Dulce Digital-Me: results of a randomized comparative trial of static versus adaptive digital interventions for Latine adults with diabetes

Abstract

Objective

To compare the effectiveness of a static, text-based diabetes education and support intervention (Dulce Digital, DD) versus a dynamic approach with personalized feedback and goal setting (Dulce Digital-Me, DD-Me) in improving diabetes outcomes.

Design and methods

Comparative effectiveness trial in 310 Latine adults with poorly managed type 2 diabetes in a Federally Qualified Health Center in Southern California, randomized to DD, DD-Me-Auto (algorithm-driven text-based personalized feedback), or DD-Me-Tel (coach delivered personalized feedback). Changes in HbA1c (primary outcome), low-density lipoprotein-cholesterol, systolic blood pressure, and patient-reported outcomes were examined across 6 and 12 months, with the primary comparison being DD versus DD-Me (combined automated and telephonic).

Results

Participants were 52.1 (±10.2) years old, 69.7% female, with HbA1c 9.3% (±1.6) at baseline. Across groups, there was a statistically significant improvement in HbA1c at 6 months (mean∆ per month = −0.17, 95% CI −0.20, −0.14; P < .001) and 12 months (mean∆ per month = −0.07, 95% CI −0.09, −0.05; P < .001). However, there were no time-by-group interaction effects indicating group differences in clinical outcomes across 6 or 12 months. The DD-Me groups showed greater improvements across time than the DD group for diabetes self-management behaviors.

Conclusions

Static and adaptive digital interventions for Latine adults with type 2 diabetes had similar and clinically significant effects on HbA1c across 12 months. Simple digital approaches can be integrated within primary care-based chronic care models to reduce diabetes disparities.

ClinicalTrials.gov registration

NCT03130699, Initial Release 04/24/2017, https://clinicaltrials.gov/ct2/show/NCT03130699?term=NCT03130699&draw=2&rank=1.

Latine individuals with type 2 diabetes responded positively to static and adaptive digital text messaging approaches in their language of choice resulting in better glycemic outcomes.

Graphical Abstract

Graphical Abstract.

Introduction

From 2001 to 2020, diabetes prevalence significantly increased among US adults, affecting 37.3 million people, or 11.3% of the US population.¹ Diabetes poses a substantial burden on society, and individuals of low-income and racial and ethnic minority backgrounds experience significant health disparities.¹ Hispanic/Latino(a/x/e) (hereafter Latine [We use the term Latine to acknowledge the Latin American (Mexican) heritage of most of our participants while recognizing gender inclusivity. Participants were not asked about their preferred terminology in this study.]²) adults in aggregate have higher rates of diabetes than non-Hispanic Whites, although there is substantial heterogeneity by heritage group (eg, higher among Mexican and Puerto Rican descent) and place of birth (eg, higher among those living in the US the longest).³ Adverse social determinants of health, including poor healthcare access and low income, as well as communication and cultural factors, impede optimal management, placing Latine persons at disproportionate risk for costly diabetes complications.⁴

Diabetes self-management education and support (DSME/S) provides a solid foundation upon which to deliver better healthcare, improve adherence, clinical, quality of life, and cost outcomes, and reduce disparities.⁴ However, many at-risk individuals are unable to access DSME/S due to practical barriers. Healthcare systems need alternative methods to extend the reach of DSME/S efficiently and effectively.

Rapidly evolving digital health technology can positively impact outcomes in diabetes.^5,6 Investigations of digital interventions for diabetes care are limited in Latine persons,⁵ although preliminary evidence suggests that simple, text message and telehealth interventions promote clinically significant improvements in glycemic control in this population.^6,7 The widespread adoption of cellular telephones, even among older adults and individuals from low-income and racial and ethnic minority groups,⁸ highlights the potential for digital health technology to circumvent practical barriers to traditional DSME/S among Latine persons. Additional trials are needed to confirm the effectiveness and acceptability of digital health interventions in this medically underserved population.

