A "Leren en Beslissen" (Learning & Deciding) project at the University of Amsterdam for Stokhos Emergency Mathematics: predicting where and when ambulance incidents occur so emergency vehicles can be positioned preemptively and response times cut.

Abstract

It is impossible to know precisely where accidents will happen in the future, but predictions can be made from historical data. Stokhos assists local ambulance departments in predicting where and when accidents occur and how much resource (personnel and vehicles) is needed. This project builds a model to find cyclical patterns in historic data spanning two years (15 May 2018 – 10 May 2020, 84,456 ambulance rides with time of arrival, position, and ride type). Using cyclical regression, Random Forest, and LSTM models, the work shows a significant cyclical pattern exists; the final LSTM reaches a mean absolute error of 10, with room to improve using more data and Random-Forest ensembling.

Problem

Stokhos advises on optimal emergency-vehicle coverage per region. Existing software distributes accidents over a whole year with no shorter timespan. The challenge: predict where accidents occur as accurately as possible over a shorter timeframe than a year, so preemptive placement of emergency vehicles improves response times and ultimately saves lives.

Data & method

Two datasets from the Zuid-Holland-Zuid safety region cover calls to the Emergency Medical Call Center, including a subjective patient-acuity decision by the ambulance crew. Naively dropping NaN records would lose ~80% of the 84,456 rows, so the pipeline began with careful raw-data analysis, feature removal/engineering, and NaN handling. The region was modelled as a grid map and fed to LSTM/RNN sequence models and Random Forest, with cyclical-pattern evaluation over multiple time spans.

Materials