Machine learning to help hospitals predict Covid-19 demand

A machine learning (ML) system is being trialled in hospitals in England to manage upcoming demand for beds and ventilators in intensive care units (ICUs) in the Covid-19 coronavirus outbreak.

The Covid-19 Capacity Planning and Analysis System (CPAS) was developed by NHS Digital data scientists and a team of researchers from the University of Cambridge, and uses data from Public Health England (PHE). It goes live this week in four hospitals with the aim to predict and plan resources needed to treat critically ill patients.

As well as the ability to forecast demand, the system is intended to help the wider NHS in ensuring that ventilators, other equipment and drugs that ICUs will need are in place when they are required, according to NHS Digital’s chief medical officer, Jonathan Benger.

“With the pressure being placed on intensive care by the current coronavirus pandemic, it is essential to be able to predict demand for critical care beds, equipment and staff,” he said.

By using data on Covid-19 patients who have been treated in hospitals, the new system has the ability to predict how many patients may require admission to an ICU, as well as how many may require ventilators and how long patients are likely to be in hospital or ICU.

According to NHS Digital, these collated sets of data used by the system can support hospitals in planning required resources, make up shortfalls, or share any excess capacity with other hospitals.

A machine learning engine called Cambridge Adjutorium developed by medical researchers at Cambridge, which had been previously used to develop insights into cardiovascular disease and cystic fibrosis, has been used as the foundation for CPAS.

Two weeks ago, the university found it was possible to do capacity planning for Covid-19 patients, and NHS Digital identified the opportunity to “industrialise the methods” and deploy the data-driven planning system through the national infrastructure that it manages.

To adjust the system to the pandemic, the Cambridge team worked with NHS Digital, initially using depersonalised data collected by PHE’s Covid-19 Hospitalisation in England Surveillance System (Chess) system. The data was used to train the Cambridge Adjutorium model and successfully demonstrated that it could predict hospital resource use accurately.

“Although the system uses data from individuals to build its models, the system does not make treatment decisions about individual patients – rather, by aggregating that data we can make more accurate predictions about larger groups, at the level of a hospital, a trust, a region or nationally," said Mihaela van der Schaar, the Cambridge professor who leads the team that developed the system.

“So while we can say with a high level of confidence that 30 out of 40 ITU beds in a hospital will be occupied next week, we are not trying to predict which patients will be in them,” she added.

The CPAS tool encompass three elements. As well as forecasting demand for resources, the capability includes the provision of statistics, so it is possible to get a demographic picture of the population of patients being admitted to ICUs, their additional medical conditions, and compare that data with regional and national information.

CPAS also provides a “simulation environment” which allows hospitals to test the effect of alternative scenarios, such as increasing the number of available beds or changes in the profile of patients admitted.

The goal is to successfully complete the trial, then evolve the robustness of the system and improve its capability and accuracy by integrating a wider range of data collected by NHS Digital alongside the Chess data.

According to NHS Digital, a number of other countries have already expressed interested in the ML-based system as a means to handle healthcare resources in the coronavirus pandemic.

If the CPAS trials prove successful, NHS Digital wants to evolve the system into a broader framework for hospital resource management. It is hoped that, after the pandemic, the platform can be used to predict hospital length of hospital stay, discharge planning and wider intensive care demand.