One Drop is a multi-condition mHealth solution that can be tailored to each user’s unique needs and preferences. The One Drop digital platform targets people with prediabetes, T1D or T2D, high blood pressure, high cholesterol, or combinations of these conditions. The One Drop mobile app can be used standalone or in conjunction with a monthly or yearly blood glucose test strip subscription, Bluetooth-enabled One Drop blood glucose meter, or connected devices (eg, Wi-Fi–enabled weight scales or Wi-Fi–enabled smart blood pressure monitors).

The One Drop mobile app is available for devices running the iOS, Android, or WatchOS operating systems. Uses can opt to enroll in the free or premium versions of the app. The free version includes extensive logging functionality with the capacity to log health data (eg, blood glucose, blood pressure, HbA_1c, and weight), intensity and duration of exercise, food eaten, and medications prescribed and taken. When users begin logging data actively (ie, manually through the app or with a synced One Drop meter) or passively (via integrations with Apple HealthKit or Google Fit), detailed reports and visualizations become available, displaying summaries of the entered data and blood glucose trends over time. Messages of support, health-related insights, and educational content are also delivered to users’ in-app inboxes or newsfeeds. The premium One Drop app subscription additionally provides users with on-demand access to health coaching with certified health professionals specializing in their conditions, machine learning–powered trends and forecasts, adjustable goal setting, and a personalized content library with hundreds of lessons.

Blood glucose forecasts were introduced to the One Drop app as a free feature for users with T2D in September 2018 and became a premium feature in August 2020. These forecasts project the direction (ie, up, steady, or down) that a user’s blood glucose will trend in the following 1 to 8 hours. Users automatically begin receiving blood glucose forecasts upon recording their first blood glucose reading and continue to receive them after each subsequent recording. When a forecast is generated, users receive a pop-up notification indicating the direction (rising, steady, or falling blood glucose) and duration (1-8 hours) of the forecast. These notifications can be paired with an actionable suggestion to help maintain healthy blood glucose levels. Figure 1 shows an example of a blood glucose forecast in the One Drop app.

Users who joined One Drop prior to September 2018 or did not have a paid subscription after August 2020 did not have access to the blood glucose forecasts. After forecasts were implemented in the program, users could become ineligible to receive forecasts if they failed to meet any of the prediction algorithm’s requirements. Users did not receive forecasts if (1) they had extremely high glucose (>600 mg/dL); (2) they had extremely high glucose variability (>80 mg/dL forecast standard error); (3) their logging frequency was consistent with continuous glucose monitor (CGM) use (>50 readings in a 24-hour period); (4) they had ever recorded bolus insulin use; (5) they recorded basal insulin use prior to March 2019 (ie, the time when the model began serving predictions to these users).

This study employed a retrospective cohort design, evaluating real-world outcomes over a 12-week period; participants were thus not recruited for study participation. There were 2 cohorts that were compared after statistically controlling for available potential confounders. The first cohort received at least one blood glucose forecast in weeks 1 through 11 and the second received no forecasts. Those included in the group who received no forecasts did not have a paid subscription after August 2020 or were ineligible to receive blood glucose forecasts based on the exclusion criteria.

On August 10, 2021, One Drop users with T2D who had one or more blood glucose measurements recorded 12 weeks following their first recorded reading were identified. Blood glucose readings were entered either manually through the app or passively through the One Drop blood glucose meter, another synced device, or integration with Apple Health Kit or Google Fit. The query for eligible users was limited to those who had started using One Drop and recorded their blood glucose for the first time in 2019 or later to minimize the possibility of comparing groups that participated in different iterations of One Drop. Additionally, only users with at-risk baseline blood glucose (estimated HbA_1c≥7%) were queried. Users with a glucose logging frequency consistent with CGM use were excluded. The baseline measurement time point was defined by averaging the first 7 days of blood glucose readings, beginning with the first recorded blood glucose reading for that individual. Follow up (in week 12) consisted of measurements recorded from days 77 to 83.

One Drop received an exemption for institutional review board approval and a waiver of informed consent from Solutions IRB, an independent ethics review company (Little Rock, AR and Yarnell, AZ) to study all deidentified data owned by One Drop. User data are stored in a secure cloud-based server. All One Drop users must actively agree to an end user license agreement upon creation of their accounts, granting One Drop permission to use data entered in the app for analysis, reporting, and research purposes.

There were 1411 users included in the analysis. Users were deemed eligible for analysis if (1) they reported a diagnosis of T2D; (2) their first week of One Drop participation was between the years 2019 to 2021; (3) they recorded ≥3 blood glucose measurements in their first week; (4) they recorded ≥1 blood glucose measurement in week 12; (5) they had a self-reported year of diagnosis; and (6) their calculated week-1 HbA_1c was 7.0%.

Users averaging more than 7 blood glucose measurements per day for the first 11 weeks of the study period were assumed to be using a CGM (>539 measurements; 7 measurements times 7 days times 11 weeks) and were excluded.

