Professor Christopher Pearce and his team at Outcome Health have developed a machine learning algorithm to calculate a patient’s risk of an Emergency Department (ED) presentation, based on general practice (GP) patient information and ED admissions data.

ED presentations

Today, more people present to Australia’s public hospital EDs than ever before. In 2016-17, there were 7.8 million ED presentations nationwide – more than 21,000 each day – compared to 6.7 million in 2012-13.[i] Australia spends $5 billion on hospital based care every year[ii], but there is no accurate or clear way of predicting who will go to ED, or understanding how to prevent those visits from happening.

EDs are a critical part of Australia’s health care system, but many of these costly admissions could be avoided if the right treatment is provided to patients in the primary care setting, by a local doctor or GP. GPs in Australia face a challenge though, as they don’t always have access to a clear summary of a patient’s full medical history, so identifying those at risk of imminent hospital presentation and admission can be difficult. A lack of shared information and unified patient records across the health care system contributes significantly to this challenge.

Tackling the problem

Previously designed predictive risk tools in this space have focused predominantly on patients at a high risk of ED presentation, based on a small number of chronic diseases. But now, thanks to HCF Research Foundation funding, Professor Christopher Pearce and his team at Outcome Health have created a predictive model that calculates the risk of a patient attending ED in the next 30 days after a GP visit. The model uses a machine learning* algorithm that considers thousands of patient variables to build the risk score. The team found that by passing more data through the machine learning, more information is able to be linked back to a potential ED presentation, and the more accurate the algorithm becomes.

The aim of the research was to use GP health record data in order to develop a tool that highlights to a doctor those patients with a higher risk of an ED presentation, at the time of an appointment. Doctors can then take relevant action in terms of different treatment options and early intervention, to help patients avoid the ED presentation and reduce unnecessary visits nationally. This in turn saves the system money and prevents illness before it becomes more serious or complicated – meaning faster and more appropriate care and better health outcomes for Australians.

Importantly, the POLAR Diversion tool is not intended to replace regular patient care, but can act as a support from which GPs can take action. More than just a predictive model, the tool can help to better inform and support the decisions GPs make.

Key algorithm results

Doctors who have evaluated the tool using their own patients have so far provided positive feedback. Many felt it was a good alert tool that could be incorporated into existing GP software alert systems, as it’s also accompanied by easy to use patient summaries which describe the factors behind an alert being triggered. Factors such as patient demographics, diagnoses, medications, pathology results and other patient measures are all weighted within the tool and shown back to the GP for their review.

On the whole, GPs felt the tool performed well and tended to agree with most of the results, but especially for older patients and those with lower risk scores. This successful algorithm has the potential to be used in general practices across Australia. The next challenge lies in developing clinicians’ trust in the tool, to maximise its potential use and effectiveness in influencing the provision of health care in Australia.

• Within 30 days of a patient visiting the GP, the algorithm correctly predicted 73.7% of ED admissions.
• Of patients who did not present to ED, 82.3% were correctly identified as not presenting to ED at any point in the year after the patient’s last recorded GP visit.

[i] Australian Institute of Health and Welfare, Emergency Department Care 2016-17: Australian hospital statistics
https://www.aihw.gov.au/reports/hospitals/emergency-department-care-ahs-2016-17/contents/summary

[ii] Analysis of 2014-15 Health Budget: Unfair and Unhealthy
https://ses.library.usyd.edu.au/bitstream/2123/11981/1/2014-15healthbudget.pdf

*Machine learning is a type of artificial intelligence that provides computer programs with the ability to learn without being explicitly programmed, and change when exposed to new data.