations.

Our Approach: A Virtual Replica of UCSD Health

We built a discrete event simulation (DES) — a computer model that recreates patient flow across all three UCSD Health campuses simultaneously. Think of it as a flight simulator for hospital operations: instead of risking real patients, administrators can test different transfer policies in the model, see what happens, and make smarter decisions before ever changing anything in the real world.

What makes our simulation unusual is that instead of inventing fictional patient paths, each virtual patient follows an actual historical journey drawn from a library of 150,337 real UCSD Health encounters — making it as faithful a replica of the real system as the data allows.

The Data

One Year of Real Patient Journeys

Our simulation is built on Admit-Discharge-Transfer (ADT) event data from UC San Diego Health — a de-identified record of every time a patient was admitted, moved between units, or discharged during a one-year observation period. Each event is timestamped and linked to a specific patient encounter, department, care tier, and campus.

The raw dataset contains 2.1 million events across 985,000 patient encounters. After filtering to patients who physically occupied hospital beds, we retained 150,337 complete patient trajectories suitable for simulation. For the length-of-stay prediction model, we held out 20% of encounters as a test set, with the remaining 80% used for training — ensuring the predictor was never evaluated on data it had seen before.

Patient arrivals vary meaningfully between weekdays and weekends — La Jolla averages 198 arrivals per weekday but only 149 per weekend day, a ~25% drop driven by fewer elective admissions. Our simulation models these patterns separately to reproduce the real rhythms of the health system.

Methods

Building a Virtual Hospital System

The simulation models each patient as a sequence of states — ED → Floor → ICU → Discharge — across any of the three campuses. We added a fourth care state, Boarding, to explicitly model patients who have been admitted as inpatients but are physically stuck in the ED waiting for a bed upstairs. This gives us 12 active states in total (4 tiers × 3 campuses), plus discharge.

Rather than assigning patients synthetic paths, each arriving patient is assigned an actual historical trajectory randomly sampled from the 150,337-encounter library. Bed capacity for each of the 12 states was calibrated from the data itself by observing steady-state occupancy in an unconstrained simulation run.

What We Built vs. What We Used

Predicting Who Needs a Long Stay

To enable smarter transfer decisions, we trained a length-of-stay (LOS) predictor using quantile regression. At the moment a patient is admitted from the ED, the model predicts whether their stay is likely to be long — using three features available at that moment: how long they've already been in the ED, how many times they've been admitted in the past 90 days, and which campus they're at.

We predict the 75th percentile of expected LOS rather than the average, because hospital LOS distributions are heavily skewed by a small number of very long stays. Patients whose predicted LOS exceeds a threshold are flagged for possible transfer to a less congested campus before they occupy a scarce inpatient bed.

How the Project Evolved

Several design decisions changed significantly after we saw early results:

• 1
  Single-campus → multi-campus. Early prototypes simulated La Jolla in isolation. We quickly realized that without modeling receiving campuses simultaneously, any transfer policy was meaningless — you can't evaluate where patients go if those destinations aren't in the model.
• 2
  Synthetic paths → real trajectories. Our first approach used transition probability matrices to generate patient paths stochastically. Switching to a library of 150K real historical journeys dramatically improved how faithfully the simulation reproduced actual UCSD Health patterns.
• 3
  Mean LOS prediction → quantile regression. A mean LOS predictor was highly sensitive to outlier long-stay patients. Moving to the 75th percentile gave a more robust and operationally useful trigger for flagging transfers.
• 4
  Single threshold → threshold sweep. Testing only one LOS threshold initially obscured the non-linear behavior we later discovered: East Campus ICU is catastrophically overwhelmed at a 3-hour Hillcrest threshold but partially recovers at 8 hours. The sweep revealed this instability.

Results & Impact

What the Simulation Reveals

After a 25–30 day warm-up period, bed utilization stabilizes into realistic steady-state patterns. The results are striking: two of the three campuses are chronically operating near or above capacity, with serious consequences for patient wait times.

See It in Action

The animation below shows a 30-second snippet of the simulation running the recommended transfer policy — routing long-stay patients out of La Jolla at a 12-hour LOS threshold. Each figure represents a patient moving through the system in real time across all three campuses.

