medwireNews: Researchers have developed a machine-learning tool that could predict a type 1 diabetes diagnosis in children on the basis of data from primary care electronic health records.
Speaking at the 2022 Diabetes UK Professional Conference, Julia Townson (Cardiff University, UK) explained that around a quarter of children with type 1 diabetes in the UK are not diagnosed until they are in diabetic ketoacidosis (DKA), with rates unchanged for 25 years despite public health campaigns, highlighting the need for improved tools for early detection.
She said that previous research identified different patterns of primary care contact among children who later go on to develop type 1 diabetes versus those who do not, leading the team to hypothesize that primary care data could be used to flag those likely to be diagnosed with the condition.
To investigate this, Townson and colleagues used a machine-learning algorithm drawing on 81 pieces of information from electronic health records studied from 2000 to 2016 to produce a single score that indicates the likelihood of being diagnosed with type 1 diabetes. The information used in the tool included flags such as family history, fatigue, urinary tract infections, obesity, and weight loss, as well as data on the frequency of recent primary care contact relative to average contact frequency for each child.
The Welsh SAIL/Brecon registry of approximately 35 million primary care contacts for 1 million children (0.21% with type 1 diabetes) was used as the training dataset, and the tool was tested using the English Clinical Practice Research Datalink (CPRD) and Hospital Episode Statistics records involving around 43 million contacts for 1.5 million children (0.10% with type 1 diabetes).
When testing their model, the researchers created an alert system based on the tool and different thresholds, and used the relationship between benefit (proportion of children who were identified by the tool and received a diagnosis) and effort (proportion of primary care contacts with a flag for type 1 diabetes) to assess its performance.
Based on an effort threshold of 10% – meaning that 1 in 10 children would have been alerted on the systems had the tool been in place – the algorithm identified 75% of children in the CPRD dataset who went on to be diagnosed with type 1 diabetes in the following 90 days, compared with just 25% of these children if the algorithm was not used.
And when the reduced threshold of 4.6% was used, 64% of children with type 1 diabetes would have been identified in the 90 days prior to diagnosis, versus 12% if the algorithm was not used.
Townson then shared an example of a child diagnosed with type 1 diabetes just before 2.5 years of age to illustrate how long before diagnosis an alert should be considered a success for the machine-learning tool. The child was diagnosed in DKA, but had alerts at 32 and 19 days prior to diagnosis. Primary care contacts prior to these timepoints did not raise alerts.
medwireNews is an independent medical news service provided by Springer Healthcare Ltd. © 2022 Springer Healthcare Ltd, part of the Springer Nature Group