Dulce Digital (DD), an educational text messaging program available in English and Spanish, developed from the evidence-based Project Dulce (PD) curriculum⁹ includes static (one size fits all) educational and supportive messaging. In a randomized controlled trial, DD improved glycemic control relative to usual care in Latine adults with type 2 diabetes (HbA1c meanΔ = −1.0% vs −0.2%, P < .05).⁹  Dulce Digital-Me (DD-Me) was developed to enhance the DD program by incorporating “adaptive” or dynamic components (eg, personalized feedback), based on input received from patients.^10,11 DD-Me is theoretically informed and uses evidence-based behavior change approaches to improve self-management.¹² This randomized trial examined the comparative effectiveness of DD versus DD-Me in improving diabetes management and patient-reported outcomes. The study evaluated the additive value of personalized assessment, goal setting, and feedback, while also examining 2 different methods of delivering these adaptive components: algorithm-driven text messaging (DD-Me-Auto) or telephonic delivery by a specially trained medical assistant health coach (DD-Me-Tel). We anticipated that the personalized approach would have a greater impact on improving diabetes outcomes relative to the already effective DD texting intervention. We did not have a priori hypotheses concerning the relative advantage of automated-text or coach-delivered feedback in the 2 DD-Me conditions but compared these groups to support future implementation.

Research design and methods

Procedures were approved by the Institutional Review Boards of Scripps Health (reviewing) and San Diego State University (relying). All participants provided written informed consent. Recruitment started on June 22, 2017, and the trial was completed on August 14, 2021. Trial methods have been detailed previously¹² and are summarized below.

Study design and randomization

The study applied an individually randomized with equal allocation (1:1:1-DD:DD-Me-Auto:DD-Me-Tel), parallel groups, open-label, comparative effectiveness design. The primary focus was on the difference between DD compared with DD-Me (Auto + Tel). Staff who conducted follow-up assessments were blinded to allocation.

Setting, population, and sample

Participants were Federally Qualified Health Center (FQHC; population 80% Latine) patients in San Diego County, CA. Inclusion criteria were Latine (self-identified as Hispanic/Latino/a); ≥18 years of age; diagnosed type 2 diabetes; and glycosylated hemoglobin (HbA1c) ≥8.0%, systolic blood pressure (SBP) ≥140, and/or low-density lipoprotein-cholesterol (LDL-C) ≥100 mg/dL in past 90 days. Exclusion criteria were Severe/life-threatening illness; pregnant or lactating; not Spanish or English literate; severe auditory, or visual problems.

A priori power analysis identified a target sample size of N = 414, allowing for up to 30% attrition, to detect a small–medium effect size difference between any 2 groups for all outcomes.¹² Specific to HbA1c, this effect size represents a clinically meaningful change of 0.5% with a 1.3% SD (with SD determined from prior studies in a similar population; d = 0.33).¹¹ However, study enrollment concluded early with N = 310 participants, due to the corona virus disease (COVID-19) pandemic. This modification was approved by the funder and IRBs. Supplementary Figure S1 depicts the CONSORT diagram.

Study orientation

Electronic medical records (EMRs) were queried to identify patients meeting the criteria, telephone screening was conducted, and those eligible attended an in-person study orientation and baseline visit. Trained bilingual interviewers obtained consent, conducted baseline assessments, and unveiled randomization assignments. Participants were provided a cellular-enabled glucose monitor, test strips, and a cellular-enabled pillbox. Participants without a cellular telephone were provided a study phone. Participants viewed a brief diabetes management education video describing the study and devices. They were given one-on-one device training and written instructions.

Interventions

Details regarding the intervention’s theoretical framing and content have been published¹² but are described briefly here. Interventions were conducted over 6 months, during which participants continued to receive routine diabetes care. Primary care physicians were provided a printed summary of blood glucose values and behavioral data prior to patient clinic visits.

All groups received core text messaging content suitable for their cultural and health literacy levels in their preferred language (Spanish or English) and consistent with the National Standards for Diabetes Education and Support¹³ and best practices for behavior change^12,14 specifically: (1) taking medication; (2) clinical indicators and reducing risk; (3) healthy eating; (4) being active; and (5) and healthy coping. The content was based on our Project Dulce^12,15 and DD evidence-based DSME/S programs,¹⁶ with updates consistent with current standards. Core content messages were sent during participant-selected time windows—initially, 2-3 per day, tapered over 6 months.

Participants monitored their blood glucose and managed oral diabetes medication using cellular-enabled devices and responded to brief ecological momentary assessment (EMA) questions delivered in their preferred language (Spanish or English) via text message assessing health behaviors and well-being in real time. Data from the pillbox (openings, as a proxy for medication taking), glucose monitor (blood glucose readings), and EMA responses were captured using a secure, device-agnostic platform¹⁷ that was also used to deliver core DSME/S text content, and personalized feedback and goal setting for DD-Me participants.

