- Researchers recently created an artificial intelligence model that predicts diabetes onset with 12 hours of blood glucose data collected from a wearable device.
- They say their model could aid the diagnosis of prediabetes and help prevent type 2 diabetes.
- How the AI model will impact rates of type 2 diabetes screening remains to be seen.
Diabetes is a chronic condition
Prediabetes, or “intermediate hyperglycemia,” is the high risk stage before type 2 diabetes when blood sugar levels are above average but below the threshold for diabetes.
Diabetes and prediabetes diagnoses typically involve blood tests, including the A1C test, a measure of a person’s average blood sugar over the last 3 months, a fasting blood sugar test, a glucose tolerance test, or a random sugar test.
New ways to screen for prediabetes and type 2 diabetes may encourage individuals to get tested.
In a new study, researchers investigated whether they could use readings from continuous glucose monitors (CGM) to diagnose prediabetes and diabetes. With just 12 hours of glucose profile data, the researchers could classify type 2 diabetes, prediabetes, and people without impaired glucose tolerance.
“I believe our method offers a lot of potential to be used as a novel tool to aid healthcare providers in their own decision-making, especially for remote or virtual care of patients. For the general public, our method could not only be used for monitoring and early screening but alerting a patient of their risk of developing diabetes.”
The findings were recently presented at the 36th Conference on Neural Information Processing Systems (NeurIPS) in New Orleans, LA.
Continuous glucose monitors (CGM) are wearable devices that measure blood glucose every 15 minutes.
CGMs help people with diabetes regularly monitor their blood sugar levels.
“Continuous glucose monitors (CGM) are gaining traction to be worn in the nondiabetic, general population for health reasons or other specific goals,” Dr. WIlliam Dixon, clinical assistant professor of emergency medicine at Stanford University and co-founder of Signos, not involved in the study, told MNT.
Dr. Dixon added that determining the presence of and risk for diabetes based on CGM data may be helpful for people who are not routinely tested or screened for the condition.
“There are also signs of impaired glucose tolerance that can be apparent even before average glucose levels reach a concerning range,” Dr. Dixon said.
For the study, researchers used data from 436 participants from India.
Each participant wore a CGM device for an average of 12 days and provided data including their sex, age, and body mass index (BMI).
The researchers defined participants’ A1C levels of 6.5% and higher as type 2 diabetes, 5.5%–6.5% as prediabetic, and under 5.5% as healthy.
Among the participants, 172 had type 2 diabetes, 87 had prediabetes, and 177 were healthy. Diagnoses were confirmed by physicians.
The researchers created AI prediction models based on different blood glucose level time durations. They compared models based on 12, 24, 72, 168, and 288-hour windows of data.
They found that CGM data was 1.21, 1.34, and 1.17 times more accurate than demographic data in identifying type 2 diabetes, prediabetes, and healthy individuals.
They also found that their 12-hour model was similarly effective as longer duration models.
After optimizing the 12-hour model, they identified 87%, 84%, and 86% of those with diabetes or prediabetes and healthy individuals.
The researchers noted that of those in the 12-hour prediction, 23 were misclassified due to unusual 12-hour readings in which they reported the same blood glucose levels over time.
The researchers concluded that CGM systems could enable quick and accurate screening of diabetic outcomes.
The researchers hope to conduct similar studies on larger cohorts to improve their prediction models.
When asked about the study’s limitations, Dr. Jeon told MNT:
“Our findings are developed based on about 400 patients’ CGM signals. Further evaluation is required using an independent larger cohort and bigger population data to generalize our method. However, we are encouraged by the results and look forward to our continued work in this area.”
Dr. Jeon noted that with Klick’s predictive diagnostics, people could learn their results from home instead of traveling to a clinic for blood tests and waiting for a few days.
Michael Lieberman, Ph.D., managing director of research and development at Klick Applied Sciences, also told MNT:
“From a public health standpoint, prediabetes is hugely underdiagnosed. The ability to easily determine with a high degree of likelihood that someone is prediabetic without a doctor’s visit could be extremely helpful in identifying people at risk of becoming diabetic.”
Dr. Lieberman added that early detection of prediabetes could provide a person’s healthcare team ample time to reverse the course of the disease before it’s too late.
Dr. John Miles, an endocrinologist from The University of Kansas Health System, not involved in the study, noted that the practical implications of this study are relatively modest. He noted to MNT:
“I am not sure we can say at this stage that continuous glucose monitoring (CGM), as performed in this study, is an improvement over existing methods for diagnosing diabetes. It is certainly true that some people don’t know they have diabetes or prediabetes. However, the fact that CGM can accurately define which category people are in (diabetes, prediabetes, or non-diabetes) as defined by hemoglobin A1c doesn’t mean that it would be a practical alternative to [A1c testing].”
“[A1c testing] using a fingerstick blood sample is currently used to screen for diabetes and would be simpler, faster, and probably cheaper than CGM in a mass screening program,” Dr. Miles concluded.