As of June 2021, 10,197 participants (aged >18 years and living in the United States) of Foodsmart who enrolled since January 2016 had entered a plausible value for HbA_1c (HbA_1c >3% or HbA_1c <15%). Of those, 643 Foodsmart participants had entered at least two HbA_1c entries, with the first and last entry at least 30 days apart. The final sample size was 643 participants who had at least two reports of HbA_1c.

Foodsmart is a digital nutrition platform that encourages sustained behavior change through nutrition education and personalized meal planning, and promotes healthy eating and nutrition through online grocery and food ordering integration. Foodsmart has two components, FoodSmart and FoodsMart, to help users learn how to eat healthy to meet their nutrition targets and order affordable, tasty, and healthy food online, respectively.

The FoodSmart component provides participants with digital dietetics information on how to better plan meals to meet their nutrition targets. Once participants enroll, they are prompted to fill out the Nutriquiz, an online dietary assessment. Participants report their usual dietary intake and meal planning habits and based on the responses, the assessment provides specific dietary recommendations and a tailored meal plan. Participants can retake the Nutriquiz assessment at any time to track their progress toward their health goals.

The second component is FoodsMart, an online food purchasing environment that promotes buying healthy groceries and meals. Personalized meal plans are converted into a grocery list and integrated into online ordering and delivery of meal kits, prepared foods, and groceries. Participants are encouraged to purchase healthy options that align with their preferences and personalized meal plan. Customized grocery discounts for healthier food options and budget-based purchasing that compares prices across integrated grocery partners help participants save money and further encourage participants to choose healthy food options.

Foodsmart is available through health plans and employers and can be accessed via the web or iOS or Android operating systems.

On the Foodsmart platform, participants were able to enter biometrics such as height, weight, HbA_1c, blood pressure, and lipids, and were able to update their biometrics at any time. Given the potential for error when entering self-reported metrics, the following values were considered as incorrect entries and were replaced with a missing value: HbA_1c ≤3% or ≥15%, BMI ≤15 kg/m² or ≥50 kg/m², and weight ≤27.2 kilograms or ≥181.1 kilograms. We only included participants who reported an HbA_1c measurement at least twice, and we used the first (baseline) and last (end) values entered. Length of follow-up was calculated as the number of months between the date of the first value and the date of the last value. We defined HbA_1c ≥6.5% as the cutoff for diabetes as defined by the American Diabetes Association [22]. The same method was applied to the end HbA_1c value to assess diabetes status at the end of follow-up. Since a glycemic target of HbA_1c <7% is recommended for nonpregnant adults, as defined by the American Diabetes Association, we used the cutoff of 7% for sensitivity analyses [5]. Changes in HbA_1c were calculated by subtracting the first reported value from the end value. Percent change was calculated by dividing the change in HbA_1c by the first HbA_1c entry.

Baseline BMI was calculated as the first weight entry in kilograms divided by height in meters squared (kg/m²). Participants’ baseline BMI was categorized as normal BMI (BMI <25 kg/m²), overweight (25-29.9 kg/m²), or obese (BMI ≥30 kg/m²). Participants were also able to report any conditions they currently had (eg, high blood pressure, high cholesterol) in the Nutriquiz.

Participants self-reported their usual dietary intake and habits in Foodsmart. Upon enrollment, participants were prompted to fill out a 53-item food frequency questionnaire called Nutriquiz (adapted from the National Cancer Institute Diet History Questionnaire I [23]). Demographic information (age, sex, height), weight, and daily dietary intake (added sugars, fiber, fruits, vegetables, whole grains, fats, proteins, water, and sodium) were also obtained using the Nutriquiz.

Based on responses from the Nutriquiz, a score (Nutriscore) was calculated to assess overall diet quality, which is based on the Alternative Healthy Eating Index-2010 and the Commonwealth Scientific and Industrial Research Organization Healthy Diet Score [24,25]. Participants were assigned a total Nutriscore from 0 to 70 based on the sum of scores for 7 components: fruits, vegetables, protein ratio (white meat/vegetarian protein to red/processed meat), carbohydrate ratio (total fiber to total carbohydrate), fat ratio (polyunsaturated to saturated/trans fats), sodium, and hydration (percent of daily fluid goal). Each of the components was scored from 0 to 10, with 10 being optimal. Change in the Nutriscore was calculated as a participant’s last Nutriscore minus the participant’s first Nutriscore. A positive change in Nutriscore indicates the participant improved their dietary quality.