DD participants received the same content, frequency, and order of messages, with no tailoring other than text delivery windows. DD-Me participants received the same DD components plus they: (1) selected the core content messaging topic they preferred to receive first (eg, eating vs activity) and (2) received adaptive feedback and goal setting informed by transmitted data. Those in the DD-Me-Auto arm received algorithm-driven text messages tailored to EMA responses in real time and weekly texted summaries of blood glucose and medication adherence. A priori, EMA responses, glucose values, and pillbox openings were categorized from “optimal” to “needs improvement,” and if/then logic and response sets were designed to motivate maintenance or improvement of healthy self-management behaviors. Those in the DD-Me-Tel arm received adaptive feedback and goal setting from the same Spanish–English bilingual, bicultural health coach during weekly phone calls. Coaching calls were facilitated by an automatically generated report with summaries of participants’ progress based on glucose values, pillbox openings, and EMA data, using the same targets used in the DD-Me-Auto group to facilitate consistent feedback.

Assessment procedures and measures

Assays of HbA1c using a National Glycohemoglobin Standardization Program-certified laboratory¹⁸ and LDL-C; assessments of SBP and body mass index; and surveys in preferred language were conducted at baseline. Participants returned 6 and 12 months later (±8 weeks) for group-based follow-up assessments. From March 2020 until the study end, participants obtained a blood draw at the nearest FQHC lab and completed follow-up surveys by phone, due to the COVID-19 pandemic. Electronic medical record clinical data obtained during the follow-up windows were then abstracted.

Self-report measures captured patient-reported outcomes, emphasizing potential pathways of intervention effects. The “provider support” subscale of the Chronic Illness Resource Survey (CIRS) evaluated patient–provider communication over the past 3 months (Cronbach’s α = 0.83 in the current sample at baseline).¹⁹ A single Summary of Diabetes Self-Care Activities²⁰ item assessed days per past week when blood glucose was tested as recommended. Diabetes medication adherence was assessed with the Adherence to Refills and Medications Scale (Cronbach’s α = 0.78 in the current sample at baseline).²¹ The Rapid Assessment of Physical Activity assessed the extent of aerobic activity from sedentary to regular active and engagement in any strength and flexibility activity.²² Healthful eating behaviors were assessed with the Food Behavior Checklist (FBC).²³ The Diabetes Distress Scale evaluated diabetes distress (Cronbach’s α = 0.93 in the current sample at baseline).²⁴ Self-reported demographic variables were obtained at baseline. Health literacy was assessed using the Single Item Literacy Screener.²⁵

Statistical methods

Analyses were performed using IBM SPSS Statistics (IBM Corporation, Armonk, NY) and MPlus V8.9.²⁶ Inferential statistical testing determined baseline group differences and identified possible covariates. Missing data analyses were conducted separately for the clinical and patient-reported outcomes. Variables associated with missing data were added as covariates and all models were estimated using full-information maximum likelihood to further account for missing data.²⁷ This approach uses all available data to estimate parameters, instead of replacing or imputing missing values, and has been shown to provide unbiased parameter estimates and standard errors under various missing data conditions.²⁸

For the CONSORT diagram, the percentages receiving the allocated intervention were calculated per participant and summarized by the allocated group. Totals of each intervention metric (EMA/Core content [all groups], automated text, blood glucose, and pillbox feedback [DD-Me-Auto only], and weekly call feedback [DD-Me-Tel only]) received per participant were calculated, and for each metric, the proportion of participants within each group who received >50% of the overall intended intervention was reported. For the DD-Me groups with multiple intervention components, those who received >50% of all components were considered to have received the intervention as intended.

Changes over time (defined as months since baseline) were examined using multilevel models as implemented by MPlus Model parameters, and standard errors were estimated using complete and partial cases, and the appropriate link function for each outcome. Both baseline–6-month (primary) and baseline–6–12-month (3 timepoints; maintenance) models were tested. Models included dummy-coded variables representing group assignments. While the primary focus was on comparing DD to DD-Me-combined, multiple dummy-coded systems were created to make all pairwise comparisons between individual DD groups. These dummy-coded variables were then used in the Time × Group interaction terms, assessing differences in the rate of change between groups. Within-group changes for the total sample and each intervention group were also tested to determine the statistical and clinical significance of changes in outcomes. Unstandardized regression coefficients (B), 95% confidence intervals, and P values are reported. Statistical significance was determined by 2-sided P < .05. Analyses followed intention-to-treat principles. Sensitivity (per-protocol) analyses were also conducted for HbA1c only, limiting the DD-Me groups to those who received at least 50% of the allocated intervention (see Supplementary Figure S1).

