Previously at my role in the NHS, I established that every four hours of emergency department delay is associated with an 8% increase in the risk of 30-day mortality. Converting this statistical insight into a reliable forecasting tool presents significant technical hurdles. To catalyse the development of high-performance solutions, we launched the SPHERE-PPL NHS Severe Patient Harm Forecasting Contest, an initiative to crowdsource and benchmark predictive models within a complex healthcare system.

This talk explores the practical difficulties of engineering a ‘leak-proof’ time-series competition using real-world data. We discuss the complexities of curating a development dataset of system-wide variables with varying frequencies. We also detail the challenges of anticipating real-world data issues, such as reporting lag and missingness, while implementing a validation framework that mitigates exploitation. We conclude with insights from the winning models and how they are being integrated into the NHS system in Bristol.