Introduction
NHS Scotland's Information Services Division first developed SPARRA (Scottish Patients at Risk of Readmission) in 2006 within a policy context, which required shifting from a healthcare system geared towards hospital-based treatment to a system founded on a preventative, anticipatory approach to the management of long term conditions.
In 2018, NHS Scotland ISD began a collaboration with The Alan Turing Institute to update and extended this model using state of the art machine learning and artificial intelligence methods. Once developed, this model will produce a score every month for 80% of the Scottish population (about 3.6 million patients), assessing their risk of emergency admission to hospital in the following 12 month period. This score is delivered automatically direct to GP surgeries and can potentially be utilised when informing primary care intervention.
Explaining the science
When people need to visit an emergency department, their health has broken down in some way. Keeping these people healthy now requires treatment-based rather than prevention-based healthcare. Avoiding these breakdowns of health is in the interest of individual patients and the overall efficiency of the healthcare system.
If people are at a high risk of health breakdown, it is often possible for GPs to intervene to reduce this risk: for instance, by increasing frequency of appointments, increasing, decreasing or changing medications, or making earlier referrals to secondary healthcare services. However, these interventions need to be made judiciously, and it is very difficult for busy GPs to make the consistent and accurate assessments of health breakdown risk necessary to decide how much to intervene for each patient.
Factors contributing to an individual patient's risk include previous interactions with the healthcare system, medications and long-term conditions, and demographics. These factors interact in complex ways: research has shown that a simple sum of risk factors is not optimally effective in predicting risk. Both lookup and synthesis of risk factors is difficult for GPs, but relatively easy for computers. The SPARRA risk score bridges the gap between risk factors and usable risk estimates.
Project aims
The next generation of SPARRA is being developed to improve predictive performance, whilst ensuring reproducibility and scalability of the entire analytical workflow. This involves both a careful application of the latest machine learning and artificial intelligence methods, as well as production of a reproducible data science environment based around Docker, which can be deployed into NHS safe haven environments and carried over to NHS National Services Scotland deployment arenas.
Initial results show substantial improvements to both precision and calibration of the model, utilising a super learner of a collection of machine learning models. This means both that more patients identified at risk are correctly flagged and that the judgement of the level of that risk is more accurate, potentially enabling better health outcomes through more accurately informed GPs.
Applications
NHS Scotland ISD will be deploying this model, complete with reproducible computational environment, directly into production to deliver risk scores to GPs on a monthly basis. Most importantly this has the potential to improve health outcomes through more targeted primary care intervention, with a secondary benefit of reducing pressure on NHS Scotland hospital admissions and attendant cost savings.
In future iterations of this project, objectives include careful development of interpretable scoring so that GPs have on hand not just a risk score, but an understandable assessment of why that risk score arises. This has the potential to act as a broader support tool aiding GPs in speeding up decision making and diagnosis about primary care interventions they deem appropriate.
Recent updates
March 2020
The new SPARRA model is in final testing and the reproducible data science Docker environment preparing for safe haven security audits. The model should go into production use in the coming months.
Organisers
Researchers and collaborators
Dr James Liley
Postdoctoral Fellow, University of Edinburgh
Samuel Emerson
Doctoral Student, Durham University
Bilal Mateen
Clinical Data Science Fellow