All analyses were performed using SPSS Version 28 (IBM Corp). Our predetermined α level was .05 for all statistical tests. Between-group differences in the year users started One Drop, age, gender, years diagnosed with T2D, insulin use, number of week-1 blood glucose recordings, sum of week-1 to week-11 blood glucose recordings, baseline average glucose, and baseline individual blood glucose variability were assessed. Differences were tested with 2-tailed independent-samples t tests for continuous variables and chi-square tests for categorical variables. Descriptive statistics were also computed for these variables. A gender of “other” was reported by 4 users in the group who did not receive forecasts and zero users in the group who received forecasts; this value was treated as missing and excluded from the chi-square analysis assessing between-group differences in gender.

Variables were omitted from models if high missingness was observed (greater than 50%), but descriptive statistics for the variable are still reported. Age was the only variable meeting this threshold (1143/1411, 81% missing).

We specified 2 analysis of covariance (ANCOVA) models to determine the forecast group effect on blood glucose outcomes. Covariates were selected for ANCOVAs if their baseline values were significantly different in the groups who received and did not receive forecasts. The first ANCOVA tested the group effect on week-12 average glucose, controlling for years diagnosed, insulin use, baseline blood glucose variability, and baseline average glucose. The second ANCOVA tested the group effect on week-12 %PIR, controlling for years diagnosed, insulin use, baseline blood glucose variability, and baseline %PIR. Interactions between the group variable and covariates were tested; significant interactions were held and interpreted, while nonsignificant interactions were dropped from the final reported models.

The secondary analysis sought to find mechanisms through which exposure to blood glucose forecasts resulted in greater reductions in average glucose concentration in week 12; it was hypothesized that blood glucose logging could be one such mechanism. This hypothesized model is illustrated in Figure 2. To test this hypothesis, 3 separate linear regression models were specified to establish that the direct effect of group on blood glucose logging behavior and the direct effect of group and logging behavior on week-12 average glucose were all significant.

PROCESS is a free software add-on for SPSS that includes over 70 predefined models [32]. Mediation models in PROCESS incorporate ordinary least squares regression and estimate indirect effects and their confidence intervals through a bootstrapping procedure that is robust against nonnormal sample distributions [33]. Model 4 was specified to estimate the indirect effect of group on week-12 average glucose via blood glucose logging with 5000 bootstrap samples. All models included years diagnosed, insulin use, baseline blood glucose variability, and baseline average glucose as covariates.

The ANCOVA revealed significant mean differences between the groups in week-12 average glucose when controlling for baseline average glucose and covariates (F₇=40.75, P<.001). All model covariates except insulin use (P=.16) were significant (years diagnosed, baseline average glucose, and baseline blood glucose variability; all P<.001). Additionally, significant interaction effects for group × baseline average glucose (F₁=5.28, P=.02) and group × baseline blood glucose variability (F₁=15.12, P<.001) were observed.

To explore the group × baseline average glucose interaction, ANCOVA models were specified to evaluate the group effect at different levels of baseline eHbA_1c. Baseline average glucose was split into eHbA_1c risk levels for interpretability. This interaction is visualized in Figure 3. Users receiving forecasts ended week 12 with significantly lower average glucose than those not receiving forecasts when baseline eHbA_1c was ≥10% (mean difference –54.52mg/dL; –1.9% eHbA_1c, P=.002). Nonsignificant results within the other three categories may be attributed to low power (1-β<.4). Among those with ≥8% eHbA_1c, reductions in week-12 eHbA_1c ranged from –0.86% to –1.9%.

The second interaction, group × baseline glucose variability, was evaluated by performing a median split on baseline blood glucose variability and specifying ANCOVA models for each level. Values below the median were categorized as “low variability” and values above the median were categorized as “high variability.” Values at the median were pruned. Among users with high baseline blood glucose variability, those receiving forecasts experienced significant reductions in week-12 average glucose (mean difference –53.22mg/dL; –1.85% eHbA_1c, P<.001) relative to those not receiving forecasts. Figure 4 visualizes the group × baseline glucose variability interaction.

The ANCOVA revealed a significant mean difference between groups in week-12 %PIR when controlling for covariates (F₆=40.27, P<.001). All model covariates except insulin use (P=.3) were significant (years diagnosed, baseline %PIR, and baseline blood glucose variability; all P<.001). Additionally, a significant interaction effect between group × baseline %PIR (F₁=4.84, P=.03) was observed.

The group × baseline %PIR interaction was evaluated by performing a median split on baseline %PIR and specifying ANCOVA models for each level. Values below the median were categorized as “low %PIR” and values above the median were categorized as “high %PIR.” Values at the median were pruned. Among users with low baseline %PIR, those receiving forecasts experienced a significant increase in %PIR (mean difference .45%, P<.001) compared to those not receiving forecasts. A visualization of the group × baseline %PIR interaction is presented in Figure 5.

As reported above, the results suggest that users receiving blood glucose forecasts experienced greater reductions in week-12 average glucose than those not receiving forecasts. We thus proceeded with our secondary analysis.

