We conducted a retrospective cohort study analyzing data from patients with type 2 diabetes enrolled in the Unika Health program between 2019 and 2024. Each patient was added as soon as they enrolled in the program, and their blood glucose levels were monitored using the Gluco+ Wireless Smart Gluco-Monitoring System and their readings were subsequently accessed on a mobile app. Unika Health is a comprehensive digital health care delivery model that combines RPM, chronic-care management, and structured lifestyle coaching to extend care between physician encounters. The program was implemented in collaboration with 74 physician practices and organizations.

Within this model, an interdisciplinary team—including primary care physicians, registered dietitians, certified diabetes care and education specialists, and care coordinators—conducts initial enrollment, device onboarding, comprehensive baseline assessment, and collaborative goal setting. Physiologic and behavioral data are transmitted continuously; algorithm-defined thresholds prompt targeted outreach, supplemented by regular one-on-one dietitian counseling, physician follow-ups, and laboratory assessments every 3 to 6 months.

To comply with HIPAA (Health Insurance Portability and Accountability Act) deidentification standards, multiple safeguards were implemented to protect participant privacy. Member identifiers were encrypted using asymmetric key encryption, with a hashed version retained to maintain analytical consistency without revealing identity. All personally identifiable information, including names and direct identifiers, was removed. Temporal variables were uniformly shifted to mask exact dates while preserving seasonal and longitudinal patterns relevant for analysis. Numeric ages were converted into categorical groups, and race information was consolidated into broader categories to minimize reidentification risk. Together, these measures ensured secure, ethical, and reproducible handling of sensitive health data.

To be included in the analysis, patients were required to meet the following criteria:

The requirement of at least one before-meal glucose measurement served as a minimum inclusion threshold to confirm device activation and program participation; it was not the engagement exposure itself. Engagement was operationalized separately as each patient’s mean weekly measurement frequency across the 6-month observation period. All readings were captured automatically from Bluetooth-connected glucose meters and were device time-stamped at the point of measurement rather than manually entered, and before-meal status was recorded as a structured field transmitted with each reading.

When multiple HbA_1c results were available within the baseline or end point windows, we selected the one closest to the enrollment date for baseline measurements and the one closest to the 6-month mark for end point measurements. Prestudy HbA_1c measurements were not accessible, as data from prior to enrollment reside in external sources outside of the program’s data infrastructure. Therefore, the first HbA_1c measurement recorded during the study period was designated as the baseline.

Participants were recruited through physician referrals to the Unika Health program. Eligible patients had a diagnosis of type 2 diabetes mellitus and were prescribed SMBG as part of their management plan.

Upon enrollment, participants received Bluetooth-connected glucose monitoring devices and access to a mobile app for tracking their measurements. The program collected several types of data:

The primary outcome was the HbA_1c measurement taken in the end point window. As shown in Figure 1, the study timeline includes a baseline window for the initial HbA_1c measurement, a 6-month observation period for glucose monitoring, and an end point window for the final HbA_1c assessment.

The primary focus of this study was remote monitoring frequency, quantified by the weekly frequency of self-administered glucose measurements. This continuous, time-varying metric served as our treatment variable within the causal inference framework, where higher frequencies indicated greater remote monitoring frequency levels. Monitoring frequency was calculated based solely on glucose measurements captured during the study period, as pre-enrollment monitoring data were unavailable. As this was a retrospective study, engagement was operationalized based on monitoring frequency captured via the connected blood glucose monitoring device; detailed analysis of app interaction features and cross-platform engagement was not conducted and was beyond the scope of this work.

Patient-specific baseline covariates included demographic characteristics (age at enrollment and sex), clinical parameters (BMI category and baseline HbA_1c), comorbidity status, physical activity levels, and lifestyle factors (alcohol consumption and smoking status).

We identified several time-varying confounders based on their potential interactions with remote monitoring frequency: (1) weekly mean before-meal glucose readings, which reflect current glycemic control; (2) program duration in weeks, accounting for stage-specific intervention effects; and (3) weekly medication adherence status and medication type (oral, insulin, or others).

The primary outcome was defined as the HbA_1c value obtained at the follow-up lab visit after the 6-month study period, which provided a direct assessment of glycemic control achieved during the period of observed remote monitoring frequency behavior.