Data and resource availability statement

The datasets generated during and/or analyzed in the current study are available from the corresponding author upon reasonable request.

Results

As shown in Supplementary Figure S1, the percent of participants who received the allocated intervention as intended varied by group, ranging from 100% in the DD group to 46% in the DD-Me-Auto group. Importantly, the percentages were lower for components that were dependent on participant engagement. Automated feedback messages and weekly glucose range summaries relied on participants responding to EMA items and checking glucose (DD-Me-Auto), and the completion of health coaching calls relied on the participant/coach successfully connecting by telephone (DD-Me-Tel).

At baseline, the intervention groups differed only in employment status (Table 1), and this variable was controlled in all multilevel models. Those with missing data on clinical outcomes were younger, and more likely to be male and English-speaking relative to those with complete data, while those with missing data on patient-reported outcomes were more likely to have lower income; therefore, these variables were also included as covariates. Supplementary Table S1 displays descriptive statistics for all outcomes at each timepoint, for all groups.

Table 1.

| Baseline sociodemographic and clinical characteristics of the total sample and intervention groups.^a

Variable	Total sample N = 310	DD n = 107	DD-Me-Auto n = 106	DD-Me-Tel n = 97
Age, y, M (SD)	52.11 (10.18)	51.22 (10.7)	53.21 (10.26)	51.90 (9.49)
Sex n (%)
Male	94 (30.3)	31 (29.0)	33 (31.1)	30 (30.9)
Female	216 (69.7)	76 (71.0)	73 (68.9)	67 (69.1)
Ethnicity, n (%)
Hispanic/Latino	310 (100%)	107 (100%)	106 (100%)	97 (100%)
Not Hispanic/Latino	0 (0%)	0 (0%)	0 (0%)	0 (0%)
Race, n (%)
White	106 (35.0)	42 (40.0)	37 (35.6)	27 (28.7)
Black or African American	3 (1.0)	2 (1.9)	1 (1.0)	0 (0.0)
American Indian or Alaskan Native	4 (1.3)	3 (2.9)	1 (1.0)	0 (0.0)
Asian	1 (0.3)	1 (1)	0 (0.0)	0 (0.0)
Native Hawaiian or Other Pacific Islander	3 (1.0)	1 (1)	1 (1.0)	1 (1.1)
Other race	181 (59.7)	54 (51.4)	62 (59.6)	65 (69.1)
More than 1 race	5 (1.7)	2 (1.9)	2 (1.9)	1 (1.1)
Education, n (%)
< High school education/GED	230 (74.9)	81 (76.4)	73 (69.5)	76 (79.2)
≥ High school education/GED	77 (25.1)	25 (23.6)	32 (30.5)	20 (20.8)
Yearly household income, n (%)
≤$20 000	189 (66.5)	64 (63.4)	62 (66.0)	63 (70.8)
>$20 000-$30 000	49 (17.3)	14 (13.9)	19 (20.2)	16 (18.0)
>$30 000	46 (16.2)	23 (22.8)	13 (13.8)	10 (11.2)
Employment, n (%)
Not employed	168 (54.4)	55 (51.9)	68 (64.2)	45 (46.4)
Employed part time	72 (23.3)	21 (19.8)	19 (17.9)	32 (33.0)
Employed full time	69 (22.3)	30 (28.3)	19 (17.9)	20 (20.6)
Marital status, n (%)
Not married/living with partner	107 (34.5)	33 (30.8)	41 (38.7)	33 (34.0)
Married/living with partner	203 (65.5)	74 (69.2)	65 (61.3)	64 (66.0)
Place of birth, n (%)
US 50 states/DC	20 (6.5)	5 (4.7)	9 (8.5)	6 (6.2)
Mexico	281 (90.6)	97 (90.7)	95 (89.6)	89 (91.8)
Other	9 (2.9)	5 (4.7)	2 (1.9)	2 (2.1)
Language, n (%)
English	21 (6.8)	6 (5.6)	8 (7.5)	7 (7.2)
Spanish	289 (93.2)	101 (94.4)	98 (92.5)	90 (92.8)
Insurance, n (%)
Uninsured	158 (51.0)	54 (50.5%)	49 (46.2%)	55 (56.7)
Insured (any)	152 (49%)	53 (49.5)	57 (53.8)	42 (43.3)
Health literacy, n (%)b
Adequate	167 (53.9)	56 (52.3)	56 (52.8)	55 (56.7)
Limited	143 (46.1)	51 (47.7)	50 (47.2)	42 (43.3)
Duration of diabetes diagnosis, y, M (SD)	10.37 (8.22)	10.22 (8.22)	10.31 (8.25)	10.59 (8.26)
BMI, kg/m2, M (SD)	32.22 (6.54)	31.84 (5.82)	33.51 (7.35)	31.37 (6.34)
HbA1c, %, M (SD)	9.31 (1.59)	9.25 (1.64)	9.31 (1.58)	9.37 (1.56)
SBP, mm Hg, M (SD)	122.35 (18.33)	121.36 (17.11)	122.69 (17.96)	123.08 (20.10)
LDL-C, mg/dL, M (SD)	94.72 (38.00)	93.07 (38.75)	91.17 (33.88)	100.32 (41.03)

