This study was an administrative claims exploratory data analysis of the previously described c-RCT (NCT01107015) [10,11], where changes in health care utilization during the trial were compared with rates of usage before the trial. Subsequently, the observed changes in health care usage were compared with changes in HbA_1c. In addition to analyzing claims data, BG data were obtained by abstraction from patients’ medical charts [11]. We planned a priori to evaluate utilization but not costs, as cost data would be sparse or skewed in a study of 92 participants [10].

The c-RCT was conducted in 26 primary care practices in 4 distinct geographic areas in Maryland, USA [11]. The c-RCT randomly assigned 26 primary care practices to 1 of 3 stepped treatment groups described below (groups 2-4) or a control usual care (UC) group (group 1) [10,11]. Enrolled patients (n=163) followed their physician randomization assignment. The study population included adult patients aged 18-64 years having a physician diagnosis of T2D at least 6 months prior to study enrollment and HbA_1c≥7.5%. Patients were only eligible for the parent study if their health services were covered by a commercial insurer [10,11]. The primary outcome of the parent study was an absolute change in HbA_1c (percentage point of total hemoglobin) comparing the UC (control) group and the maximal treatment group at baseline and 12 months [10,11].

Group 1 received usual care (UC), group 2 received coaching only (CO), group 3 received coaching and patient care provider portal (CPP), and group 4 received coaching and patient care provider portal with decision support (CPDS) [10,11]. The maximal treatment was a mobile phone–based and Web-based self-management patient CPDS [11]. Providers in the CPDS group had access to analyzed patient data linked to the standard of care and evidence-based guidelines, while providers in the CO group received data from their patients if patients chose to share it. Providers in the coaching and patient care provider portal group received unanalyzed patient data [11].

For the current secondary data analysis, we analyzed all-cause and diabetes-related adjudicated (paid) claims data for 56% (92/163 participants in the parent study) participants in each study group, (UC, n=28; CO, n=17; CPP, n=13; and CPDS, n=34). All-cause utilization was defined as any claims-based health care utilization inclusive of diabetes and any other diagnosis on the claim. We had access to commercial insurance claims data from 2006 through 2010 for study participants covered by the state’s largest commercial insurer. Other participants were covered by multiple commercial insurers and not included in this analysis. Claims data were aggregated into 12 categories: physician office visits, general practitioner, cardiologists, outpatient services, lab claims, prescription claims, endocrinologists, ophthalmologists, podiatrists, emergency department visits, inpatient visits, and total inpatient hospital days. The physician office visits category was defined as available outpatient records for an office visit validated by the commercial insurance company. General practitioners were defined as any medical visits listed as family practice, general practice, or internal medicine. Category classifications (ie, lab vs prescription claims) were confirmed by the insurer.

Four utilization categories were excluded from the analysis because data were sparse: endocrinologist visits, emergency department visits, inpatient hospital visits, and total inpatient hospital days. Ophthalmologist and podiatrist visits were also infrequent. However, due to their importance in standard diabetes care, we included them in revised models. Data were organized by date and grouped into the 12 months prior to randomization (prerandomization period) and 12 months after randomization (postrandomization period). Patient medical charts were used to collect HbA_1c values at baseline and at 3, 6, 9, and 12 months [11].

The Institutional Review Board of the University of Maryland, Baltimore, approved this study. A data and safety monitoring board was designated to review the study procedures and adverse events. After enrollment was closed, errors in consent were found and all participants, both physicians and patients, were asked to sign consent forms again, as recommended by the Institutional Review Board. All patients in the final analysis were reconsented.

Administrative claims data for the participants’ study year were compared with the previous year using generalized linear mixed-effects models to examine the effects of treatment group differences overall, over time, and grouped by time interaction. Specifically, the prerandomization year was the reference year. The treatment effect thus analyzes whether health service visit frequencies were different by group, comparing the year before to the study intervention period. The time by treatment group effect represents the group differential changes from the prerandomization period to the postrandomization period. To address the clustering of physician practices in the parent study, random effects were used to account for within-practice clustering and within-patient correlation.

Due to infrequent ophthalmologist and podiatrist visits (which caused nonconvergence models), these claims’ data were analyzed using repeated measures analysis of variance. Considering the skewed frequency of claims, a Poisson distribution was selected for the outcome.

To examine the impact of utilization on HbA_1c, general linear models were used. Two tests were conducted for each type of visit: 1 model included only a baseline HbA_1c and visit count effect (postrandomization period only) and the other included both of these effects as well as a study group effect. The study group effect examined whether the number of visits, baseline HbA_1c, and group membership (physician office and general practitioner) predicted 12-month HbA_1c. The tests without group-effect modeling examined whether the number of visits and baseline HbA_1c predicted 12-month HbA_1c. The statistical software SAS version 9.3 (SAS Institute, Cary, NC) was used for all analyses. The level of significance was set at ≤.05.