Our statistical analysis began with a descriptive characterization of the study population. First, we summarized baseline cohort characteristics, including demographic distributions and clinical features. To understand patterns of remote monitoring frequency, we then conducted cluster analysis on longitudinal measurement behavior data, which identified 3 distinct profiles: 3 tiers with increasing monitoring frequency from tier 1 to tier 3, where tier 1 represents low engagement, tier 2 indicates medium engagement, and tier 3 corresponds to high engagement.

To estimate the causal effect of remote monitoring frequency on glycemic control, we used a formal causal inference framework that accounts for the complex temporal relationships between variables. The causal graph was developed based on two sources: (1) established frameworks in the marginal structural model (MSM) literature for handling time-varying treatments and confounders [15], and (2) prior applications of MSM in diabetes care research, particularly studies examining medication adherence and glycemic outcomes [16]. While our study focused on remote monitoring frequency rather than medication adherence, we adopted a similar causal structure to account for time-varying confounding in diabetes self-management behaviors.

To address time-varying confounding, we implemented MSMs with inverse probability weighting (IPW). This approach allowed for a more accurate estimation of the causal effect of glucose monitoring on glycemic control, especially when factors that change over time—such as interim glucose readings or medication use—can both influence and be influenced by monitoring behavior [7].

MSMs include 2 main stages. First, we calculated statistical weights for each patient at each time point, reflecting how likely their observed monitoring behavior was, given their prior health status and characteristics. These weights helped balance the study population, making it more comparable to a randomized trial and reducing bias from time-varying confounders. To ensure stability, we limited the influence of extreme weights.

The weighting model included the following baseline covariates, selected for their clinical association with both monitoring behavior and glycemic outcome: age group, sex, BMI category, assigned treatment, baseline comorbidities (hypertension, hyperlipidemia, mixed hyperlipidemia, and diabetes complication categories), smoking status, drinking status, and physical activity level and frequency. The denominator model additionally included time-varying covariates—previous-week mean glucose, program week, and weekly medication status (oral, insulin, and other)—to address time-varying confounding. We assessed covariate balance before and after weighting using standardized mean differences (SMDs), with an SMD<0.1 considered balanced.

After weighting, the strongest measured confounder, baseline HbA_1c, improved in balance (SMD reduced from 0.29 to 0.14), though several covariates retained SMDs above 0.1, indicating residual imbalance. Because balance was not fully achieved by weighting alone, baseline covariates were retained in the outcome model; this doubly robust specification yields a consistent estimate if either the weighting or the outcome model is correctly specified.

In the second stage, we used these weights in a regression model to estimate the effect of monitoring frequency on HbA_1c, adjusting only for baseline characteristics such as age, sex, BMI, comorbidities, and treatment type. We specifically avoided adjusting for variables measured after the start of monitoring, as this could distort the estimated effect.

Because it is not practical to model every possible monitoring pattern, we summarized each patient’s monitoring behavior using their average weekly frequency, following conventions in MSM literature [15]. All analyses were conducted using Python (version 3.8; Python Software Foundation) and the statsmodels (version 0.13.2) package [17].

This study constituted a secondary analysis of preexisting data originally collected by the Unika Health program during routine clinical care and program operations. The data were fully deidentified and anonymized in accordance with HIPAA deidentification standards. The research team at the University of California, Irvine was granted explicit permission to access and use the deidentified dataset for research purposes under a data use agreement with Unika Health, ensuring the data were used solely for the research purposes intended. Informed consent was obtained from all participants as part of their original enrollment in the Unika Health program. Patient privacy and confidentiality were maintained throughout, and no identifying details were published in this manuscript.

A total of 1436 patients with type 2 diabetes or prediabetes comprised the study cohort (Table 1). The cohort had a nearly equal gender distribution (50.1%, 719/1436 female) and was predominantly older, with 95.4% (1370/1436) aged 46 years or older. Nearly half (45.9%, 659/1436) were classified as obese, with an additional 31.8% (456/1436) as overweight. Hypertension was present in 82.6% (1186/1436) of the patients, and hyperlipidemia in 47.7% (685/1436), with 25.9% (372/1436) having mixed hyperlipidemia.