Abbreviations: BMI, body mass index; DD, Dulce Digital; DD-Me-Auto, Dulce Digital-Me_Automated; DD-Me-Tel, Dulce Digital-Me_Telephonic; GED, General Education Development Test; HbA1c, glycosylated hemoglobin; LDL-C, low-density lipoprotein-cholesterol; SBP, systolic blood pressure.

aSample sizes differ due to missing data for individual variables.

^bHealth literacy was assessed using the Single Item Literacy Scale (SILS; adequate or limited).

Clinical outcomes

Table 2 displays the results of multilevel models. There were no differences for the primary outcome in the rate of HbA1c change between groups over 6 or 12 months (ie, no significant group × time interaction effects). However, as shown in Table 3, an analysis of the total sample showed a significant improvement in HbA1c across the groups for the baseline–6-month timeframe (mean∆ per month −0.17%, 95% CI −0.20, −0.14; P < .001) and in the baseline–6–12-month timeframe (mean∆ per month −0.07%, 95% CI −0.09, −0.05; P < .001). Unexpectedly, there was a clinically small but statistically significant increase in SBP in the baseline–6-month timeframe (mean∆ per month + 0.40 mm Hg, 95% CI −0.01, 0.78; P < .04) but not for baseline–6–12 months. There was no intervention effect for LDL-C. Within-group analyses of changes over time also showed significant improvements in HbA1c in each intervention group when examined separately (Table 3; Figure 1).

Table 2.

| . Time-by-group interaction effects for all outcomes, baseline to 6 months (top of the table), and baseline to 12 months (bottom of the table).^a

