Digital health application integrating wearable data and behavioral patterns improves metabolic health

37.3 million adults in the U.S. have diabetes; 95% of those have T2D¹. Another 96 million U.S. adults have prediabetes; 5–10% of that population becomes diabetic every year², and it is estimated that 70% of individuals with prediabetes will convert to T2D over their lifetime³. The annual cost of diabetes in the USA is $327b. Individuals with diagnosed diabetes incur $16,752 in health care expenses per year, which is 2.3 times greater than individuals without diabetes⁴.

Three major studies in the USA, Finland, and China have shown that lifestyle interventions, including dietary weight loss and physical activity, can prevent diabetes progression in high-risk individuals by 51–58% from 2.9 to 6 years after the original intervention^5–7, with continued protection of 27–43% for up to 20 years^7–9. Two additional prospective trials in India and Japan showed similar protection at 3–4 years post-intervention^10,11.

Cornerstone features of the highly successful Diabetes Prevention Program include: individual case managers or “lifestyle coaches”; frequent contact with participants; a structured, 16-session core-curriculum that teaches behavioral self-management strategies, followed by additional classes and one-on-one meetings by phone or in person at least once per month; self-monitoring of weight, dietary intake, and physical activity; supervised physical activity sessions; tailoring of materials and strategies to address ethnic diversity; and an extensive network of training, feedback, and clinical support¹². Due to its success, this approach has been widely adopted as the standard of care in individuals at risk for T2D. The interventions needed to implement this program, while cost effective, are resource-intensive and require contact with health care professionals/systems that may not be accessible to all, with reported participation rates as low as 2.6% due to lack of physician referrals and socioeconomic barriers¹³. During the first 4 years of implementation of the National DPP effort, only 39% of participants were retained at 44 weeks, and only 35.5% achieved the 5% weight loss goal¹⁴. Furthermore, while personalization is recommended, the current approaches offer no formal method by which to accomplish such personalization.

The explosion of new technologies that enable continuous glucose monitoring and activity tracking, as well as the widespread use of smartphones (currently used by 85% of US adults)¹⁵, provides a unique opportunity to leverage these technologies to significantly enhance the efficacy and practicality of lifestyle interventions. Technology-enabled diabetes self-management approaches have gained traction as supplements to traditional diabetes self-management models and implementation of DPP-based diabetes prevention approaches. Phone-based apps allow users to log fingerstick^16–18 glucose values, body weight, food eaten, and physical activity, with some apps accessing large nutritional databases that allow patients to determine calories and carbohydrates consumed when the user enters food items or scans a barcode. Wearable devices log steps, minutes of activity, miles attained, and heart rate changes, as well as estimate calories burned¹⁹.

Continuous glucose monitors (CGM) are wearables that, via alarm features and real-time feedback to the user about glucose trends, were initially shown to lower both hyper and hypoglycemia in T1D and multidose-insulin-treated T2D^20–27. Recently, significant but modest benefits (HbA1c −0.29% vs FSBG) were also shown in non-insulin-treated T2D²⁸. There has been no formal evaluation of CGM as a tool for enhanced lifestyle modification as a method to treat or prevent T2D. Several small proof of concept studies have been published, showing that individuals with T2D using CGM chose lower glycemic index foods²⁹, increased physical activity³⁰, decreased caloric intake, lost weight, and demonstrated decreased postprandial glucose levels³¹. Among individuals with prediabetes, there is only one published study addressing the role of CGM in promoting behavior change, and, while it showed greater dietary self efficacy, neither weight nor glycemic measures were reported³².

