EMS Predictive Relocation — static vs. dynamic, on real 911 data

0% faster

Average response, same fleet & same calls

Static

—

Dynamic

—

Ambulance-minutes saved

—

Within 8-min standard

—

p90 response

—

Median response

—

Fleet modelled

—

Significance —

Data window —

Where to stage right now

Drag to see how the recommended posts shift through the day — each hour has its own real busy-fraction and demand pattern, not one fixed all-day table.

Where to stage the fleet

Low call demand

Medium

High demand

Ranked staging post

Demand & compliance table

Live simulation

Proof & limits

Busy fraction q

—

Coverage standard

—

The model

MEXCLP (Maximum Expected Coverage, Daskin 1983) — the published method real EMS System Status Management runs on. It ranks candidate posts by expected coverage while accounting for the probability that a covering unit is already busy. The output is a compliance table: idle units continuously re-fill it from the top.

Demand forecast (ML, held-out validated)

—

Routing:

Zones by call volume & compliance rank

	Zone	Volume	Calls	Rank

Calls handled

Avg response so far

—

Units busy now

—

Delayed (no idle unit)

Last event

—

Real call arrivals and locations, replayed from the busiest real day in the window. One disclosed assumption: per-call service duration (~50 min average, not present in the source data).

Response-time distribution

Static Dynamic The whole curve shifts left — not just the mean.

Cumulative average responseover the full call stream

Static Dynamic The gap opens early and holds — not a lucky tail.

Ambulance-minutes returned to the city

Cumulative minutes returned

Realism check against measured performance

What this does not claim

This is a simulation on real data, not a fielded deployment. The fleet size and post locations are the model's, not any real agency's, so the comparison is static vs. dynamic under identical conditions — not model vs. a real department. Per-call service time (~50 min) is a disclosed assumption, applied identically to both strategies so it cannot bias the comparison. MEXCLP is a 1983 published method; the contribution here is the end-to-end pipeline on real CAD data with real road routing, and the honest measurement of what it buys.

Point the model at any city

The model is not hard-coded to Seattle or Cincinnati. Drop in a CAD / 911 export with a latitude, a longitude and a timestamp column, and the whole pipeline re-derives itself in your browser: demand zones, fleet size, the MEXCLP compliance table, both simulations and the significance test. No upload, no server — the file never leaves this machine.

Drop a CAD export here, or click to choose a file

CSV with latitude / longitude / timestamp columns (column names are auto-detected).
Try data/cincinnati/calls_60d.csv from this repo to watch it rebuild Cincinnati from scratch.

Head-to-head — — —

Static fixed home posts — how most agencies dispatch today —

Dynamic idle units continuously re-fill the compliance table —

Both lanes clear the same real calls in the same order off one shared clock — a lane advances by spending its own response minutes, so less time-to-patient means further down the track.

Minutes saved so far

—

Predictive EMS Relocation

Point the model at any city

Drop a CAD export here, or click to choose a file