Researchers launch forecasting challenge to help predict severe patient harm in NHS hospitals

Researchers have launched a new data science challenge aimed at improving the ability of NHS hospitals to anticipate and prevent severe patient harm.

"The dedication of the NHS to finding new, cost-effective ways to improve patient care is really inspiring. They are working so hard, and it's wonderful to help them find new ways to improve patient outcomes and care." Dr Will Pearse

The SPHERE-PPL NHS Severe Patient Harm Forecasting Contest invites researchers and data scientists from around the world to develop an algorithm that can forecast the number of estimated avoidable deaths in hospitals up to ten days in advance. The competition was put together by Dr Will Pearse and Alex Rabeau in the Department of Life Sciences.

By providing early warnings of rising risk, the project aims to help hospital managers take proactive steps to protect patients.

Dr Pearse said: ‘I've really enjoyed working with the NHS on this project. Their dedication to finding new, cost-effective ways to improve patient care is really inspiring. The NHS are working so hard, and it's wonderful to help them find new ways to improve patient outcomes and care.’

Predicting risk before harm occurs

Pressure in NHS hospitals, particularly in emergency departments, can have serious consequences for patient outcomes.

Evidence suggests that every four hours of delay in emergency department admission is associated with an 8% increase in 30-day mortality risk. Researchers estimate that delays in care may contribute to around 25 potentially avoidable deaths each month.

The new forecasting challenge aims to address this problem by harnessing advances in data science and predictive modelling.

Participants will develop models capable of forecasting the number of estimated avoidable deaths over short-term (one–five day) and medium-term (six–ten day) horizons.

Accurate predictions could allow hospital teams to anticipate periods of elevated risk and take preventative action, such as reallocating resources, adjusting patient flow, or implementing surge planning.

A large-scale healthcare dataset

To support the challenge, researchers have released a dataset combining estimates of avoidable deaths with 220 healthcare system indicators, covering hospital activity and system pressures. This dataset includes only aggregate performance metrics and contains absolutely no patient identifiable information.

The dataset available for candidates' model development covers March 2023 to September 2025. A further dataset from October 2025 will be made available for testing model performance on unseen data, to ensure robust model evaluation.

"We really look forward to seeing the range of algorithms which can be developed for this competition. We are very pleased to partner with Imperial and the Turing Institute to explore how the latest techniques perform on our important real-life problem." Dr Richard Wood NHS Bristol

Dr Richard Wood, Head of Modelling and Analytics, NHS Bristol, North Somerset and South Gloucestershire Integrated Care Board, said: ‘We really look forward to seeing the range of algorithms which can be developed for this competition. The NHS is a complex system and we suspect the metric investigated here is likely determined by an equivalently complex set of interactions across multiple variables and over multiple timeframes. Rather than impose some parametric structure, we believe this sophisticated problem instead lends itself quite naturally to an AI/ML solution, and so we are very pleased to partner with Imperial and the Turing Institute to explore how the latest techniques perform on our important real-life problem.’

Participants are asked to submit forecasting models implemented in R or Python, which will be evaluated using Mean Squared Error (MSE) across the different prediction horizons.

Delivering tools for real-world use in healthcare management

The competition is designed not only to advance forecasting methods, but also to deliver practical tools for healthcare management.

It is intended that the winning model will be integrated into operational use by NHS managers in Bristol, providing daily forecasts to support proactive decision-making.

Researchers hope the initiative will demonstrate how advanced analytics can contribute to improving patient safety across the healthcare system.

Open challenge to the global community

The contest is open to researchers, data scientists, and forecasters from academia, industry, and the public sector.

Final algorithms must be submitted by 5 June, with the final evaluation period concluding on 20 June.

Full details, including the dataset and participation instructions, are available via the project’s GitHub repository.