The benefits of CGM as a behavior modification tool would be magnified if information relating food and activity choices were linked to glycemic responses and shared with the user, who could then learn from the combined data. Current technology, including wearable devices that simultaneously and continuously track multiple health metrics and mobile apps that can integrate data from wearable devices, have the potential to revolutionize behavioral approaches to diabetes treatment and prevention. While the current approaches to lifestyle management have proven successful, they are not tailored to individuals, who may respond differently to nutrients and activity. Indeed, several studies showed that glycemic responses between individuals differ following exposure to the same foods^33–35, which may reflect a variety of interindividual biologic differences such as microbiome³⁵, genetics³⁴, and underlying physiology such as insulin resistance and beta cell function³⁵. Adherence to diets and physical activity also varies between individuals due differences in personal preferences and environmental factors³⁶.

Here, we sought to determine if a novel digital technology-based program, in which millions of CGM and other health data points were used to provide individualized feedback and tailored recommendations based on a user’s personal data patterns could improve lifestyle choices and metabolic health. We hypothesized that this approach, by inducing behavior modification that included healthier eating and physical activity, and took into account personalized glycemic responses and preferences, would promote weight loss, increase physical activity, and reduce hyperglycemia, which are important for treating and preventing T2D as endorsed by the American Diabetes Association and European Association for the Study of Diabetes for treatment of T2D³⁷.

This real-world study demonstrates that use of digital technology in combination with CGM can facilitate lifestyle interventions that yield improvement in glycemic measures, in the largest cohort to date.

Three studies have previously examined the use of CGM for lifestyle change T2D^29–31 Cox et al.²⁹ included only four patients and did not have a control group. Allen et al.³⁰ utilized solely physical activity intervention. Yoo et al.³¹ randomized patients with poorly controlled T2D (baseline HbA1c of 9%) who were treated with insulin (60%) or noninsulin therapies (40%), to real-time CGM vs. self monitoring blood glucose (SMBG) and demonstrated significantly greater reduction in HbA1c (0.50%, p = 0.004) at 12 weeks in the CGM group, along with reduction in total daily calorie intake, greater weight loss, and increased physical activity. Wada et al.²⁸ examined individuals with non-insulin-treated T2D with baseline HbA1c of 7.8%; the researchers did not measure lifestyle changes, but showed that use of flash CGM as compared to SMBG resulted in a reduction in HbA1c at 24 weeks (0.29%, p = 0.02).

Our results extend the prior studies by demonstrating that use of CGM with a digital app designed to enhance healthy lifestyle behaviors by informing users of the impact of food and activity choices on glycemic responses yields both behavior changes and glycemic improvement in individuals with prediabetes and even earlier stages of dysglycemia detected by CGM. Combination of CGM data with heart rate and activity information, as well as content and personalized insights, is a novel approach to management of metabolism, and yielded noteworthy results.

Glycemic improvement was observed in the group as a whole, as was reduction in glycemic variability, but the greatest improvement was evident in those whose glucose was not within the optimal range at baseline, defined as 70–140 mg/dL for those without diabetes and 70–180 mg/dL for those with T2D. This is the first study to evaluate personalization of lifestyle recommendations based on previously observed glucose patterns from CGM. The use of digital technology for diabetes treatment and prevention has been previously studied, with applications ranging from remote coaching to dissemination of educational programs such as the DPP⁴⁰ Importantly, none of these studies have evaluated integration of CGM with lifestyle measures or coaching, especially in addition to presentations to subjects of the projected glucose impacts of specific foods/meals and/or activity patterns.

Like prior digital health apps, the one used in this study served as a virtual coach, offering recommendations and reminders that addressed many components of medical nutrition therapy as recommended by the American Diabetes Association⁴¹ These include choosing healthy foods, and tailoring recommendations to consider individual preferences and environmental constraints (access, budget, living situation). Most importantly, however, this technology system (CGM and mobile app) uniquely integrated glycemic excursions obtained from CGM with actual food consumed, which allowed for personalized rather than generic coaching. Consultation with a RD in order to develop an individualized eating plan is associated with HbA1c reductions of 0.3-2.0%⁴² In the present study, which spanned only 28 days and did not include any live coaching, GMI decreased by 0.43% in those with baseline GMI > 7%, and by 0.22% in those with baseline GMI 6–6.9%. In addition, hyperglycemic events (glucose spikes), glucose variability, and TIR improved in all subgroups ranging from healthy nondiabetic to T2D, although the small T2D subgroup reached statistical significance only for glycemic variability. Hypoglycemic events <70 mg/dL also decreased in the group as a whole, although, in subgroup analyses, statistical significance was reached only in the healthy nondiabetic subgroup.

