This study utilized the Dario digital health platform (Dario Health) for chronic conditions to support self-management of diabetes. The Dario platform combines a blood glucose meter with a phone app that is available for both Android and iOS devices (Figure 1). The glucose meter consists of a small, pocket-sized holder for strips, a lancet, and the meter. The blood glucose meter is removed from the holder and plugged directly into a smart mobile device, effectively converting the smart mobile device into a display screen for the meter. All data were transferred and stored in compliance with the Health Insurance Portability and Accountability Act requirements using Amazon Web Services database solutions.

The monthly average BG (AvgBG) was defined as the means of all of a user's BG measurements taken over a 30-day interval for each user starting from the first 30 days after the initial measurement. Additionally, all collected measurements were within the valid range according to device specifications (43). It was used as the core outcome metric to reflect the monthly aggregated interval changes over 12 months. Demographic and medical information were collected for each user: Age, gender, BMI, weight, ethnicity, year of diabetes diagnosis, insulin treatment, physical activity level, stress level (0 = no stress and 10 = very stressed), alcohol consumption (number of drinks per week), smoking (0 = never and 3 = yes), and number of reported comorbidities. Digital activities were also added, such as the number of BG measurements and lifestyle tags (the sum of lifestyle activities such as: meal taggings, meal reference tagging, carbs taggings and calories burned). Those variables were analyzed vs. baseline levels in a monthly aggregation.

The inclusion criteria were Dario platform users with reported T2D in high-risk range (44–46)

[first 30 days with AvgBG ≥180 mg/dL equivalent to A1c 8.0 (47)] who measured their BG on the Dario platform between 2020 and 2024 for a minimum of two separate months consecutive or non-consecutive (N = 22,414). The users purchased the device via a direct-to-consumer (D2C) channel. Exclusion criteria: registered users that did not measure BG with Dario between 2020 and 2024 for at least two separate months (insufficient data for analysis); Out of specified range measurements (43); users did not report diabetes type 2; users with AvgBG lower than 180 mg/dL in the first 30 days after the initial measurement.

This study is based on an analysis of an existing database already collected in the digital platform. All data were anonymized before extraction for this study. Ethical and Independent Review Services, a U.S.-based professional review board, issued the institutional review board exemption for this study (#23517) (48).

We applied a retrospective longitudinal cohort study design utilizing a database Dario Health D2C population. Data were obtained from the user population defined above that included their age, weight, height (BMI), duration of known diabetes and reported insulin usage.

A classical linear longitudinal model assumes a single-slope growth pattern for changes in an outcome variable across time. Sometimes, such a simple model does not fit the empirical data. In contrast, piecewise-based mixed-effects models allow exhibiting different linear trends across time. Here, based on the visualization of the distribution of the monthly average BG, which revealed a notable change in the trend around the 4-month mark a piecewise linear mixed-effects model was used to define the time-related fluctuations and allow flexibility in the modeling of variable change trajectories across time. The decision was primarily guided by the observed pattern in the visualization and supported by previous analyses conducted on the digital health platform (49, 50).

Next, a Generalized Linear Mixed Effects Tree analysis was used (51), allowing for the identification of different relationships in subgroups of data while accounting for both fixed and random effects in longitudinal data. The tree algorithm learns where each node is associated with different fixed-effects regression coefficients while adjusting for random. This allows the detection of users' subgroups with varying fixed-effects parameter estimates, keeping the random effects constant throughout the tree. The tree algorithm was incorporated with piecewise parameters to detect subgroups with different piecewise trajectories according to the potential moderating factors in 3 different models- the demographic moderating factors included age, gender, BMI and ethnicity; Medical factors included the number of comorbidities, insulin usage, BMI and year of diagnosis; and the BG monitoring moderating factor included only the monthly number of BG measurements. The grouping decision is rooted in the theoretical and clinical relevance of each factor within its respective context. The inclusion of BMI in both models is based on its distinct conceptual role: as a baseline demographic characteristic and as a current medical factor allowing us to explore its moderating effects from different perspectives. This approach may assist in identifying a specific nonlinear trajectory for the subpopulations characterized by the moderating factors.

Overfitting occurs when a machine learning model learns not only the underlying patterns in the training data but also the noise or random fluctuations, resulting in poor generalization to new, unseen data. This is a common issue, particularly with complex models relative to the amount of data available. To mitigate overfitting, the data set was randomly partitioned (by users) into 80% training set and 20% test set- which will be used to measure the predictive validity of the chosen model during the training.

