2024-05-07 10:54:23
London – A machine learning algorithm for predicting hypoglycemia in hospital using only capillary blood glucose (CBG) levels works great, according to study data.
In further analysis, researchers used the model to assess the relative importance of different glycemic characteristics in predicting steady-state hypoglycemia. They found that extreme and variable CBG measurements have the greatest prognostic value.
Dr Chris Sainsbury, consultant in diabetes and endocrinology at Gartnavel General Hospital in Glasgow, Scotland, presented all the results at this year’s Diabetes UK Professional Conference (DUKPC) 2024 with his colleagues Drs. Greg Jones, diabetes consultant, and Dr. Deborah Morrison, general practitioner with specialist training in diabetes, introduced [1].
“We have shown that the model has very good predictive power for hypoglycaemic events,” said Sainsbury Medscape. The next step is to demonstrate that hospital staff can respond effectively to these warnings to avoid hypoglycaemia before it becomes clinically relevant.
We have shown that the model has very good predictive power for hypoglycaemic events. Dr. Chris Sainsbury
Particularly high risk of hypoglycaemia on the 7th hospital day
The accuracy of the model for predicting a hypoglycemic event was assessed using the area under the receiver operating characteristic (ROC) curve. The model calculates a number between 0 and 1 that predicts the risk of a hypoglycemic event during the next 24 hours and reaches a maximum on average on the 7th day of the hospital stay. The value increased from approximately 0.78 on day 2 to 0.85 on day 7 and then remained at approximately 0.85 until day 31.
The model is designed for use with hospitalized patients. It only works with CBG readings. In comparison, other models rely on a range of clinical data obtained from electronic hospital records. This complexity greatly limits the usefulness and transferability of analyzes between different hospitals.
Hypoglycaemia – a challenge for hospital teams
“Inpatient hypoglycaemia is a major problem in people with diabetes, particularly in patients receiving insulin or sulphonylureas. “During an overnight stay, eating habits sometimes change, which can lead to hypoglycaemia,” explains Sainsbury.
During a 24-hour stay, eating habits sometimes change, which can lead to hypoglycaemia. Dr. Chris Sainsbury
Hospitalized hypoglycemic events are associated with higher morbidity and mortality. It was not just an “inconvenience”, as doctors would have previously believed, but “an event that we really need to prevent,” Sainsbury stressed. He pointed out that regarding one in four beds in his hospital were regularly occupied by diabetic patients.
“If you’re on a diabetes unit and you have hypoglycaemia, the staff know what to do,” Sainsbury stressed. “But many are in other departments where such incidents are less frequent.”
Detect patients at risk with machine learning
That is why Sainsbury and his colleagues want to be able to identify patients at high risk of a hypoglycaemic event and treat them preventively. They have developed an artificial intelligence (AI) algorithm for this.
When training their model, the researchers first had to consider whether to use CBG data alone or in combination with other clinical data. They chose stationary CBG measurements for reasons of portability, but also because they are transferred wirelessly to patient records in real time.
They retrospectively analyzed data from 259,274 patients collected by the National Health Service (NHS) in the Greater Glasgow and Clyde region between 2009 and 2022. This resulted in almost 5 million CBG records. Internal validation was performed using a separate data set of 70,353 patients.
Several versions of XGBoost, a machine learning model, were trained to predict the probability of a hypoglycemic event, defined as a CBG <4 mmol/L between hospital days 2 and 31.
The researchers analyzed a number of features in the CBG data for each admission day and used them to predict the risk of a hypoglycemic event. These included maximum and minimum CBG levels on the day of admission, the number of tests performed and the patients’ age and gender.
Sainsbury and his colleagues were then able to validate the model, first using internal data from their hospital and then data from a similar group of patients in Edinburgh. They found that the algorithm was transferable between different clinics.
Identify the most meaningful features
The researchers report further results of the study in a second poster at the congress. Using the same data and model, they analyzed which specific features of the CBG data were most useful in predicting hypoglycemia.
“We should understand which particular characteristics predict the risk of hypoglycaemia, as this may be a way to prevent hypoglycaemia in the future,” explained Sainsbury.
Age was found to be the strongest predictor of hypoglycemia on day 1. On day 7, the most important characteristics were:
-
the lowest CBG value in the 48 hours before hypoglycemia,
-
the glycemic variability and
-
the trend for blood sugar in the last 24 hours.
“But this poster also shows that aspects that are important at the start of a patient’s hospital stay become less important as the patient progresses,” Sainsbury said. That’s understandable. “When a patient is first admitted, they are usually very ill and following a few days… they should start to recover. The patient’s physiology will change during this time. CBG levels and associated characteristics will also reflect this and affect additional risk differently.”
From research to application
If the AI detects impending hypoglycaemia, it will warn staff later in the practice – and measures can be taken well in advance of the event. The researchers plan to test their model first on diabetes wards and later on other wards to evaluate its practical application. Sainsbury wants to compare usage in 3 hospitals with different levels of intervention.
“We hope that we can then make targeted changes based on clinical protocols – for example, the dose of a sulphonylurea or an insulin can be reduced,” Sainsbury said.
In the longer term, the team wants to develop a “just-in-time” training, delivered via a smartphone app or other resources. “An employee will not only receive the hypoglycemia forecast, but also instructions on how to minimize the risk of this event,” the expert emphasized.
Optimize the management of diabetes patients
Gerry Rayman, consultant diabetologist at Ipswich Hospital and head of the NHS Getting It Right First Time programme, welcomed possible improvements in the management of hypoglycaemia in inpatient care.
“Approximately 1 in 5 diabetics experience mild hypoglycemia, and 1 in 50 experience hypoglycemia severe enough to require acute injectable emergency treatment. Hospitalized hypoglycemia is associated with longer length of stay and higher mortality. It is therefore important to prevent this unpleasant and often frightening complication , Rayman said Medscape.
About 1 in 5 diabetics experience mild hypoglycemia, and 1 in 50 experience hypoglycemia severe enough to require immediate injectable emergency treatment. Gerry Rayman
“The ability to predict at-risk individuals at an early stage based only on capillary blood glucose levels will be very valuable to clinical staff. We look forward to seeing more data on outcomes as this innovative tool is used in practice,” said Rayman.
The ability to predict at-risk individuals at an early stage based only on capillary blood glucose levels will be very valuable to clinicians. Gerry Rayman
This article was originally published on Medscape.com . As part of the translation process, our editors may also use text editing software including AI.
1715080486
#model #predicts #hypoglycaemia #hospital
