Emergency admissions (EA) to hospital continue to rise at a time when NHS budgets are under significant pressure, often with long-term conditions where timely and effective primary care could reduce the likelihood of emergency care.

NHS Scotland first developed SPARRA (Scottish Patients at Risk of Readmission) in 2006 to predict the risk of EAs. James Liley and the team at Public Health Scotland / The Alan Turing Institute have updated the existing model, to develop its accuracy in predicting the risk of EAs, and the level of that risk, in the upcoming year for 80% of the population. These predictions will be automatically shared with GPs to allow them to manage their high-risk patients and prevent admissions to hospital.


EAs represent a breakdown of health and a transition from preventative to curative medicine. Avoiding emergency hospital admissions is a major concern, not only because of the detrimental impact on an individual’s health, but the high cost of emergency admissions compared with other forms of care.


Analysing data extracted from Electronic Health Records (EHRs), including previous hospital admissions, medications, long-term conditions, and demographics – the team developed a model to predict the risk of an EA for the following year for 80% of Scotland’s population.

The model uses multiple contemporary machine-learning methods, each with different strengths, to provide substantial insights into the epidemiology underlying EAs. The predictions will be delivered directly to Scottish GPs, enabling them to take action for high-risk patients who are likely to have an EA, such as by increasing appointments or reviewing their medication.


The team’s model outperforms the existing model and is on its way to being deployed across primary care. Their work allows unprecedented insight into the epidemiology of EAs in Scotland, and provides a clinically-useful predictor of health status breakdown. The model found that deprivation has a severe effect on EA risk; the most deprived 10% of individuals in Scotland had an EA risk equivalent to up to 40 extra years of age when compared to the least deprived 10%. This helps doctors in Scotland to make clinical decisions, and demonstrates the potential for the use of modern computing in medical practice.

“A common problem in health data science is just using one predictor model. We used an array of different models, including linear and tree-like structures, each having different strengths, and combined them to make the most accurate result we can. Our work is a rare example of a model that has become a tool to help support GPs, with the understanding that professional judgements are critical when interpreting our predictions.”

This project is supported by Public Health Scotland and funded by a HDR UK grant.