Machine Learning for Real-Time Aggregated Prediction of Hospital Admission for Emergency Patients

Welcome to the repo for real-time prediction of admissions at UCLH.

About the project

The project has a number of elements

1. Extract and clean data
2. Train ML models to predict individual probability of admission
3. Aggregate the individual probabilities into a predicted distribution for number of beds needed
4. Run a ML pipeline in real-time to retrive patients current in ED, make predictions about individual probability of admission, aggregate these into predictions for number of beds needed, save predictions and email them to bed planners
5. Prepare figures for publication in an academic manuscript

See the published version of the manuscript at: https://www.nature.com/articles/s41746-022-00649-y. The authors are Zella King, Joseph Farrington, Martin Utley, Enoch Kung, Samer Elkhodair, Steve Harris, Richard Sekula, Jonathan Gillham, Kezhi Li, and Sonya Crowe. All the code in this repo was written and run by Zella King.

About the repo

This repo is organised into the following folders

• flow-mapping - for extract from Star and initial processing
• predict-admission - scripts for running machine learning and all output files
• real-time - for real-time prediction including input needed for models in real-time

Each folder has its own README to explain what it does

Each folder is structured as follows (but note that data-raw and data-output are ignored by git and model-output and media have been emptied):

• code - contains all scripts used in the final analysis
• data-raw - contains datasets created by the scripts (ignored by git so not visible outside the UCH environment)
• data-output - contains outpout from ML models, feature lists (ignored by git so not visible outside the UCH environment)
• model-output - contains results like predicted probability distributions and evaluation results
• media - output as charts

What to look at first

For the reader interested in the machine learning aspects of this project, the most important files to look at are

• predict-admission/README.md - this will introduce you to the steps involved in the analysis and how we approached the ML
• prediction-admission/generate-timeslices.R - shows the approach to creating timeslices, as reported in our paper
• the various ML scripts in predict-admission folder - run-ML.R (which runs XGBoost), run-RF.R (Random Forest) and run-LR.R (Logistic Regression)

For the reader interested in the aggregration of individual level probabilities into predictions over numbers of beds, and the seven-step pipeline reported in the paper, the most useful files to look at are

• predict-admission/generate-prob-dist-and-pt-estimates-using-survival-analysis.R - this shows the steps involved in implementing the seven-step pipeline (see commented function get-prob-dist() in that file)
• some of the functions in real-time/app/utils.R, particularly those mentioned in predict-admission/README.md