Depression, an underdiagnosed comorbidity for those with chronic illness [1], impairs functional status and worsens clinical outcomes, including morbidity and mortality; it also increases cost [2-5]. Diabetes doubles the risk of depression relative to the general population; 10% to 15% of adults with diabetes also have major depressive disorder [6,7]. The relationship between diabetes and depression might be bidirectional [8,9]. Depression with diabetes may significantly worsen the course of both disorders, leading to reduced functioning and quality of life [8-11].

Low-income, culturally diverse populations with chronic illnesses have an even higher risk of depression [7,12,13]. Hispanics and Latinos have a higher prevalence of diabetes compared with non-Hispanic whites [14] and are less likely to meet glycated hemoglobin (HBA_1c) and serum cholesterol goals [15]. Racial and ethnic minority populations also experience disparities in terms of mental health care, including appropriate mental health diagnosis, counseling, antidepressant medication prescriptions, and depression care follow-up [16-23]. Hispanics and Latinos are less than half as likely as non-Hispanic whites to receive guideline-level depression care [20].

Research has shown that there are effective ways to reduce these disparities. For example, there is growing evidence that a team-based collaborative depression care model is effective in improving care for low-income patients (including racial and ethnic minority populations) with chronic illnesses [24-27]. The US Preventive Services Task Force recommends depression screening, and an adaptive treatment approach has been shown to be effective in helping patients find successful antidepressant options or psychotherapy [28,29].

Safety-net primary care clinics are the preferred venue for underserved patients to access depression care because it is a common point of service delivery [30-32]. However, these settings encounter a complex mix of patient, provider, and health system factors that can impede the adoption of evidence-based collaborative depression care and result in persistent disparities in depression outcomes [33].

Safety-net systems organize and deliver a significant level of health care and other related services to uninsured, Medicaid, and other vulnerable populations [34]. Safety-net primary care providers often find it challenging to engage patients with major depression, particularly when it is accompanied by chronic illnesses, because the patient must participate in active and ongoing depression symptom assessment and management in addition to managing the other medical conditions [16,33,35-37]. Concurrently, minority patients are less likely to voluntarily report depressive symptoms, may view depression as a moral weakness or character flaw rather than an illness, and may be more likely to ascribe symptoms of depression to a physical illness [33,38]. Therefore, low-income, predominantly minority patients in safety-net care systems often miss out on depression diagnosis and follow-up assessments [33,39].

The Diabetes-Depression Care-management Adoption Trial (DCAT) is a translational study in a large safety-net system of primary care settings. The study compared a technology-facilitated care (TC) model with a usual care (UC) model and a supported care (SC) model to assess whether technology could facilitate the adoption of collaborative depression care for patients with type 2 diabetes. The DCAT TC model is an information and communication technology (ICT)–facilitated care management approach that harnesses automated telephone assessment (ATA) technology and integrates it with a disease management registry (DMR) system to automate key aspects of depression care. The UC model is standard care in a safety-net system in which primary care physicians (PCPs) develop individualized treatment plans for depression and diabetes care. The SC model is a team-based collaborative care management approach that involves care team staff members to provide depression and diabetes care.

The DCAT study is expected to fill two important gaps in current collaborative depression care implementation research. First, existing studies largely rely on labor-intensive, team-based SC approaches to implement collaborative depression care [24-27,40-43]. There is evidence that this SC model is effective and can be cost-effective compared with UC [24-27,44-47]. However, it is challenging for SC teams to integrate depression comorbidity care when patients are presented with other medical conditions because of the intensive labor and time needed to proactively screen for depression, follow up on treatment, and monitor and manage long-term care [33,41,48]. By relieving providers in resource-constrained safety-net clinics from many labor-intensive tasks such as collecting, summarizing, and reviewing individual or aggregate patient data to facilitate care, automation can facilitate the adoption of collaborative depression care. Therefore, DCAT tested ICT that automated critical information collection and processing for depression care tasks, including (1) Depression assessments and symptom monitoring, (2) Patient self-management behavior prompting, (3) Optimization of treatment follow-up, and (4) provider collaborative communication.

Second, existing research has not fully addressed ways to develop a patient-centered approach to implement collaborative depression care for low-income, predominantly minority safety-net patient populations. Prior studies have demonstrated that the team-based approach can effectively implement collaborative depression care in safety-net environments [24-27]; DCAT built on this evidence by applying the ICT to further address language, time, and stigma barriers [33,41,48] affecting safety-net patients. DCAT accomplished this by customizing calls with the patient’s preferred language and call time, making multiple attempts (if needed) to connect with the patient, and establishing a private and machine-only venue to report sensitive depression measures to reduce social desirability bias [49-51]. About 80% of the patients agreed or strongly agreed that the DCAT-tested ICT was easy to use (86.2%, 94/109), nonintrusive (87.1%, 95/109), and private and secure (75.9%, 82/108) [51].