Based on data collected, the observed glycemic improvements likely resulted from three lifestyle changes: (1) healthier food choices; (2) weight loss; (3) increased physical activity. While only 526 of the 1066 participants logged sufficiently to be included in the detailed dietary analysis, food log reviews showed a decrease not only in total calories, carbohydrates, sugar, and saturated fat, but also in the proportion of calories from carbohydrates, sugar, and saturated fat, with an increase in the proportion of calories from protein, healthy fats (fats other than saturated include mono and polyunsaturated), and fiber. Although we did not quantify changes in the intake of meals that led to glucose elevation, the observation that there were fewer hyperglycemic events and that TIR improved strongly suggests that glucose-elevating meals were decreased.

Weight loss is a cornerstone of diabetes treatment and prevention. In the present study, use of the CGM and mobile app promoted weight loss in >75% of participants who continued to log weight. The amount of weight lost was greater in those with T2D and those with higher baseline starting weight, who at 12 weeks lost up to 10.4 and 18.9 lbs, respectively. At 12 weeks, the % loss from initial body weight was 1.5, 2.3, and 5.1% in healthy nondiabetic, prediabetic and T2D participants, respectively; and was 2.0, 3.2, and 6.8% in those weighing 150–199, 200–249, and >250 lbs at baseline. Because not all participants logged weight throughout the study, this may selectively represent those who were more successful in losing weight, who owned a personal scale, or who were more motivated in general. Thus, the weight loss benefits in those who did not continue to log weight is not known. Nonetheless, it appears very likely that use of the CGM with app and self-weighing at least once after baseline is associated with progressive weight loss for at least 12 weeks of use. Longer term studies will be needed to determine whether this is sustained.

Physical activity also lowers glucose levels, and the participants in the SOM program demonstrated increased physical activity as measured both by self report (logging) and by an objective measure (HR > 110 bpm). Multiple previous studies have demonstrated that use of accelerometer or heart rate monitor increases physical activity^43,44 Thus, it is not discernible whether simply wearing the HR tracker/accelerometer or receiving personalized activity recommendations from the app contributed to the observed increase in physical activity. Further, the nature and intensity of the activity, as well as nuances in HR change that more specifically address intensity and conditioning, were not available for this analysis. Ultimately, the observation that the technology-based intervention as described leads to behavior modification and clinical benefits is important. It should be noted that participants in both the healthy and pre-diabetes cohorts already had relatively high levels of baseline activity, which could have skewed the data vis a vis their inclination toward performing additional exercise; nevertheless, significant increases in physical activity among both groups was observed.

It is worth noting the significance of the glycemic improvements observed in the healthy nondiabetic group. That the observation group exhibited glycemic excursions that could be improved by diet might come as a surprise to many. CGM data on nondiabetics is scant in the literature. While the TIR is high in this group at baseline, in the current analysis, which is the largest cohort of nondiabetic individuals with published data on CGM, glycemia still improved in multiple metrics. Whether improved glycemia in this population prevents diabetes or improved health outcomes is not currently known, as diabetes prevention studies have been conducted in high risk individuals with prediabetes. The data presented in fact suggests that with the advent of CGM an even earlier stage of dysglycemic is detectable and can improve with lifestyle interventions. This is important as future studies should examine the long term health risks (T2D, cardiovascular disease) as well as the impact of diabetes prevention strategies in this group of individuals. Such studies, which will take years to conduct, will determine whether the improvements in glycemic profile would be beneficial to health and potentially reduce the risk for T2D.

