Collaborative: Emergency hospital admissions: towards more equitable and actionable risk prediction.

MRC’s Collaborative awards provide students with experience of collaborative research with a non academic partner, enabling the student to spend a period of time with the non-academic partner (usually no less than three months over the lifetime of the PhD).

Background

Emergency admissions (EAs), where a patient needs urgent unplanned in-hospital care, account for over half of all admissions in Scotland [1]. EAs are, on average, longer than elective stays. EAs can also reflect health inequalities and differential healthcare access, as EA risk has been observed to be higher for those living in the most deprived areas [1]. Anticipatory intervention in primary care can prevent avoidable EAs, benefiting patients and the NHS. To achieve this, it is essential that risk stratification tools aid targeting interventions to the right patients and at the right time. 

SPARRA (Scottish Patients at Risk of Readmission and Admission) [2] is a risk score used in Scotland for over a decade to predict EA risk. Individual-level scores help General Practitioners (GPs) to identify patients who may benefit from proactive, preventative care and anticipatory care planning; aggregated scores can help to estimate future demand. Public Health Scotland (PHS) calculates SPARRA monthly for ~80% of Scottish residents. This largely excludes individuals with fewer hospital contacts, leading to population coverage gaps [3] that may compound health inequalities. Moreover, SPARRA predicts overall EA risk regardless of the underlying cause (e.g. asthma exacerbation or appendicitis). This limits actionable information, as not every EA can be predicted [2] or prevented [4]. A key subset is ambulatory care sensitive conditions (ACSC) admissions, those potentially avoidable via primary care intervention. 

Co-developed with PHS, this PhD project has direct translational potential. By characterising population coverage gaps and identifying which ACSC admissions SPARRA can and cannot predict well, we will develop methodological and data drive approaches to address these challenges. Ultimately, our findings will set the foundations for the next iteration of SPARRA, to support more equitable prevention and ease NHS system pressures. 

Aims

The overarching aim for this project is to strengthen SPARRA’s value as a digital tool for proactive care in the NHS, via more equitable and actionable risk prediction. Using national healthcare databases (~5M individuals) held by PHS, we will:

1. Map population coverage gaps. We will compare SPARRA’s current coverage to census estimates, identifying under-represented groups at local and national levels. We will then explore how coverage may be improved, including whether novel data assets (e.g. out-of-hours GP contacts, hospital prescribing) can capture systematically excluded groups. 
2. Characterise the epidemiology of ACSC admissions in Scotland. We will quantify temporal and geographic patterns and the differences between sub-groups (e.g. by deprivation), including how many EAs are potentially avoidable. We will then explore how well SPARRA predicts different ACS admissions, identifying which potentially preventable EAs it can and cannot capture. 
3. Improve ACSC-specific prediction. Building on Aim 2, we will investigate how SPARRA could be adapted to capture ACSC-specific risk. We will then explore how information about ACSC-specific risk and coverage could be used to enhance SPARRA dashboards and data products, providing more actionable insight to GPs and other primary care practitioners.  

Training Outcomes

Supported by a cross-disciplinary supervisory team and hands-on work at PHS, the student will gain an in-depth understanding of an important public health issue and how risk prediction tools are deployed in the NHS. They will develop a broad skillset that will equip them as future leaders in health data science:

• Statistics/AI: risk prediction model development and evaluation, interpretability, scalability. 
• Computing: R programming, dashboard development, reproducibility.
• Health data: electronic health records, phenotyping algorithms, data wrangling.
• Public health: health inequalities, prevention, risk communication. 
• Data governance: use of unconsented patient data, ethical compliance, statistical disclosure.
• Transferable skills: effective oral and written communication, cross-sector collaboration. 
• Patient/public engagement: patient/stakeholder groups, public-facing events e.g. Open Doors Day. 

References: 

[1] Public Health Scotland (2025). Acure hospital activity and NHS beds information, annual report. https://publichealthscotland.scot/publications/acute-hospital-activity-and-nhs-beds-information-annual/acute-hospital-activity-and-nhs-beds-information-annual-year-ending-31-march-2025/

[2] Liley et al (2024). Development and assessment of a machine learning tool for predicting emergency admission in Scotland. npj Digital Medicine. PMID: 39443624

[3] Wallace et al (2016). Reducing emergency admissions through community based interventions. BMJ. PMID: 26822070

[4] Lewis et al (2011). Choosing a predictive risk model: a guide for commissioners in England. https://www.nuffieldtrust.org.uk/resource/choosing-a-predictive-risk-model-a-guide-for-commissioners-in-england