Case study · 04
AI hospital routing
Patient routing that weighs specialty fit and ED crowding, not just distance.
- Role
- Georgetown MHSA capstone
- Status
- Completed May 2026
- Read
- 1 min
- Stack
- PythonRSwiftStatistics
d = 0.64
effect size vs closest-hospital routing
Problem
Ambulances usually go to the closest hospital, even when a farther one has the right specialty (stroke, cardiac) or a less crowded ED.
Approach
- 01Two separate layers. A decision utility picks the hospital: a trauma-score-based severity multiplier and specialty match, minus transport time and ED delay.
- 02A separate risk score (specialty value minus transport and ED delay) is used only afterwards, to compare AI routing against closest-hospital routing on equal terms.
- 03Ran both policies over 500 simulated cases and tested the difference.
- 04Analyzed the results in R and Python, with a Swift iOS demo.
The closest hospital isn't always the right one.
Results
| Before | 109.09 closest |
|---|---|
| After | 112.99 AI |
| AI better | 216 of 244 |
|---|---|
| Rest | AI worse (28) |
- Mean score 112.99 vs 109.09, a gain of 3.91 points (about 3.6%). t(499) = 15.59, p < 0.001, Cohen's d = 0.64.
- The two policies chose a different hospital in 244 cases. AI was better in 216 of them (88.5%) and worse in 28 (5.6%).
- Example: a stroke patient routed to a stroke center about 7 minutes farther scored 138.4 vs 124.6.
What I'd do next
- Validate against real dispatch and hospital-capacity data.
Simulated cases, not real patients.