We used descriptive analyses to examine the baseline demographic characteristics, HbA_1c levels, and diet quality of the study population as a whole and according to whether participants had diabetes at baseline or not. We reported categorical variables as number of participants (percentage of study population) and continuous variables as mean (SD). We used chi-square tests to assess whether categorical variables are independent of baseline diabetes status, and two-sample t tests to evaluate differences in continuous variables.

Univariate and multivariable logistic regression were used to estimate the odds ratios (ORs) and 95% CIs of having diabetes at baseline. The multivariable logistic regression model was mutually adjusted for gender, age category, baseline BMI category, baseline Nutriscore, high blood pressure, and high cholesterol.

Among participants who had diabetes, we calculated the mean changes in HbA_1c overall and by time of follow-up (>6 months, >12 months, >24 months). We used paired t tests to test whether the changes were statistically significant. Additionally, we calculated the mean percent change for HbA_1c. In a sensitivity analysis, we used a threshold of HbA_1c ≥7% to calculate mean changes in HbA_1c.

We also calculated the percentage of participants with diabetes at baseline who returned to normal HbA_1c levels by the end of follow-up, and stratified by follow-up length. We conducted a sensitivity analysis using a threshold of HbA_1c ≥7%.

To further explore the performance of HbA_1c, we examined changes in weight and HbA_1c stratified by whether participants with diabetes at baseline achieved normal HbA_1c levels (HbA_1c ≥6.5%) by the end of follow-up.

We considered P values less than .05 to be significant for all tests. R Studio (version 1.4.1106) and R (version 4.0.5; R Foundation for Statistical Computing) were used for all analyses.

The study was declared exempt from institutional review board oversight by the Pearl Institutional Review Board given the retrospective design of the study and the less than minimal risk to participants.

Baseline characteristics of the total study sample and those stratified by baseline diabetes status are shown in Table 1. We found that 43.5% (280/643) of participants had diabetes at baseline. There were 643 participants included in the analysis, of which 64% (411/643) were female and 61% (391/643) were between 40 and 59 years old (Table 1). The mean weight was 93.9 (SD 23.8) kilograms, the mean baseline Nutriscore was 31.4 (SD 8.5) points, and the mean change in the Nutriscore was 3.2 (SD 7.1) points. The mean follow-up length was 10.4 (SD 7.1) months and ranged from 1 to 38 months. Compared to participants who did not have diabetes, participants who did have diabetes were significantly more likely to be male, to have a higher weight and BMI, to have a lower baseline Nutriscore, and to self-report having high blood pressure. Participants with diabetes at baseline were also more likely to have a higher increase in Nutriscore, a longer follow-up duration, and self-reported high cholesterol compared with participants without diabetes at baseline, although the differences were not statistically significant.

To better understand what type of participant was likely to have diabetes at baseline, we examined the association between baseline characteristics and odds of having diabetes in univariate and multivariable logistic regression models (Table 2). In the univariate regression models, participants who were female were 40% less likely to have diabetes at baseline than participants who were male (OR 0.60, 95% CI 0.43-0.82, P=.002). Participants classified in the overweight BMI category were 86% more likely to have diabetes at baseline than participants classified in the normal BMI category (OR 1.86, 95% CI 1.10-3.19, P=.02). Participants classified in the obese BMI category were 151% more likely to have diabetes at baseline than participants classified in the normal BMI category (OR 2.51, 95% CI 1.58-4.09, P<.001). Participants who self-reported having high blood pressure were also 46% more likely to have diabetes at baseline than participants who did not self-report having high blood pressure (OR 1.46, 95% CI 1.06-1.99, P=.02). Participants with a higher baseline Nutriscore were less likely to have diabetes at baseline (OR 0.98, 95% CI 0.96-1.00, P=.03).

After adjusting for all other variables in the multivariable logistic regression model, we found that being female was associated with 44% lower odds of having diabetes at baseline (OR 0.56, 95% CI 0.39-0.79, P=.001). Additionally, participants who were obese were 134% more likely to have diabetes at baseline than those in the normal BMI category (OR 2.34, 95% CI 1.40-3.97, P=.001).

Figure 1 presents the mean and percent changes in HbA_1c levels among participants who were classified as having diabetes for the overall group and by length of follow-up, at >6, >12, and >24 months. The mean changes in HbA_1c overall and at >6, >12, and >24 months were –0.46, –0.37, –0.45, and –0.70 points, respectively. Percent changes in HbA_1c overall and at >6, >12, and >24 months were –6%, –5%, –6%, and –9%, respectively. All changes were statistically significant (P<.05) using paired t tests. For users with an HbA_1c ≥7.0%, mean change in HbA_1c was –0.62 points (P<.001), and percent change was –7.6%.

