Continuous glucose monitoring (CGM) can not only the track blood sugar of people with diabetes but also identify those at risk of the condition when combined with artificial intelligence.
Research showed just 12 hours of glucose profiles provided the predictive model with enough information to distinguish diabetic, prediabetic, and healthy individuals.
The tool could be used for monitoring and early screening of diabetes, particularly in scenarios involving remote or virtual care, and give care teams the chance to reverse the course of disease.
The findings were presented at the NeurIPS 2022 thirty-sixth Conference on Neural Information Processing Systems in New Orleans, Louisiana.
‘We were most surprised by our machine learning model’s ability to make accurate assessments on such a limited amount of clinical information,” said lead study scientist Jouhyun Jeon, PhD, a principal investigator at Klick Applied Sciences.
She told Inside Precision Medicine that the initial goal had been to half the standard monitoring time to seven days.
“So getting down to just 12 hours of CGM recordings while maintaining high accuracy was a really phenomenal, unexpected result that has the tremendous potential to accelerate diabetes prevention,” she said.
Prediabetes is defined by abnormal glucose homeostasis, such as impaired fasting glucose and impaired glucose tolerance.
Despite being increasingly identified as a critical metabolic state, about 90 per cent of U.S. adults with prediabetes remain unaware of their condition.
To determine whether AI could help, the team used CGM to measure blood glucose in 436 individuals recruited from four states in India as they went about their daily lives.
Of these, 172 were identified as having Type 2 diabetes, with glycated hemoglobin levels more than 6.5%, 87 had prediabetes, with levels between 5.5% and 6.5%, and 177 were defined as healthy individuals, with levels below 5.5%.
The CGM measured blood glucose levels every 15 minutes for a maximum of 14 days, as well as providing alerts on hypo- or hyper-glycemic events and giving information on glucose variability.
CGM signals from 70% of the participants were used to train and validate predictive models, while the remaining 30% of participants were used to test model performance.
A baseline model that included demographics such as gender, age, weight, height, and body mass index had an overall balanced accuracy (BCC) of 0.48 and an area under the curve (AUC) to detect prediabetic patients of 0.49.
CGM-based models improved overall BCC by 140% and AUC by 122% compared with this.
CGM-based models were also 1.21, 1.34, and 1.17 times more able to identify Type 2 diabetes, prediabetes, and healthy individuals, respectively.
The 12-hour model performed competitively compared with those of a longer duration. It showed an overall BCC of 0.67, which was equivalent to the performance of the model using the full-time CGM signals and had an overall AUC of 0.78. It had an AUC of 0.69 for identifying prediabetes.
Klick Applied Sciences managing director of research and development Michael Lieberman, PhD, admitted that work was still needed before the model could be deployed for the everyday management of patients.
“We see these results as a highly exciting first step, but one that will require further validation before anyone can start thinking about seeing this used in clinical practice,” he told Inside Precision Medicine.
“We are working with academic and commercial partners to take those next steps and are excited for the potential real-world applications.”