The current findings have several limitations inherent in real world studies. First, there was no control group—data presented are based on the change from baseline to end-of-program and thus could represent a “placebo” effect from simply being enrolled in a program. A future randomized trial with a comparator group would extend and confirm the present findings. Second, the requirement to use a CGM means that it is likely that the participants in this study demonstrated high levels of self-efficacy, and thus it is not clear that results would translate into less self-motivated individuals who were recommended by a health care provider to engage in a similar intervention. Third, some of the data captured depends on adherence to logging. It is possible that participants did not log all food or activity; thus, alternate metrics were used, such as the ratio of macronutrient to total calories and HR > 110 min, which generally supported the logged metrics. Logging weight is also subject to success bias and thus may only be interpreted to reflect the weight loss of those who weighed themselves at least once after starting the program. Fourth, ethnicity was not collected in this cohort. It is important to recommend ethnically and culturally appropriate lifestyle interventions, particularly with regard to dietary recommendations, and thus improvement of this application would include capture of ethnicity and ethnic/culturally-sensitive recommendations if desired by the user. Fifth, the duration of intervention was 28 days. Longer duration studies will be required to ascertain the durability of behavior changes and glycemic benefits. Finally, weight reporting in this study was not comprehensive as weight loss was not a specific goal of our study and therefore weight monitoring was optional. Weight logging was not required, nor were users reminded to log or enter their weight. It is not appropriate to compare our study to the DPP in terms of weight reporting, as our study’s focus was on logging food and activity and measuring glucose levels with CGM.

In our participants without a diagnosis of T2D, we used thresholds for Time in Range (TIR) that were previously reported in the literature (70–140 mg/dL) or justifiable based on translation to Hemoglobin A1c (Glycemic Management Indicator (GMI)). Currently, there are no TIR thresholds for patients without diabetes, but we believe that they will be established in the future³⁸ It is not clear whether glucose levels outside a prespecified range (eg 70–140 mg/dL) or above a given threshold (>180 mg/dL) in individuals who do not meet traditional criteria for diabetes lead to clinical consequences such as microvascular disease. It is also not clear whether glucose levels on CGM predict increased conversion rates to T2D in those who do not meet traditional OGTT criteria for prediabetes. Long-term studies will be required to ascertain these important questions. However, both mild glycemic excursions into prediabetic and diabetic range, and glycemic variability, which has been associated with increased risk for cardiovascular disease⁴⁵ are more readily ascertained with CGM than with older methods of glucose measurement, and it is likely that future studies will reveal answers to the questions posed above regarding the prediction of clinical events according to specific CGM metrics in both T2D and earlier stages of dysglycemia.

In summary, we demonstrate that an app-based platform that integrates food and physical activity logging with CGM and HR data, and provides personalized lifestyle recommendations based on user input enhances healthy lifestyle practices and improves metabolic health in individuals with and without T2D. This nee technology improves glucose profiles even in individuals who by current standards are normoglyemic, and promotes weight loss in overweight and obese individuals, highlighting its potential for early intervention. Importantly, the ability to offer intervention without human coaching (with or without AI-enabled personalization) enables affordable scaling to large numbers of people, including those who are underserved or living in remote areas––thus helping to reduce the continually increasing prevalence of prediabetes and diabetes. In this respect, it will be important to optimize user engagement, including customization to varied ethnicities and socioeconomic levels, and to consider the needs of different age groups, including the elderly who often have more difficulty accessing in-person health care services, and to youth and young adults, who tend to be less adherent to traditional models of diabetes prevention and may selectively benefit from technology-based models. Overall, this approach and similar technology-based approaches have the potential to improve metabolic health at early stages, and may increase the efficacy of current practices to prevent and treat T2D through lifestyle modification.

The methods were performed in accordance with relevant guidelines and regulations and approved by the Advarra review board.