The study cohort had a mean weekly glucose measurement frequency of 4.55 (SD 3.25) measurements (median 3.82, IQR 2.00-6.22). Cluster analysis identified three engagement profiles with distinct measurement patterns: the high-engagement group (7.5%, 108/1436) averaged 12.59 (SD 3.03) measurements/week, the medium-engagement group (34.3%, 493/1436) averaged 6.19 (SD 1.46), and the low-engagement group (58.1%, 835/1436) averaged 2.55 (SD 1.35). Figure 2 illustrates the longitudinal trends in home glucose readings across these engagement groups.

Monitoring frequency was highest in the opening weeks of the program and declined modestly thereafter in all tiers (Figure 3). The high-engagement group began at approximately 10 to 13 measurements per week and stabilized near 11 to 13 measurements per week; the medium-engagement group remained near 6 measurements per week; and the low-engagement group declined from roughly 2.4 to about 1.1 measurements per week. As no pre-enrollment meter data were available, “baseline” frequency refers to the initial weeks of the observation period rather than a preprogram measurement.

Glycemic control improved across all engagement tiers, with a monotonic dose-response gradient in HbA_1c reduction (Table 2). The high-engagement group achieved the greatest mean reduction (1.01, SD 1.67 percentage points; 95% CI 0.69-1.32), followed by the medium-engagement (0.71, SD 1.50; 95% CI 0.58-0.85) and low-engagement (0.38, SD 1.40; 95% CI 0.29-0.48) groups. All three within-patient reductions were statistically significant (all P<.001, 2-tailed paired t test), with 95% CIs excluding zero. The high-engagement group also presented with the highest baseline HbA_1c (mean 8.04%, SD 1.76), consistent with greater room for improvement.

Among patients with high baseline HbA_1c (≥9%; n=248), engagement patterns showed even more pronounced effects on glycemic control (Table 3). The high-engagement group demonstrated the highest control rate, defined as achieving HbA_1c<7% at the 6-month end point (36.4%, 8/22), and the greatest mean reduction in HbA_1c (3.12, SD 1.97 percentage points). In contrast, the low-engagement group showed the lowest control rate (21.8%, 31/142) and the smallest mean reduction (2.17, SD 1.98 percentage points).

The treatment model produced stabilized weights with a mean of 1.05 (SD 0.46) and a median of 1.03 (IQR 0.87-1.16; range 0.09-12.94), indicating a well-behaved distribution centered near 1.0. The treatment model achieved an R^² of 0.63.

To examine whether the effect of monitoring engagement on glycemic control is consistent across different end point time points, we additionally analyzed outcomes at 3 months and 12 months. Table 4 presents the glycemic outcomes by engagement tier at each time point.

The dose-response relationship between monitoring frequency and glycemic control remained consistent across time points. At 3 months, each additional weekly measurement was associated with a 0.04 (95% CI 0.03-0.05) percentage point reduction in HbA_1c (β=–0.04; P<.001), and at 12 months with a 0.03 (95% CI 0.01-0.06) percentage point reduction (β=–0.03; P=.003).

Our analysis of 1436 patients with type 2 diabetes and prediabetes enrolled in the Unika Health RPM program demonstrated that higher glucose monitoring adherence is causally associated with substantial improvements in glycemic control, as measured by HbA_1c reduction. Consistent with our primary objective of quantifying the impact of varying engagement levels on glycemic outcomes in a real-world setting, we identified a clear dose-dependent relationship: each additional weekly glucose measurement was associated with a 0.05 percentage point decrease in HbA_1c (P<.001). Patients in the highest engagement group (mean 12.59 measurements/week) experienced a mean HbA_1c reduction of 1.01 percentage points, compared with 0.38 in the lowest engagement group (mean 2.55 measurements/week)—a difference of 0.63 percentage points—with the intermediate group showing a progressively greater improvement (Table 2). These findings were robust after adjustment for baseline confounders using MSM with IPW.

The observed dose-response relationship underscores the importance of sustained patient engagement in digital health interventions for diabetes management. Prior studies evaluating RPM programs have reported HbA_1c reductions of 0.28 to 0.54 percentage points compared with usual care [7-10]; however, these studies have largely treated engagement as a binary variable—enrolled versus not enrolled—rather than characterizing the gradient of active participation. Our work provides the first rigorous quantification of how incremental increases in remote monitoring frequency translate to measurable glycemic improvements in a large real-world cohort, advancing beyond general program efficacy to isolate the specific contribution of engagement intensity.