DCAT-related improvements have been reported in several publications, including trial design [33], TC technology design and evaluation [52], patient acceptance of the technology [51], patient engagement [50], provider implementation reactions [53], depression risk profiling and prediction [54,55], and cost-effectiveness analysis (unpublished data, 2018, [56]). This paper reports DCAT-related depression symptom and diabetes care outcomes after 6 months. In addition, to provide the patients’ perspective on treatment benefits, this paper includes patient-reported outcomes, including physical and mental well-being, functional impairment, and satisfaction with care [57].

The main hypothesis of the paper is that, compared with the UC group, both the TC and SC groups will have statistically significant greater improvement in depression symptoms, diabetes care processes and outcomes, and patient-reported outcomes. TC uses an ATA system to ease the adoption of collaborative depression care rather than direct clinical intervention with patients; therefore, although the researchers have no hypothesis on how TC outcomes will compare with SC outcomes, this paper also explores whether the TC group will have better outcomes than the SC group.

DCAT is a translational study conducted in collaboration with Los Angeles County Department of Health Services (LACDHS), the second largest safety-net care system in the United States. Institutional review board approval was obtained from the University of Southern California, the Olive View–University of California Los Angeles Medical Center, and the Los Angeles Biomedical Research Institute.

Textboxes 1 and 2 show the eligibility and ineligibility criteria for patients.

Every enrollee received a set of educational and community resource materials in Spanish or English.

The enrollment period was from April 2011 to May 2012. Patients with type 2 diabetes were identified for recruitment from the DMR database and clinic records. Patients provided verbal consent to bilingual research assistants during study eligibility screening.

Measures were taken at baseline and at 6, 12, and 18 months by independent English-Spanish bilingual interviewers. Primary outcomes included three depression outcomes, five diabetes care measures, and six patient-reported outcomes measuring physical and mental well-being, functional impairment, and satisfaction with care (see Table 2).

The target sample size was based on power analysis for two primary outcomes: reduction of depressive symptoms (measured by PHQ-9 score) and depression remission. Power analyses were conducted using nQuery (Statistical Solutions Ltd, Boston MA) [65] to estimate effect sizes of the treatment with nonrandomized pre- and postintervention comparisons and longitudinal statistical approaches for repeated measures to compare trends in depression-related outcomes. The calculations assumed an alpha of .05, power of 0.80, attrition rates less than 20% at each 6-month follow-up assessment up to 18 months, and 25% to 30% depression prevalence among patients with diabetes [25]. For pre- and postintervention comparisons across all three program conditions, a sample size of approximately 500 in each study group would allow detection of a small effect size of 0.1.

All analyses were carried out according to the intention-to-treat rule consistent with standard practice in most clinical trials. The propensity score method has proved to be an effective approach to analyzing observational or quasi-experimental studies [66-70]. A propensity score is defined as the probability that a patient is likely to receive treatment or control given the patient’s baseline characteristics. Patients with the same propensity scores are like those in a randomized controlled trial.

The classical propensity score method is only applicable to two-way comparisons. Thus, we used a generalized propensity score (GPS) method designed for comparing two or more interventions versus one comparison group [71,72], wherein the GPS is defined as the conditional probability that a patient is likely to be in a specific group given this patient’s baseline characteristics. As recommended [72], a multinomial logistic regression was used to estimate GPSs. The model used study group as the dependent variable and the measured baseline characteristics shown in Table 3 (see “Results”) as the independent variables. We subsequently checked the distribution of the estimated GPSs because comparisons between groups are suspect if substantial separation occurs between study groups [72-74].

Comparative treatment effects were estimated by linear or logistic regression models that used outcomes at 6 months as the dependent variable and study group, care team, outcome variables at baseline, two of the three estimated GPSs, insulin use, HBA_1c, age, gender, and preferred language as the independent variables. Regression that includes estimated GPSs as covariates has been shown to be an effective tool to adjust sample biases in observational or quasi-experimental studies [71,72]. Three care team variables were used to adjust for differences among providers. Two of the three estimated GPSs adjusted for imbalance in baseline characteristics. Insulin use, HBA_1c, age, gender, and preferred language were included because their effects on outcomes were of clinical interest; and their inclusion is consistent with prior findings in behavioral and clinical factors associated with depression in patients with diabetes [75]. The coefficients of study group predicted comparative treatment effects while controlling for other covariates. All statistical analyses were conducted at 0.05 significance level (two-tailed) using SAS (SAS Institute Inc, Cary NC) software, version 9.3.