To our knowledge, this administrative claims analysis is the first study assessing a c-RCT mobile phone diabetes intervention’s impact on health service utilization. This study expands on the study by Quinn et al [11] by assessing the intervention’s effect on changes in health service utilization and whether those changes explain the clinically significant change in HbA_1c reported in the primary study. We found that mobile phone-based treatment and behavioral coaching intervention decreased physician office visits, general practitioner visits, and lab claims over time while prescription claims increased. The issue of whether an increase or decrease in claims as a result of the intervention is desirable or undesirable is unclear. Increases in medication prescriptions may be a good outcome, suggesting that appropriate physician intensification occurred and participants were more appropriately taking their medications [17]. Decreases in physician visits and labs saves money but could either indicate less need for care, or if too infrequent, may suggest mHealth participants are inappropriately substituting the in-person care with attention from the intervention.

Our finding of no significant effects of utilization visits to explain the change in 12-month HbA_1c obtained in the parent study [11] may be due to the intervention’s multiple behavior change strategies. In a follow-up mixed methods analysis of patient engagement in the parent study, we learned that some patient behaviors (glucose monitoring, healthy eating, and taking medication) contributed to greater changes in HbA_1c [15]. It may also be that our sample of claims data was too small or too short in follow-up to evaluate the impact of utilization on changes in HbA_1c over a 1-year period.

The generally positive findings reported here need to be put into context with the mixed economic results from diabetes management interventions reported in other studies. Nundy examined the impact of a 6-month mHealth demonstration project among adults (n=74) with type 1 and type 2 diabetes who were members of an academic medical center’s employee health plan. Although those authors observed pre-post improvements in glycemic control (P=.01) and a significant decrease in per person outpatient visits, only 20% of the eligible population participated in the study [19]. Another study found improvements in clinical measures but no impact on health care utilization or cost [20]. None of these studies are directly analogous to this study, which highlights the difficulty in determining successful integration of mobile phone diabetes interventions in clinical settings as a reimbursable service.

The results of our study will inform payers attempting to understand the potential of mobile phone diabetes management technology. Payers have been reluctant to reimburse for mHealth visits, partly due to lack of evidence that mHealth interventions make a difference in utilization and related costs. Payers’ views are short term, not long term, because members may not be enrolled in their plan for the next insured year. The prevention or delay of complications require years. Therefore, a short-term outcome, such as change in HbA_1c as demonstrated in our primary study, as an indicator of prevention may be more appealing to payers than specific utilization effectiveness of an mHealth intervention.

Payers are “experimenting” with health and wellness apps, but current reimbursement payments for mHealth care are largely limited to remote rural areas. Payers have a financial interest in minimizing their risk by actively promoting the health of their policyholders. However, most health care technologies are used in the current fee-for-service (FFS) model, including our evaluation [21]. In a FFS system in which there are only codes for medical devices and human to clinician visits, “there are few, if any, reimbursement codes that exist for frequent high-value patient touchpoints driven by technology rather than humans.” [22]. The potential benefits of mobile phone diabetes care from the perspective of payers may be driven by the transformation of the incentives from pay for service to pay-for-value or performance. Mobile phone diabetes management, such as our intervention focusing on behavioral change based on digital contacts between patients and providers that improves clinical outcomes and utilization metrics, is well-placed in a bundled payment model. The bundled payment model could include a per-patient amount for a bundle of services to be provided, preferably with payment based on agreed-upon clinical outcomes achieved by the mobile phone digital contacts, instead of an FFS payment system based solely on the number of office contacts.

We advise caution in generalizing our findings. Participants in the study were insured by a single commercial insurer and may have experience with and access to resources including individual group practice guidelines, access to specialists, and variations in insurance plan coverage and coinsurance different from the rest of the population with T2D. We attempted to address these differences by enrolling multiple community physicians to participate in the study and randomization at the practice level with patient enrollment following physician randomization assignment. Administrative claims data may not adequately capture service utilization by persons with T2D. For example, diabetes is infrequently the primary diagnosis for emergency department visits or hospitalizations but is often an underlying condition (eg, to myocardial infarction). The analysis was unlikely to see more severe conditions requiring hospitalizations because of exclusion criteria and therefore unlikely to observe high utilizers where diabetes is severe and uncontrolled, (eg, gangrene or kidney failure), although 45% (41/92) of participants had HbA_1c>9% at enrollment. Improvements in utilization, both increases and decreases depending on the health service, may have occurred after the utilization analytic year. Some utilization was required of the study treatment (ie, visiting primary care provider and receiving prescription for HbA_1c tests).

Our program of studies [11,15,17,18,23], including the analysis reported here, demonstrates that a mobile phone diabetes technology achieved a clinically significant change in BG control and that important service utilization increased (pharmacy) or decreased (physician office visits). These findings may help persons with T2D diabetes engage with health service providers and participate in decisions to receive services, lab tests, and prescribed medications recommended by treatment guidelines.