Notably, even patients in the low-engagement group experienced a meaningful 0.38 percentage point reduction in HbA_1c at the 6-month end point. This finding suggests that enrollment in a comprehensive, primary care–integrated RPM program may confer benefits beyond the direct effects of glucose monitoring frequency alone. The increased self-awareness, behavioral modifications, and access to personalized health coaching associated with program participation likely contribute to positive outcomes independent of monitoring frequency, consistent with theoretical frameworks of self-management support in chronic disease [18,19].

The engagement patterns observed in our study—ranging from 2.55 to 12.59 measurements per week—reflect the wide variability in SMBG adherence documented in national surveys, where many patients monitor less frequently than clinically recommended [20]. Our results suggest that even within this variability, a graded and meaningful glycemic benefit exists, reinforcing the clinical value of supporting any level of engagement while prioritizing strategies to increase monitoring frequency.

Our adjusted models further confirmed that baseline HbA_1c, age, treatment modality, and the presence of complications were independently associated with glycemic control during program participation (Table 3) [21,22], consistent with established diabetes risk profiles. Additionally, longer program duration was independently associated with a modest but statistically significant decline in HbA_1c, suggesting that sustained participation yields cumulative benefits beyond initial engagement levels.

Collectively, these findings indicate that the intervention’s impact derives from both the supportive, primary care–integrated RPM structure and the degree of active patient engagement. The intervention’s grounding in established clinical workflows and physician oversight aligns with current guidelines for diabetes management [23], and its remote delivery model extends access to ongoing, proactive guidance in real-world settings.

Several limitations should be acknowledged. First, despite adjustment for a broad set of confounders and the use of MSMs with IPW, unmeasured confounding cannot be completely ruled out. Factors such as diabetes education, health literacy, psychological factors, social support, and health care access may influence both monitoring frequency and glycemic outcomes. Additionally, we lacked complete patient histories and comorbidity data due to information residing in disparate electronic medical records to which we did not have direct access.

Second, our measure of engagement captured monitoring frequency but not the timing, quality, or clinical response to individual readings. We could not assess whether patients monitored at optimal time points or acted upon their glucose values. While digital glucose meters were used, the possibility of measurement error or data entry issues could not be entirely excluded.

Third, while medication status across oral agents, insulin, and other therapies was incorporated as a time-varying covariate in the weighting model, we did not have sufficiently granular data on therapy intensification or regimen switching within each engagement tier. Residual confounding by unmeasured treatment changes therefore cannot be excluded.

Fourth, measured engagement with the monitoring program may also correlate with other unmeasured positive health behaviors—such as increased physical activity or dietary improvement—that could independently contribute to the observed outcomes, precluding full attribution of glycemic benefit to monitoring frequency alone.

Fifth, our study population consisted of patients enrolled in a digital health program, who tend to be younger, more educated, and more technologically comfortable than the broader diabetes population. Generalizability to RPM programs with substantially different structures, intensities, or patient populations may therefore be limited.

Our findings demonstrate that higher engagement with a digitally enabled, primary care–integrated RPM program is causally associated with substantially greater glycemic improvement among patients with type 2 diabetes and prediabetes. Beyond confirming program efficacy, this work highlights that the magnitude of benefit is meaningfully modifiable through engagement intensity—a finding with direct clinical and policy relevance. Given that suboptimal glycemic control remains a significant public health challenge, and that diabetes is a potent modifiable risk factor for cardiovascular disease, kidney disease, neuropathy, and retinopathy, scalable strategies that both expand access to integrated RPM programs and actively foster sustained patient engagement have the potential to meaningfully reduce the burden of diabetes-related complications at a population level.

Future work should extend beyond intermittent glucose monitoring to continuous glucose monitoring (CGM), which enables real-time, granular tracking without the limitations of single-point measurements. Rigorous clinical trials validating CGM integration within coordinated care delivery models—including hospital and clinical settings—are warranted. Additional research directions include leveraging AI-enabled devices for targeted behavioral interventions and analyzing text- and video-based coaching interactions to enable increasingly personalized care. Such advances, embedded within the supported self-management framework demonstrated here, may further amplify the benefits of digitally enabled diabetes care.