A total of 1704 patients were screened, of which 101 patients met the exclusion criteria, 128 patients refused to participate, 12 patients did not sign the Health Insurance Portability and Accountability Act agreement, and 57 patients were excluded after not completing the baseline assessment. Men had a significantly lower enrollment rate than women (84.0% [536/638] vs 89.02% [949/1066], respectively; P=.003), which was partly associated with poor alcohol use scores (4.9% [31/638] for men vs 0.56% [6/1066] for women).

Among the 1406 patients enrolled in DCAT (484 in UC, 480 in SC, and 442 in TC), 1309 patients (416 in UC, 461 in SC, and 432 in TC) had complete data in the measures used in estimating the GPSs after interviews at baseline. As shown in Table 3, there were no significant differences in baseline depressive symptoms measured by the PHQ-9 score, anxiety symptoms measured by the Brief Symptom Inventory score, functional disability measured by the Sheehan Disability Scale (SDS), and the overall mental status measured by the 12-item Short Form Survey mental score. SC and TC patients had higher HBA_1c compared with UC patients because the SC and TC patients were enrolled from the DMP program designed for patients with severe diabetes. Other significant differences were diabetes self-care score and psychological stress measures (economic stress, number of stressors, sum of stress level, diabetes emotional burden, and diabetes regimen stress). The unbalanced samples were expected because of the quasi-experimental design. A Consolidated Standards of Reporting Trials diagram outlining participant flow is shown in Figure 1. See Multimedia Appendix 1 for comparison of baseline characteristics of samples included in versus excluded from the regression analysis. No significant differences were identified.

With the final sample size of 1087 (341 in UC, 380 in SC, and 366 in TC) to evaluate intervention effects, the study has the statistical power of 0.80 to detect an effect size of Cohen d=0.12, a small effect size. Regression analysis with GPS adjustment results regarding 6-month outcomes in DCAT are shown in Tables 4 and 5. The GPSs were estimated by the previously described multinomial logistic regression model. The distributions of the estimated GPSs across study groups were similar; thus, treatment effects can be predicted based on the estimated GPSs instead of actual group assignment.

Compared with UC, both SC and TC were significantly associated with decreased PHQ-9 scores (least squares estimate, LSE: UC=6.35, SC=5.05, TC=5.16; P value: SC vs UC=.02, TC vs UC=.02) and reduced prevalence of depression as measured by PHQ-9 ≥10 (SC vs UC: adjusted odds ratio, AOR=0.45, 95% CI 0.23-0.88, P=.02; TC vs UC: AOR=0.33, 95% CI 0.17-0.65, P=.007). Only TC was significantly associated with improved depression remission relative to UC (AOR=2.98, 95% CI 1.08-8.25, P=.04), although SC came close. There were no significant differences in depression outcomes between the SC and TC groups.

Regarding diabetes care measures, no significant differences existed between SC and UC. However, TC was significantly associated with reduced total cholesterol level (LSE: UC=176.40, TC=160.46; P=.01) and increased odds that the patient would have an HBA_1c test (TC vs UC: AOR=3.40, 95% CI 1.58-7.31, P<.001). The latter was positively correlated with depression remission (AOR=2.67, 95% CI 1.15-4.17, P=.004). There were no significant differences in diabetes care measures between the SC and TC groups.

Both SC and TC were significantly associated with improved SDS scores relative to UC (LSE: UC=3.21, SC=2.61, TC=2.59; P value: SC vs UC=.04, TC vs UC=.03). SC was significantly associated with improved satisfaction with care for emotional problems compared with UC (LSE: UC=3.25, SC=3.64; P=.01), but only TC was significantly associated with improved satisfaction with diabetes care (LSE: UC=4.01, TC=4.20; P=.05) and satisfaction with care for emotional problems among patients with depression, as measured by PHQ-9 ≥10 at baseline (LSE: UC=3.20, TC=3.70; P=.05). There were no significant differences in patient-reported outcomes between the SC and TC groups.

