Patients at highest risk of a quick return to emergency departments can be spotted and supported using machine learning, according to new research.
The insight could be used to help reduce reattendance rates, easing pressure on emergency departments.
Southampton researchers trained their machine learning model on thousands of visits to the emergency department at University Hospital Southampton.
It is hoped the approach can now be tested on datasets from more NHS Trusts. It could then begin to support decision-making processes in emergency departments across the UK.
Higher and lower risk reattendance
The use of emergency departments has been growing steadily over the last decade. This has contributed to increased overcrowding and extended waiting times.
The Southampton research aims to curb this trend by helping reduce the number of short-term reattendances. This is where a patient returns to an emergency department within 72 hours of being discharged.
The initial results show that machine-learning models can be used to classify patients, with moderate performance, into low and high-risk groups.
Scientists have published their findings in Scientific Reportsexternal link.
Dr Francis Chmiel, lead author, says: “This model could serve as a check prior to a patient exiting from the emergency department or earlier in the process to identify possible care actions. The explanation can be extremely valuable by indicating why the patient is likely come back to the hospital, generating pathways to prevent that outcome.
“Our hope is that models like this one can be used to reduce pressure on our NHS, to personalise and therefore enhance the care of our NHS patients. This will support the health of both our local population in Southampton, but nationally as well.”
Learning through real-word data
The new research combined expertise from the University of Southampton and its IT Innovation Centreexternal link with University Hospital Southampton and the NIHR Southampton Biomedical Research Centreexternal link.
Computer scientists built a gradient boosting decision tree model. It was trained using 35,447 attendances by 28,945 patients in Southampton. This was then evaluated on a hold-out test set featuring 8,847 attendances by 7,237 patients.
Dr Dan Burns, a Research Engineer at the IT Innovation Centre within the School of Electronics and Computer Science, says: “We extracted a number of variables such as the number of times a person had visited the emergency department in the last 30 days, lifestyle factors such as whether they live alone or have a history of smoking, and the number of conditions on their health record. We used these insights to form an algorithm that could predict whether a patient would be readmitted within 72 hours.
“More importantly, in health, we are often seeking explanations behind why the algorithm produced a given prediction. We therefore also applied an explanation algorithm known as SHAP (Shapley Additive exPlanations) to generate a way of explaining which variables contributed most to a given decision by the algorithm for each individual prediction. These were grouped into explanation ‘scores’ that were clustered to be more communicable to clinicians.”
Converting risks into a decision support tool
Prior research has shown several factors are suggestive of short-term reattendance risk. These include living alone, depression, initial diagnosis and historical emergency department usage.
In their exploratory analysis, the Southampton team found that higher reattendance rates were observed for incidents during the night. They also found that certain complaints, such as abdominal pain, had a higher reattendance rate than the hospital average.
Knowledge of these risk factors is important to clinical staff, but more support could be available in the future optimise decision-making before discharge.
Comments