All tree models are built using the lmertree function from R package “glmertree” (52, 53), under the assumption of random intercept for the users, meaning that the model considers that there is a difference in outcome between individuals. Modeling with a random intercept assumes that each individual has a unique baseline outcome, capturing inherent differences between users. This accounts for variability across individuals while focusing on the overall trajectories.

Pruning in decision tree models is a crucial technique for reducing model complexity and enhancing generalization. The pre-pruning process was deployed as all models held “minsplit” (the smallest number of observations in the parent node that could be split further) of 100, and the “maxdepth” (the maximum distance between the root and any leaf) to ensure no more than three levels of decision nodes and by that to force the model to focus on the most important splits and avoid overfitting to minor data patterns. Together, these parameters help to balance between underfitting and overfitting, ensuring that the model remains both interpretable and robust. Five-fold cross-validation for monthly average BG levels were deployed on the training data. Finally, the models were tested with an untouched (test) data set, whereas the assessment of the model's performances was evaluated with root mean square error (RMSE), comparing RMSE based on the training and test data.

Gender, BMI and ethnicity did not significantly moderate monthly average BG fluctuations. For these non-significant results, the effect sizes and confidence intervals indicate that the corresponding P-values are likely greater than 0.05, especially if the confidence intervals include zero. These findings highlight the broad applicability of digital health interventions across diverse demographic groups. The decision tree segmented users into three age subgroups (node 3: ≤35, node 4: 36–60, node 5: >60; Figure 3) and each exhibited a significant improvement in monthly average BG (mg/dL) in the first 4 months (node 3: B = –5.54, 95% CI −7.45 to −3.64, p ≤ .001; node 4: B = –6.50, 95% CI −6.99 to −6.01, p ≤ .001) greater in the older age group (node 5: B = –7.15, 95% CI −7.69 to −6.60, p ≤ .001).

Over the following 8 months, ≤35 age group showed no change in monthly average BG (node 3: B = 0.16, 95% CI −0.76 to 1.08, p = .74), 36–60 group had a slight reduction (node 4: B = −0.26, 95% CI −0.48 to −0.04, p ≤ .01) and >60 group showed the most significant improvement (node 5: B = −0.41, 95% CI −0.64 to −0.18, p ≤ .001).

Following the decision tree's results another LME model was evaluated. In this model, the interaction between the 3 age categories and lifestyle tags (the sum of meal taggings, carbs taggings and calories burned) was tested. It was found that users in the age category >60 demonstrated a greater increase in lifestyle tags during the first 4 months (B = 2.81, 95% CI 0.38 to 5.25, p = .02) compared to age group ≤35 (Figure 4). The interaction wasn't significant when comparing the age group of 36–60 to ≤35 (B = 0.87, 95% CI −1.54 to 3.29, p = .49), meaning that they were statistically equivalent.

To assess the robustness of our findings to key modeling assumptions, we conducted comprehensive sensitivity analyses testing alternative model specifications, data inclusion criteria, and analytical approaches. We tested alternative piecewise breakpoints (months 3, 4, 5, 6), different age and BMI modeling approaches (continuous, quartiles, categories), outlier removal (excluding BG >500 or <50 mg/dL), consecutive vs. non-consecutive months, different baseline BG levels, and different training/test splits (70/30 to 85/15). The 4-month breakpoint (months 1–4) provided the best model fit among the tested alternatives (lowest AIC score), confirming this as the best-fitting specification (slope months 1–4: −6.80 mg/dL/month, p < .001; slope months 4–12: −0.33 mg/dL/month, p < .001). Alternative breakpoints (3, 5, 6 months) also showed statistically significant results for months 1–4, but with inferior model fit. Age was remodeled into categories (≤35, 36–60, ≥60) and quartiles. All age groups showed similar improvements during months 1–4 across different age modeling approaches, indicating that initial response was consistent regardless of age. Age significantly moderated sustained improvements during months 4–12 (quartiles: B = −0.423, p = 0.041; categories: >60 years, B = −0.41, p < .001), with older users showing greater sustained improvements, validating the decision tree findings. BMI was remodeled into categories (Normal<25, Overweight 25–30, Obese I 30–35, Obese II 35–40, Obese III >40). Results showed similar improvements in all categories for both time periods (all p > 0.05).

Results were highly consistent across alternative data inclusion criteria and model specifications. Outlier removal (excluding 0.9% of data) produced nearly identical estimates (both segments p < .001). Analyses restricted to consecutive months only (n = 9,887 users) showed stronger but consistent effects (months 1–4: −8.81 mg/dL/month; months 4–12: −0.80 mg/dL/month, both p < .001), suggesting that sustained engagement is associated with better outcomes. When testing different training/test splits (70/30 to 85/15), slope estimates for months 1–4 ranged from −6.64 to −6.82 mg/dL/month (all p < .001), and test RMSE remained stable (73.1–73.5 mg/dL).