Analysis of 6-month DCAT outcomes revealed that both the TC and SC groups were significantly associated with better outcomes compared with UC in terms of depressive symptoms reduction. Using the PHQ-9 score, which ranges from 0 to 27, the 1.3 (SC group) and 1.2 (TC group) points of improvements compared with the UC group are clinically meaningful given that the baseline PHQ-9 score is only about 6.5 points. The magnitude of improvements is consistent with a recent collaborative depression care study that included both depressed and nondepressed patients [80]. This finding supports the hypothesis that the two intervention groups would be associated with better depression care outcomes.

It was not surprising to find the ATA technology did not improve depression outcomes of the TC group over the SC group in this case because it was designed to facilitate the adoption of collaborative depression care rather than direct clinical intervention with patients. Clinically, DMP care teams in both groups were trained in and practiced the LACDHS depression protocol and treatment guideline. The SC DMP providers monitored patients using traditional mechanisms (specifically by calling patients), and they coordinated care among themselves; therefore, the study-related depression care resulted in additional new workload for the SC providers. For the TC care team, the technology, albeit at a low-intensity of contact, helped alleviate the workload for depression symptom and treatment adherence monitoring.

The ATA technology also prompted providers to follow up and alerted emergency responders to immediately contact patients with suicidal ideation. That could be the reason why only the TC group was significantly associated with depression remission and increased patient satisfaction among depressed patients. The results are encouraging that a well-designed technology can be an effective aide to the adoption of collaborative depression care. Full justification of the TC model via a complete cost-effectiveness analysis is presented elsewhere [56]. The cost-effectiveness analysis revealed that the intervention models improved quality-adjusted life years, depression-free days, and medical costs. The TC model was cost-effective compared with SC and cost-saving compared with UC. The 6-month and cost-effectiveness results suggest the TC model is promising in facilitating better and cost-effective care for depressed patients.

Moreover, only the TC model was significantly associated with improved diabetes care processes, indicated by reduced total cholesterol level and increased odds that the patient would have an HBA_1c test. One possible explanation for these improvements in diabetes care is that as depressive symptoms are increasingly monitored and timelier addressed, patients may become more willing to take active care of their diabetes; this explanation is supported by the significant correlation between the odds of having an HBA_1c test and the improvement in depression remission. Another possible explanation is that providers may address patients’ diabetes care needs in addition to depression care needs when they respond to the task reminders generated by the technology.

Compared with the UC group, both the SC and TC models were significantly associated with better improvement in 6-month patient-reported functioning in family, work, and social life, as measured by the 3-item SDS. The 0.6 points of improvements in SC and TC groups compared with the UC group are meaningful as most patients at baseline had only minimal functional impairment (average baseline SDS score was 2.1 points; SDS score >6 indicates functional disability [63,64]), which implies that the room for improvement is small. This finding suggests that the two enhanced care delivery models not only improved depressive symptoms but also translated such symptom improvement into better perceived life functioning.

While the TC model delivered positive results, and most patients in the TC groups reported high acceptance of the ICT tested in DCAT [51], there is significant room for improvement in using the ATA technology. Specifically, only half of the scheduled calls were answered successfully because of phone connectivity issues or lack of time for the patient to answer the calls [50,52]. One challenge may have been that during the study, most patients in LACDHS did not use cellphones. Now that cellphones are more readily available, attention should be turned to other ICT (such as SMS text messages and smartphone apps) to improve patient contact and to capture patient-reported outcomes. Such technology has greater portability and versatility, may extend the ATA capabilities in reaching and engaging patients, may potentially increase the model effectiveness, and may reduce costs.

Researchers can expect unpredictable consequences after making changes (such as the DCAT implementation) in complex systems such as LACDHS. As discussed in another DCAT study [53], two particularly important implications that emerged were the strengthened role of social workers (in both SC and TC) and the importance of suicide-alert responders (in TC). Every DMP site had a colocated clinical social worker, an evidence-based method of improving quality of depression care [81,82]. The clinical social workers were an underutilized resource before the study; during the DCAT trial, they proved instrumental in adopting depression care. Furthermore, the suicide-alert responders appeared to play a much larger role than anticipated. Providers facing typical barriers in mental health care (including lack of familiarity with guidelines, lack of self-efficacy, and lack of outcome expectancy [83]) were reassured by the availability of an organizational resource for the patient to fall back on in the “worst case scenario,” namely severe suicidal ideation. Taken together with the strengthened role of the social worker, the interventions seemed to have leveraged the available mental health resources into a more cohesive, integrated model of mental health care in a primary setting. In other words, the SC and TC models used existing diabetes disease management teams and leveraged available mental health resources to implement depression symptom monitoring and treatment protocols, provider collaborative communication, and patient relapse prevention.