We calculated the percentage of participants with diabetes at baseline who achieved a normal (<6.5%) HbA_1c level overall and by cumulative length of follow-up time. Among all participants with diabetes, 21.4% (60/280) achieved normal HbA_1c levels, using a threshold of 6.5%. Among participants whose follow-up time was longer than 6, 12, and 24 months, the percentage of participants who achieved normal HbA_1c levels was 21.0% (43/205), 22% (21/97), and 39% (7/18), respectively. In a sensitivity analysis, for participants with an HbA_1c ≥7%, 31.2% (62/199) of them achieved an HbA_1c level less than 7%.

To better understand how weight and HbA_1c changed according to end diabetes status, we examined changes in weight and HbA_1c stratified by whether participants with diabetes at baseline achieved normal HbA_1c levels (Table 3). Reductions in weight and HbA_1c were greater for those who achieved normal HbA_1c levels at the end of follow-up versus those who did not.

In this study of 643 participants who used the Foodsmart platform, we found that 43.5% (280/643) had diabetes at baseline, as defined by their baseline HbA_1c level. Foodsmart participants with diabetes at baseline were more likely to be male and have a higher weight and BMI. On average, HbA_1c decreased by 0.46% among participants with diabetes over a mean duration of follow-up of 10.7 (SD 7.3) months. Among participants with diabetes at baseline, 21.4% (60/280) of those participants achieved a normal HbA_1c level by the end of follow-up. These findings suggest that use of the Foodsmart platform may be associated with improved glycemic control among users with diabetes.

In line with our findings, prior studies evaluating the association between diet interventions and clinical biomarkers showed that various nutrition therapies significantly improved glucose regulation and reduced HbA_1c levels in patients with diabetes mellitus. For instance, Esposito et al [18] conducted a randomized trial to evaluate the effects of a low-carbohydrate Mediterranean diet versus a low-fat diet on HbA_1c levels among individuals with type 2 diabetes. The trial was conducted in Italy and included 215 participants with type 2 diabetes who were classified as obese, had never previously taken antihyperglycemic medication, and had HbA_1c less than 11%. After two years, those on the low-carbohydrate Mediterranean diet had a decrease in HbA_1c of 1.1%, while those on the low-fat diet had a decrease of 0.5%. In our study, participants who had a follow-up time period greater than 2 years were observed to have a 0.7% decrease in HbA_1c. Esposito et al [18] also found that, at the end of their study, 37% and 24% of participants returned to normal HbA_1c levels (using a threshold of 7%) after following the Mediterranean and low-fat diet, respectively. In our study, 31% of participants returned to normal HbA_1c levels (using a threshold of 7.0%). The effects of the low-carbohydrate Mediterranean diet were likely greater than what was observed with Foodsmart because the trial had more specific and stringent guidelines for the participants’ diets than there were with Foodsmart, as participants can make changes that best fit their lifestyles. Therefore, with assistance from Foodsmart in making flexible dietary changes, participants can see changes in their HbA_1c levels that are similar to that of a strict low-fat diet.

In our study, we found that men were more likely to have diabetes, which is in line with a previous study conducted by Nordström et al [26]. The authors analyzed data from The Healthy Aging Initiative, a population-based prospective study of men and women 70 years of age or older in northern Sweden, and found a significantly greater prevalence of diabetes in men than women. They hypothesized that this was due to differences in visceral fat mass among men and women and found that when visceral fat mass was adjusted for, male sex was no longer associated with diabetes. Therefore, their findings suggest that differences in the prevalence of diabetes between males and females may be due to differences in visceral fat mass, which is known to be a strong predictor of diabetes [27,28]. Additionally, we found that people with diabetes were more likely to be obese, which has been established by several studies that have found obesity to be a risk factor for diabetes [29-32]. Finally, we found that there was greater weight change among those who achieved normal HbA_1c levels. This is consistent with findings from Gummesson et al [33], who conducted a systematic review aimed at understanding the association between weight loss and HbA_1c for overweight and obese patients with type 2 diabetes. They found a dose-response relationship between weight loss and reduction in HbA_1c in their participants, which may explain why we see a larger weight change for those who have a greater reduction in HbA_1c and return to normal levels in our study.

