patientflow: predicting short-term hospital bed demand using real-time data – The Healthcare Service Analytics & Decision Intelligence Atlas

Summary

PatientFlow is a Python package for real-time short-term predictions of hospital bed demand. It uses snapshots of current in-hospital patients as well as expected future arrivals, converts patient-level predictions into aggregated group-level bed-count forecasts, and delivers actionable outputs for bed managers, such as a likely range of the number of beds needed in the next 8 hours for a given specialty. The package is designed for operational, short-term use as opposed to long-term strategic planning.

Originally developed for UCLH’s emergency admissions team, PatientFlow uses real-time or near real-time Electronic Health Record (EHR) data to convert individual-patient probabilities of admission, discharge or transfer into aggregated forecasts of bed occupancy or demand.

By training a machine learning model (specifically, an XGBoost model) on snapshots of unfinished patient visits, using this data to predict whether patients will be admitted/discharged/transferred in that time period, this model can be used to make predictions on new data.

Source: https://www.youtube.com/watch?v=ha_zckz3_rU

Talks

This first of a series of two talks covers the theory of the work.

This next talk dives more into the details of the package.

Project status

Status: Silver

Rationale: This work currently describes itself as an alpha release. It has been used by UCLH and a hospital in Sweden, but it is not in widespread use beyond that, so may still require modification for your own use case.