In summary, the 6-month DCAT findings suggest that both the TC and SC delivery models are significantly associated with improved depression outcomes and life functioning, and that the TC model offers additional promise in terms of improved depression remission, diabetes care processes, and patient satisfaction. Given the rapid rise of diabetes during the past several decades—especially among low-income, minority populations—and the immense opportunity to improve diabetes-related measures and outcomes, a growing number of health plans and health care organizations are trying to manage their diabetes population through disease management programs. The SC and TC models demonstrated that an important and valuable way to support providers is to add evidence-based collaborative depression care and facilitate adoption of ICT in diabetes DMPs designed to reduce disparities in commonly comorbid diabetes and depression care. When enhanced by ICT, DMPs may be able to greatly improve overall care, cost, and effectiveness of health care delivery for underserved patients. DCAT SC and TC models improved diabetes and depression outcomes in the second largest US safety-net health system; other resource-constrained programs may replicate these models to improve comorbid diabetes and depression outcomes.

The main limitation of this study is its employment of a quasi-experimental design, which introduces bias because of the differences in both patient characteristics and care teams at each facility. To mitigate the bias, in the regression analysis we adjusted for patient differences through propensity scores and the assignment of the six care teams at eight facilities (each facility was staffed by only one team; two of the six teams served two facilities). Although care facilities were matched by geographic location and patient sociodemographics among the three study groups, the quality of care can vary from facility to facility; therefore, the regression analysis included a check in which care team assignments were replaced for each facility. This analysis did not change the direction and significance of intervention effects; however, the adjustment may not be sufficient. Differences in the facilities, the DMP care teams, and the unmeasured patient sociodemographics, diabetes and comorbid conditions, and psychological stress measures may differentiate diabetes and depression care needs and outcomes [84]. Providers should consider these differences when applying the technology.

Another limitation may be the predominantly Latino sample, which raises concern about the generalizability of findings. Applying the DCAT TC model to other groups should be done cautiously and with further evaluation.

The third limitation is the focus in this paper on clinical outcomes. However, the DCAT TC model was designed to accelerate the adoption of evidence-based collaborative care to improve the overall care process. Full analyses of 18-month clinical outcomes and the cost-effectiveness of the TC model will be reported elsewhere.

The fourth limitation is lack of data to understand the practical mechanism by which the SC DMP and the technology-enhanced TC DMP led to improvements in depression. Possible improvement mechanisms for future research include better treatment initiation or adjustment, receipt of PST, greater patient adherence, or referrals and visits to other mental health providers.

DCAT adds to the growing number of telehealth studies that are employing technology to improve depression care in primary care settings for patients with chronic diseases [85-92]. A key strength of the DCAT TC model over earlier studies is that it used automated calls, which reduces provider depression monitoring workload and allows more time for clinical encounters such as timely adjustment of treatment. The TC model is especially effective in a resource-constrained environment such as safety-net care systems, improving care for predominantly minority and low-income patients.

Applying automated remote monitoring ICT, electronic clinical decision support, and even artificial intelligence to facilitate chronic disease management is an emerging research topic. Prior studies revealed that ATA is valid in conducting depression screening and suggested the technology can be incorporated into the care management model [93,94]. However, evidence is limited regarding the comparative effectiveness of the technology. Kroenke et al [95] tested ATA with care management for pain and depression in patients with cancer; Ratanawongsa et al [96] and Handley et al [97] tested ATA-facilitated diabetes management for low-income Medicaid and safety-net patients. The DCAT study uniquely addressed depression care for patients with type 2 diabetes in a safety-net setting where comorbidity of the two diseases is common and where many barriers such as culture diversity, financial stress, and limited provider resources impede the adoption of evidence-based depression interventions. Results from DCAT are consistent with prior studies [91-97] and support ATA as a promising technology to facilitate care management, even for sensitive conditions such as depression, for diverse populations and in primary care settings.

Both SC and TC models are associated with improved 6-month depression outcomes and reduced functional disability among adult patients with type 2 diabetes. However, the TC model is more likely to achieve greater improvements in depression remission, as well as measures of patient satisfaction and diabetes care quality. This paper provides encouraging evidence that a well-designed automated ICT system is an effective facilitator that can support delivery of evidence-based collaborative depression care to patients with type 2 diabetes in a resource-constrained urban safety-net primary care setting. This is a promising solution to reduce health disparities, improve patient experience of care, and improve the health of low-income, minority populations.