How to explore the full interactive animation:

1. Click the button below to open the simulation in a new tab.
2. Use the play/pause control at the bottom to start the animation.
3. Hover over any patient figure to see their current state and wait time.
4. Watch how the queue lengths at each campus change over time.

30-second preview · La Jolla 12-hour transfer policy

Explore the Full Interactive Animation →

Where Is the System Breaking Down?

La Jolla — average daily bed utilization over the 6-month baseline simulation.

East Campus — average daily bed utilization over the 6-month baseline simulation.

Proportion of simulation time each resource spent above 80% utilization. La Jolla and East Campus are the most chronically constrained. ED resources across all campuses are near zero — the bottleneck is inpatient capacity, not the ED itself.

• La Jolla's floor unit exceeded 80% utilization 73% of the time; its ICU 64%, boarding 59%.
• East Campus ICU exceeded 80% utilization 63% of the time, despite having only 8 beds.
• East Campus ICU produced an average wait time of ~220 hours — catastrophic queueing when a tiny resource reaches capacity.
• La Jolla's boarding unit saw a non-zero wait in 89% of all queue events.
• Hillcrest spent comparatively little time above 80% — suggesting meaningful available capacity.
• The binding constraint is inpatient beds, not the ED.

Average wait time per resource (hours) among all queue events in the baseline simulation. East Campus ICU stands out dramatically — patients waiting for one of only 8 beds face severe delays once that resource fills.

Transfer Policies: What Actually Helps?

We tested whether routing long-stay patients away from their admitting campus at the moment of ED admission could reduce these bottlenecks. The answer depends critically on which campus the transfers originate from.

La Jolla boarding unit and ICU utilization under three LOS-based transfer thresholds vs. baseline. All three thresholds eliminate boarding congestion and meaningfully reduce ICU pressure.

Sensitivity of mean wait time to LOS transfer threshold, by source campus. Transferring from La Jolla reliably reduces system-wide congestion. Transferring from Hillcrest can overwhelm East Campus ICU under aggressive thresholds — a non-obvious load-shifting effect that shows why transfer policy design requires system-wide thinking.

How to Interpret These Results

The utilization percentages and wait times here are simulation outputs, not live measurements from UCSD Health operations. The most meaningful signal is comparative: how does a transfer policy perform relative to the no-transfer baseline, and does that pattern hold across multiple threshold conditions? Consistent patterns across thresholds are more robust findings than any single number.

Limitations

• Each source campus experiment was run in isolation — real-world deployment would involve simultaneous decisions across all campuses.
• The model does not include staffing constraints, clinical contraindications, or patient consent requirements.
• The LOS predictor uses only three features; its wide residual spread means predictions are rough guides, not precise forecasts.
• Capacity parameters were derived from typical steady-state occupancy rather than licensed bed counts.

What Comes Next

Three directions stand out for future work: enriching the LOS predictor with clinical features available at admission; modeling simultaneous multi-campus transfer logic to capture system-wide interaction effects; and incorporating real-time capacity checks into the transfer trigger so that transfers are only authorized when the receiving campus has available beds.

The Team

About Us

This project was conducted through the Joan & Irwin Jacobs Center for Health Innovation at UC San Diego as part of the DSC 180B Capstone sequence. Data was provided by UC San Diego Health. Simulation was implemented in Python using SimPy and vidigi.

Prior Work & Acknowledgements

This project builds on a body of prior DES research in hospital operations. Key works that informed our design and methodology:

• Day et al. (2012) — Demonstrated that adding a healthcare provider to ED triage could reduce mean length of stay by 15%; validated using DES. Emergency Medicine Journal.
• Ahmed & Alkhamis (2009) — Used simulation-optimization to show that strategic resource allocation at patient entry can cut wait times by 40% without added cost. European Journal of Operational Research.
• Kuo et al. (2014) — Showed that fitting LOS distributions from EHR data (rather than assumptions) meaningfully improves DES fidelity. Flexible Services and Manufacturing Journal.
• Boyle et al. (2021) — Presented a framework for generalisable, data-driven ED simulation models. European Journal of Operational Research.
• Paul et al. (2010) — Systematic review emphasizing that poorly calibrated simulation models can lead to misleading policy recommendations. Simulation.