AI-Supported Coaching Shows Promise for Improving Glycemic Control

Among patients who had been living with diabetes for an average of 9 years, 71% of those who were randomized to the AI-enabled intervention achieved the endpoint of a target hemoglobin A1c (HbA1c) level < 6.5% at 12 months compared with just 2.4% of those randomized to the usual-care group (P < 0.001).

Kevin M. Pantalone, DO (Cleveland Clinic, OH), the study’s lead author, said the results of this bundled intervention exceeded the researchers’ expectations.

“To actually be able to achieve this aggressive target at the same time that a significant number of patients were reducing their number of therapies—including many of the newer and more novel modern therapies like GLP-1 receptor agonists and SGLT2 inhibitors—is very telling in just how powerful these types of interventions can be on patient behavior and improving diabetes outcomes,” Pantalone told TCTMD.

Once loaded onto the patient’s smart phone, the app uses machine-learning algorithms to monitor continuous glucose data from a monitor the patient wears. Those data are used for real-time guidance, anticipating changes in blood sugar with certain meals, and making suggestions to patients about how and what to eat.

“I think the key here is that it’s providing personalized recommendations in real time,” Pantalone noted. “When I see patients in clinic and I say, ‘watch your diet and exercise,’ or I send them to a nutritionist, they do get the information, but then three months go by and they’re at the grocery store shopping and there’s no one there to help them make decisions. Or they’re sitting down eating dinner with their loved ones and there’s no one there guiding their decision making, so what this system really provides is personalized recommendations in real time, which is what helps to change behavior.”

“I think the key here is that it’s providing personalized recommendations in real time.” Kevin M. Pantalone

Another important aspect is that by using this type of guidance, patients have significant freedom in that their food choices are guided by personal preferences, Pantalone and colleagues say.

Patient satisfaction with the system was so high that the majority of those who participated in the year-long study wanted to keep using it. While those patients received access to it as part of the trial, the system is commercially available and some insurers will cover it with a request from the patients’ physician, Pantalone said.

Bundled Intervention Wins Seen

For the single-center study, published last week in NEJM Catalyst, Pantalone and colleagues randomized 100 patients to the bundled intervention (Twin Health) and 50 to standard of care. The average age was 58.5 years, mean BMI was 35.1, and average HbA1c level was 7.2%. Participants were required to have access to a smartphone compatible with the app and sensors and had either an HbA1c level of 7.5% to 11%, an HbA1c of 6.5% or greater but less than 7.5% on a glucose-lowering medication, or an HbA1c < 6.5% on at least one nonmetformin glucose-lowering medication.

After installing the app, participants were provided with four electronic devices that communicated with the app via Bluetooth: a continuous glucose monitor (Dexcom G6 or G7), an activity sensor (Garmin vívosmart 4), a smart weight scale (RENPHO), and a blood pressure meter (TaiDoc TD-3140).

In addition to recommendations about food choices, the app would recommend physical activity, sleep, and deep breathing exercises, which the participants were encouraged to follow, as well as to log their actions. A color-coded food-selection system was used in the app, along with the patients’ preferred foods, to guide both grocery shopping and daily meal choices. No calorie restrictions were used, but 5,000 steps per day of exercise was encouraged initially, gradually increasing to a goal of 7,000 steps or more. Resistance training for 20 minutes a day, three times a week was also encouraged. Patients could communicate with human health coaches in daily chats and also had access to video calls at prescribed times.

To reflect real-world practice, any management decisions regarding the patient’s diabetes and medications were managed by their primary care provider with no direction from the researchers.

Compared with the usual care group, those in the intervention group had greater mean changes from baseline to 12 months in both mean HbA1c and mean body weight, with positive associations between changes in BMI or body weight and changes in HbA1c over time (P < 0.001).

In a posthoc analysis, the intervention group saw reductions in the use of GLP-1 receptor agonists, dipeptidyl peptidase 4 (DPP-4) inhibitors, sulfonylureas, sodium-dependent glucose cotransporter (SGLT) 2 inhibitors, and insulin, with few corresponding changes in the usual care group.

By 12 months, the intervention group had lost an average of 8.6% of their body weight versus 4.6% for the usual care group (P < 0.001).

Pantalone said the findings support the idea that the intervention works, not only through impressive weight loss but also by helping patients stop medications that aid in weight loss.

“A common question that we get from patients is when are they going to be able to stop these medications,” Pantalone said. “ I generally [tell them] they’re not going to stop these medicines and in fact it’s more likely that over time we’re going to have to add additional medicines to help to maintain blood sugar control. It’s very discouraging [and] people feel defeated. What I saw with this intervention was that it really empowered people to want to get off these medicines and to actually give them the tools and the knowledge necessary to accomplish that. Watching it happen was very, very impressive.”

The researchers say this type of intervention could have wide applicability in diverse populations with diabetes, while placing minimal time burdens on primary-care resources. Longer term studies, including one that will follow patients from this study who chose to keep using the AI-enabled intervention, are needed, however, to understand if the approach is associated with sustained glycemic improvement over time.