Results remained statistically significant and directionally consistent across all alternative specifications, demonstrating that conclusions are not dependent on specific modeling choices or data inclusion criteria. The pattern of stronger effects in more engaged users (those with consecutive data) is expected and clinically meaningful, as it suggests that sustained engagement is associated with better outcomes. This pattern aligns with our primary finding that monitoring frequency moderates outcomes.

The study leveraged ML methods to identify user-level factors influencing blood glucose levels among people managing diabetes using digital health platforms. A significant reduction in monthly average BG was observed during the first four months of engagement, with a milder decrease in subsequent months.

The magnitude of blood glucose reductions observed in this study is clinically meaningful and comparable to established benchmarks for behavioral and pharmacologic interventions. During months 1–4, mean blood glucose declined by 28.7 mg/dL, which corresponds to an estimated reduction of approximately 1.0% in HbA1c, based on validated conversion formulas (47). Reductions of this size are similar to those achieved with first-line glucose-lowering therapies in early treatment escalation (54, 55–57). Even the smaller additional decline of 8.4 mg/dL observed during months 4–12 corresponds to an estimated 0.3% reduction in HbA1c, which remains clinically significant. Evidence from large cohort studies and meta-analyses demonstrates that HbA1c improvements as modest as 0.3%–0.5% are associated with meaningful reductions in microvascular complications, improved treatment stability, and lower healthcare utilization (58, 59). Importantly, the sustained improvement during months 4–12 reflects maintenance rather than regression an imperative outcome in diabetes management, as behavioral benefits often diminish over time in real-world settings. These outcomes reinforce the potential of digital behavioral interventions to deliver health benefits comparable to established therapeutic approaches and highlight their value as an adjunct to routine diabetes care (60, 61).

The findings highlight how user characteristics such as age, insulin use, disease duration, and frequency of blood glucose self-monitoring can moderate glycemic changes over time. These factors, in combination with lifestyle tags that reflect digital engagement, interact to influence overall outcomes. Particularly, older adults (>60 years), users with frequent BG monitoring (>12 times/month), and those using insulin with a shorter disease duration (<5 years) showed the most substantial improvements in BG control.

Our results highlight two critical insights. First, the most substantial improvement in average BG occurred during the first four months of platform use across all user subgroups, consistent with earlier findings showing an initial engagement-driven benefit from digital self-monitoring tools (8, 10). Second, older adults (≥60 years) and users who engaged more consistently in lifestyle tags within the app such as meal tagging, meal reference tagging, carbohydrate tagging, and tracking calories burned, showed more sustained improvements in blood glucose throughout the year. This suggests that consistent self-monitoring is a key factor in long-term glycemic management.

This study is subject to several limitations. The observational design limits causal inference: associations between engagement patterns and BG improvements may be influenced by unmeasured confounders such as user motivation or socioeconomic status. Although the models included key demographic and behavioral variables, psychosocial factors such as mental health status, digital literacy, or readiness for change were not available and could also impact the results. Selection bias may also be present, as the sample consists of direct-to-consumer (D2C) users who voluntarily opted into the program, potentially limiting to broader patient populations. Additionally, the absence of clinical gold-standard measures such as HbA1c limits our ability to assess glycemic control based on laboratory tests. Attrition bias is another concern, as participants with incomplete data may differ systematically from those who remained engaged, influencing the observed outcomes. In the present study, more than one-third of users had available measurements at month 12, a level of data availability comparable to that reported in prior long-term digital and mobile health studies. In addition, sensitivity analyses demonstrated that the primary findings were robust to alternative data inclusion criteria. Observational nature introduces potential temporal confounding, where changes over time may be influenced by external factors unrelated to the intervention.

The absence of a control group limits comparisons, making it difficult to attribute outcomes solely to the intervention rather than natural disease progression or external influences. Lastly, potential multicollinearity among several predictors in the machine learning models should be acknowledged. Factors such as age, insulin use, diabetes duration, BMI, and engagement frequency are known to be interrelated in the context of diabetes management. For example, longer disease duration is often associated with a higher likelihood of insulin therapy, and younger age groups may demonstrate different engagement patterns. ML approaches such as tree-based models used in this analysis are generally robust to correlated predictors, as they split the data using the most informative variables and assess each feature's contribution through hierarchical partitioning rather than assuming linear independence. Future studies with larger samples and explicit modeling of interaction terms, or dimensionality-reduction techniques, may help clarify the independent vs. synergistic contributions of these variables.