Available at http://ezemsplanner.com/.
ezEMS Planner was developed as a Hackbright Academy four-week project. It uses publicly available data from DataSF to predict the ambulance arrival time at a given address and to calculate statistics about the ambulance system in SF.
The prediction is based on a random forest regression model fitted to data recorded between 2000 and 2019. Of a total of approximately 5.1 million records in the database, 3.3 million are listed as medical incidents and they required a total of 1.7 million ambulance units being dispatched—1.4 million from the fire department and 300,000 million private units. The app uses this filtered dataset of 1.7 million entries recorded between 2000–2019.
All the data was stored in a PostgreSQL database, whose tables are depicted in the diagram below:
The following features were extracted from the information in the database and later used to fit the random forest regression model:
Temporal Features
Year
Day of Year_sin
Day of Year_cos
Day of Week_sin
Day of Week_cos
Hour_sin
Hour_cos
is_Weekend
(boolean)is_Holiday
(boolean)
Spatial Features
Latitude
Longitude
Tract
(GEOID10 from 2010 US Census)Nearest Hospital
(distance)Nearest Fire Station
(distance)
Dispatcher Input
Original Priority
(3, 2, or E)Unit Type
(private or fire department)
Some of the temporal features above were decomposed into sine and cosine components, to account for their circularity (e.g., 23:00 is closer to midnight than to 21:00, but Python would have no way of knowing this if the time was simply converted to a number, because 23 is closer to 21 than it is to 0.)
The data was split into 75% training data and 25% testing data, and a grid search was conducted over a set of 176 combinations of random forest regression model parameters (11 n_estimators
values between 50 and 300, 8 min_samples_split
values between 2 and 400, and 2 max_features
values of "log2" and "auto"). The best-fitting model had n_estimators = 175
, min_samples_split = 200
, and max_features = "log2"
.
The graph above shows the importance of the features in the best-fitting model. Input features related to the same human-interpretable parameter were combined (e.g., Hour_cos
and Hour_sin
).
The uncertainty on the predicted ambulance arrival time is approximately 3.5 minutes, which is an improvement of 24% over a naive guess. Additional features, such as live travel time to the incident, ambulance post locations, commuter-adjusted population estimates, and homelessness data would likely improve the accuracy of the prediction.