Respository for storing the OSM OHCA location prediction model and deployment pipeline

This repository contains the code for the OSM OHCA location prediction model. The model uses H3 to create a hexagonal grid over training data and uses a H2O AutoML regressor to predict the count of OHCA incidents in provided hexagons.

Preview 😊

• Green hexagons - hexagons with an AED or a hospital inside
• Blue hexagons - top 10 hexagons with the highest predicted number of OHCA incidents withouth an AED
• Red hexagons - other hexagons without an AED

All hexagons have an opacity set to the predicted number of OHCA incidents in the hexagon.

Prerequisites for running the deployment pipeline

• Python 3.11
• Java 8 or higher

For building neighbourer, the program that adds neighbour data to hexagons, you will also need:

• H3 version 4.10 (C library)
• CMake

Running the deployment pipeline

Install python dependencies:

pip install -r requirements.txt

Compile neighbourer:

cd neighbourer
cmake CMakeLists.txt
make

The deployment goes through the following steps:

1. aquire_data.py - Downloads and saves case data as two files - osm_data.csv with data about the terrain features and warszawa_osm.csv into the data directory from:

• The Overpass API for Data about terrain features
• VBOHCA - Spatiotemporal Data Set for Out-of-Hospital Cardiac Arrests in Virginia Beach
• Cincinnati Fire Incidents (CAD) (including EMS: ALS/BLS)
• Police Dispatched Incidents, Montgomery County, MD

2. create_main_df.py - Creates two dataframes, one for training and one for testing, and saves them into the data directory, their names are main_hexagon_df.csv and target.csv. It also uses the neighbourer - a program that sums values of neighbours for every hexagon and adds appropriate columns. The program is written in C++ and uses the H3 library.

3. train_model.py - Trains the model and saves it into the models directory

4. predict.py - Uses the target.csv file to predict the count of OHCA incidents in each hexagon and saves the result into the data directory as predictions.csv

5. visual.py - Creates a map of the predictions and saves it into the data directory as map.html

Performance

The best model currently achieves a root mean squared error of 0.91 on predicting the number of ohca cases in a hexagon across 3 years on the whole dataset.

Special thanks

We would like to send special thanks to people and organizations that have shared data that powers our model 🤗. We also encourage other local goverments to share their data in accessible ways for research and open source projects!

• The city of Virginia Beach and the creators of VBOHCA
• The city of Cincinnati
• Montgomery County
• The County of Berks

Authors

• Radosław Dąbkowski
• Michał Ambroży
• Igor Witczak
• Koło Naukowe Geoinformatyki