The forecast group was significantly and positively associated with total blood glucose logs (path a; b=20.72, t₁₄₀₅=2.86, P=.004). Total blood glucose logs were significantly associated with week-12 average glucose (path b; b=–0.13, t₁₄₀₅=–6.86, P<.001). The total effect of group on week-12 average glucose was also significant (path c; b=–21.74, t₁₄₀₅=–4.27, P<.001). As all 3 paths were significant, we proceeded by regressing week-12 average glucose on group, controlling for total blood glucose logs. The direct effect was reduced, but remained significant (b=–19.22, t₁₄₀₄=–3.82, P<.001).

Results from the bootstrapping procedure produced an estimated indirect effect (path c’) of group on week 12 average glucose through total blood glucose logs with a 99% CI that did not include 0, indicating a significant mediation effect. While the indirect effect was significant, the direct effect also remained significant, indicating that blood glucose logging is a partial mediator of the relationship. Mediation results are summarized in Table 2. The conceptual model is presented in Figure 6, along with unstandardized path coefficients.

The data generated for this study were collected from users of an mHealth app, and thus have real-world generalizability; however, the rapid evolution of the product, the available data, and a lack of experimental controls are limitations. First, there may have been covariates to consider that were not available for analysis, such as age, race, ethnicity, socioeconomic status, health motivation, CGM use, and other factors that may have differed across the groups and impacted outcomes. The impact of these variables on the relationship between blood glucose forecasts and diabetes outcomes is a potential avenue for future research.

Additionally, the groups systematically differed at baseline due to prespecified criteria that excluded some users from receiving forecasts. Those who received forecasts had fewer years diagnosed with T2D, lower glucose, higher %PIR, lower glucose variability, and were less likely to be taking insulin than those who did not receive forecasts. While we controlled for these variables, the bias inherent to this study design may still have been present. The results should be interpreted with this potential for selection bias in mind.

Finally, because the One Drop users included in this study participated at different times over a three-year period, it is possible that they participated in different iterations of the One Drop app that differentially impacted their app experiences, creating potential confounders. The One Drop app is continually updated and improved based on clinical science, behavior science, and research performed internally and externally. Further, participation in this study was limited to 12 weeks. While other studies of mHealth interventions have evaluated glucose outcomes at 3 months [37,38], insight into the sustained impact of forecasts on glucose beyond a 12-week period is limited. To address these limitations, which are characteristic of real-world evidence, future long-term, prospective randomized studies are needed to confirm the causal impact of forecasts on self-monitoring behavior and glucose management.

While the study design and the nature of real-world data likely introduced bias in the results, real-world studies confer unique benefits that extend beyond the confines of a controlled study. Studies using real-world data exchange the internal validity of an RCT for external validity [39], allowing our results to be generalized to other populations with T2D using mHealth apps. The stringent requirements for inclusion in RCTs may exclude participants that would normally be seen in a real-world clinical setting [40]. When used in conjunction with evidence-based clinical practice, real-world evidence has shown that mHealth apps can lead to significant improvements in glycemic management over the course of 1 year [41]. Aside from supporting an existing care network, mHealth apps may also have a niche in providing care for hard-to-reach populations [42].

A major innovation in data-powered health insights is the predictive modeling of specific outcomes, such as blood glucose levels. Although predictive models do not necessarily reveal causes and effects, these models have been used for discovery, hypothesis testing, risk prediction, and the identification of counterfactuals and effective interventions [43]. Despite promising evidence on machine learning models, such as one study that demonstrated a significantly reduced glycemic response when a machine learning model was paired with a dietary intervention, blood glucose forecasts remain a nascent technique. There are currently no agreed-upon protocols for machine learning models in precision health [44,45], and a number of different models have been used as frameworks for machine learning training and development [46]. Further, there is no well-defined approach to estimate carbohydrate intake, the effect of stress and activity on blood glucose level, or the portability of machine learning models to capture inter- and intraindividual variability [45,47].

Until now, the majority of blood glucose forecast research has been focused on T1D [48]. This has limited the scope and real-world potential for blood glucose forecasts, as most (90.9%) diabetes cases in the United States have been diagnosed as T2D [49]. Blood glucose forecast research for T2D has mainly focused on forecasting the future incidence of disease or adverse glycemic events [50]. In addition, blood glucose forecast studies for either T1D or T2D have largely been limited to a short forecast horizon of 2 hours or less [48]. The current study advances the literature in studying a forecast horizon of 8 hours in an at-risk T2D population.

One Drop is one of the first mHealth apps to provide blood glucose forecasts to its users; our findings are the first to provide evidence for the effectiveness of delivering blood glucose forecasts as part of an mHealth intervention. The results suggest that exposure to blood glucose forecasts may be effective for individuals with T2D who have a high level of blood glucose, with forecast exposure associated with reduced glucose, a higher percentage of blood glucose %PIR, and increased self-monitoring of blood glucose. Further, blood glucose logging was a partial mediator of the relationship between forecast exposure and glucose reduction, highlighting a potential mechanism through which forecast exposure is associated with reduced glucose. Taken together, this novel evidence highlights the potential for One Drop blood glucose forecasts to improve glycemic management in individuals with T2D.