Patients with diagnosed diabetes incur mean medical expenditures of $16,750 per year, of which about $9600 is attributed to diabetes [3]. Glucose-lowering drugs such as sodium–glucose cotransporter 2 (SGLT2) inhibitors and glucagon-like peptide 1 (GLP-1) receptor agonists account for a large proportion of these expenditures, with an estimated mean annual cost of $2727 per patient [34]. Significant savings can be achieved if patients meet the target American Diabetes Association HbA_1c level of <7% [5,22]. A 1%, 1.25%, or 1.5% reduction in HbA_1c for a commercially insured patient could result in savings of $801, $1033, and $1266 per patient per year, respectively [35]. In this study, participants with diabetes at baseline who achieved an end normal HbA_1c level reduced their HbA_1c by 1.7% on average, achieving a clinically significant change of 0.5% [36]. Given the high cost of medications, prevention and management of diabetes through eating healthier could be an attractive, low-cost alternative. Unfortunately, we do not know whether participants were on glucose-lowering medications before or during enrollment on the Foodsmart platform. Despite this, we can estimate the difference in costs between prescription medications and Foodsmart. Improved glucose management would cost $327 more per person with diabetes annually relative to current care, largely due to use of antihyperglycemic medications [37]. In comparison, as of 2021, the Foodsmart platform on average costs $12.30 per eligible member annually. Using the results above, a 1% reduction in HbA_1c would cost $26.98 on average. On the other hand, using metformin or liraglutide (a GLP-1 receptor agonist) to reduce HbA_1c levels by 1% would cost on average $120 and $8640, respectively. Given that the cost of metformin is 4 times higher than the cost of using Foodsmart, and assuming participants on the Foodsmart platform were not on glucose-lowering medication, the cost of a digital platform like Foodsmart would be significantly more affordable than standard treatment with diabetes medications [38-40].

There are some important limitations to note for this study. The first is that HbA_1c levels were self-reported and were not clinically validated. However, these values should still be fairly accurate, particularly for participants with diabetes who used the app to track their HbA_1c levels. Since participants were not required to enter HbA_1c levels, we have reason to believe people who did—in particular, participants with diabetes—had purposefully entered their HbA_1c levels rather than entering an arbitrary HbA_1c level, which would lead to greater inaccuracy. Additionally, follow-up time was based on when the biometrics were entered, but did not necessarily line up with when the labs were conducted. Another issue is potential selection bias for participants with diabetes who choose to use the platform and are included in the study. For example, those with diabetes who use the app, particularly as a tracker, might be more inclined to want to make changes to their lifestyle. They may have made changes outside of what they did in the app that resulted in changes in HbA_1c. Therefore, we cannot definitively conclude that Foodsmart's platform caused these changes in HbA_1c, but there could be an association between using the platform and HbA_1c changes. A randomized controlled trial must be conducted to determine if there is a causal link. In addition, there are other potential factors influencing diabetes status at baseline and changes in HbA_1c that we might not be able to evaluate because certain types of data are not collected in the Foodsmart app. For example, we do not have participants’ personal or family medical histories to understand their influence on diabetes status [41]. We are also unable to assess how the use of diabetic medications may influence HbA_1c, as well as other medication-induced fluctuations in HbA_1c. However, either prevalence of use of these medications or the incident hyperglycemia as a result of these medications in the US population is fairly rare [42]. Some other influencing factors for HbA_1c that we did not collect include sleep and amount of exercise [43,44]. We also did not account for socioeconomic factors, such as educational level, which might confound the associations seen and the accuracy of the self-reported biometrics, as stated earlier [45]. Additional studies are required to obtain more information about these covariates. We also did not account for the frequency of use of the Foodsmart platform, which could affect the associations found. Finally, due to missing data for several biometrics (such as BMI) and only single values input for HbA_1c, our study had a small sample size relative to the total number of participants who use the Foodsmart platform.

This study also has many strengths. To our knowledge, this is the first study that evaluated the real-life impact of behavior change with online food ordering, diet, and meal planning through a digital intervention and its impact on diabetes and HbA_1c levels. Using Foodsmart’s large user base, this study was able to draw real-world associations between changes in dietary habits and HbA_1c levels and the use of a commercial digital health platform. Furthermore, participants on the Foodsmart platform had a broad range of durations of enrollment; this allowed us to measure changes in HbA_1c over different lengths of time, including time spans of greater than 2 years.

This study evaluated changes in self-reported HbA_1c levels among participants with diabetes who were using a digital nutrition intervention with personalized recipe recommendations, meal planning, food ordering, and grocery discounts and price comparisons. Future research through a randomized controlled trial will be needed to assess the causal effect of the Foodsmart platform on dietary changes and improvements in HbA_1c levels, the difference in cost between pharmaceutical and digital interventions, and which specific components of the dietary score are associated with a reduction in HbA_1c levels.