The Turing’s partnership with NHS Scotland hit its stride in the last 12 months bringing the latest advances in data science and machine learning to bear on the complex challenge of predicting who will require emergency hospital admissions. If someone is known to be at a high risk of a health breakdown and emergency admission, GPs can often intervene to reduce this risk by, say, increasing appointments, adjusting medication or making targeted referrals.
Such an anticipatory approach to the management of long-term conditions makes sense: an ounce of prevention is worth a pound of cure. But it is difficult for busy GPs to make the accurate assessments of a given patient’s risk of emergency admission. Scottish Patients At Risk of Readmission and Admission (SPARRA) is a tool created in 2006 that predicts an individual’s risk of emergency admission within 12 months. The current iteration, SPARRA version 3 (v3), has been in place since 2012, with risk scores now based on a wide range of hospital-based patient data and calculated for all Scottish residents who have had some hospital or prescribing activity in the previous three years.
How did it start?
“Public Health Scotland bring their knowledge of health and care, and the data assets we have available, and the Turing bring in new data science expertise to form a really strong collaboration which will make a real difference to the way data science informs and influences the delivery of health and care services in Scotland.”
Scott Heald, Associate Director and Head of Profession for Stats Public Health Scotland
More than four million people have a SPARRA score at any given time, and this information is provided monthly to their GPs. It was late 2017 when the Turing first collaborated with NHS Scotland with a successful Data Study Group. In 2019, with a deeper partnership firmly established, the collaboration kicked into a high gear and in early 2020, SPARRA v4 was completed and is now ready for deployment across Scotland. It transforms SPARRA v3 with the addition of a cutting-edge suite of machine learning modelling techniques, including random forest, gradient boosting and neural networks.
These models all sit within a ‘superlearner’ algorithm, which learns to make predictions based on the entire suite of technologies. Feeding into these models is a new engineered set of features generated by a latent Dirichlet allocation model, which automatically groups long-term condition diagnoses and pharmaceutical prescriptions into related ‘topics’ and scores each observation against them. All of the computation is performed securely within NHS data safe havens to preserve patient privacy.
Tests on historic data illustrate that SPARRA v4 has the potential to pre-empt significantly more readmissions. In fact, when SPARRA v4 is compared with v3, 1,000 additional emergency admissions (of the 10,000 patients deemed most at risk) could have been pre-empted. What is more, the calibration of the SPARRA tool has now improved remarkably. In SPARRA v3, of the 10,000 patients deemed most at risk of emergency admission, the difference between the model’s overall predicted number of admissions and the reality was about 1,000, a significant gap.
“In SPARRA v4, even for the 30,000 patients deemed most at risk of emergency admission, the difference between the overall prediction and reality was less than 100. A remarkable improvement in accuracy which will give healthcare providers increased confidence in the tool,” said the project’s Principal Investigator, Turing Health and Medical Sciences Programme Fellow Louis Aslett of Durham University. The plan was to deploy v4 in early 2020, although this has been delayed while NHS Scotland deals with the more immediate concerns of COVID-19.
What does the future hold?
As well as improving the lives of more Scottish citizens, the enhanced SPARRA system should deliver a substantial secondary benefit of reducing pressure on Scottish hospital admissions, and the cost savings that go with that. The SPARRA model will continue to be refined during 2020 and future iterations of this project will include clinically relevant context to support the risk scores, improving interpretability for GPs. Greater interpretability will enable GPs to assess better which patients at high risk of hospital admission will respond best to primary care intervention, keeping them out of hospital.