Outcome variable	DD-Me-combined (Auto + Tel) vs DD B (95% CI) P	DD-Me-Auto vs DD B (95% CI P)	DD-Me-Tel vs DD B (95% CI) P	DD-Me-Auto vs DD-Me-Tel B (95% CI) P
Clinical outcomesb	Baseline–6 months
HbA1c (%)	0.00 (−0.06, 0.06) P = .95	−0.03 (−0.10, 0.04) P = .42	0.03 (−0.03, 0.10) P = .32	0.06 (−0.01, 0.13) P = .08
LDL-C (mg/dL)	0.84 (−0.63, 2.31) P = .26	0.54 (−1.08, 2.16) P = .50	1.13 (−0.67, 2.93) P = .22	0.59 (−1.18, 2.36) P = .52
SBP (mm Hg)	−0.46 (−1.17, 0.26) P = .21	−0.39 (−1.21, 0.43) P = .35	−0.52 (−1.37, 0.32) P = .23	−0.13 (−0.97, 0.71) P = .76
Patient-reported outcomesc
Patient–provider communicationd	0.00 (−0.03,0.03) P = .90	0.001 (−0.03,0.04) P = .78	0.00 (−0.04, 0.04) P = .92	0.01 (−0.03, 0.04) P = .71
Blood glucose monitoring (d/wk)e	−0.02 (−0.10, 0.06) P = .61	−0.03 (−0.11, 0.05) P = .47	−0.01 (−0.11, 0.09) P = .83	−0.02 (−0.12, 0.08) P = .67
Aerobic exercisef	0.00 (−0.03, 0.03) P = .96	0.00 (−0.04, 0.03) P = .72	0.00 (−0.04, 0.04) P = .91	0.00 (−0.04, 0.04) P = .94
Strength exerciseg,h	0.01 (−0.10, 0.14) P = .87	0.01 (−0.12, 0.13) P = .86	0.01 (−0.14, 0.15) P = .93	0.01 (−0.14, 0.15) P = .94
Flexibility exerciseg,i	0.07 (0.07, 0.25) P = .25	0.09 (−0.04, 0.22) P = .17	0.03 (−0.10, 0.16) P = .63	0.06 (−0.07, 0.19) P = .38
Healthy diet behaviorsj	0.03 (−0.14, 0.19) P = .76	−0.03 (−0.21, 0.16) P = .78	0.08 (−0.11, 0.27) P = .39	−0.11 (−0.30, 0.08) P = .26
Diabetes distressk	−0.01 (−0.02, 0.03) P = .63	0.02 (−0.01, 0.05) P = .20	−0.01 (−0.04, 0.02) P = .57	0.03 (−0.01, 0.06) P = .06
Medication adherencel	−0.20 (−0.33, −0.007) P = .002	−0.019 (−0.34, −0.04) P = .02	−0.23 (−0.37, −0.08) P = .002	0.04 (−0.10, 0.18) P = .56
Clinical outcomesb	Baseline–6–12 months
HbA1c (%)	0.02 (−0.02, 0.06) P = .37	0.00 (−0.04, 0.05) P = .96	0.03 (−0.01, 0.08) P = .12	0.03 (−0.01, 0.07) P = .12
LDL-C (mg/dL)	0.05 (−0.69, 0.79) P = .90	0.04 (−0.78, 0.64) P = .92	0.06 (−0.87, 0.98) P = .91	0.01 (−0.92, 0.94) P = .98
SBP (mm Hg)	−0.18 (−0.53, 0.16) P = .30	−0.26 (−0.63, 0.12) P = .17	−0.11 (−0.52, 0.30) P = .60	0.15 (−0.22, 0.52) P = .42
Patient-reported outcomesc
Patient–provider communicationd	0.01 (−0.02, 0.03) P = .72	0.01 (−0.02, 0.03) P = .62	0.00 (−0.03, 0.03) P = .92	0.01 (−0.02, 0.04) P = .69
Blood glucose monitoring (d/wk)e	−0.02 (−0.09, 0.04) P = .49	−0.03 (−0.11, 0.04) P = .37	−0.01 (−0.09, 0.07) P = .82	−0.03 (−0.11, 0.06) P = .55
Aerobic exercisef	−0.01 (−0.03, 0.02) P = .64	−0.01 (−0.04, 0.02) P = .58	0.00 (−0.04, 0.03) P = .80	0.00 (−0.04, 0.03) P = .80
Strength exerciseg,h	0.03 (−0.07, 0.12) P = .58	0.04 (−0.07, 0.15) P = .52	0.02 (−0.10, 0.13) P = .79	0.02 (−0.10, 0.14) P = .74
Flexibility exerciseg,i	0.06 (−0.01, 0.13) P = .11	0.11 (0.02, 0.20) P = .02	0.01 (−0.07, 0.09) P = .82	0.10 (0.003, 0.19) P = .04
Healthy diet behaviorsj	0.13 (0.001, 0.25) P = .05	0.07 (−0.08, 0.21) P = .36	0.19 (0.04, 0.35) P = .01	−0.13 (−0.29, 0.04) P = .13
Diabetes distressk	0.01 (−0.03, 0.02) P = .63	0.00 (−0.02, 0.03) P = .72	−0.02 (−0.04, 0.01) P = .16	0.02 (−0.01, 0.04) P = .06
Medication adherencel	−0.14 (−0.24, −0.04) P = .008	−0.11 (−0.23, 0.001) P = .05	−0.16 (−0.28, −0.05) P = .006	0.05 (−0.06, 0.16) P = .40

Abbreviations: Auto, automated; DD, Dulce Digital; DD-Me, Dulce Digital-Me; HbA1c, glycosylated hemoglobin; LDL-C, low-density lipoprotein-cholesterol; SBP, systolic blood pressure; Tel, telephonic.

^aUnstandardized regression coefficients are presented and indicate the difference in rate of change between groups shown in column headings.

^bAnalyses control for age, gender, employment, and language.

^cAnalyses control for income.

^dPatient–provider communication was assessed using the healthcare team subscale of the Chronic Illness Resource Survey (range, 1-5).

^eBlood glucose monitoring (d/wk) was assessed using the blood glucose question from the Summary of Diabetes Self-Care Activities measure (range, 0-7).

^fAerobic exercise was assessed using the aerobic subscale of the Rapid Assessment of Physical Activity (5-point scale, 1-5).

^gBinary logistic regression analyses were conducted for this outcome. Unstandardized regression coefficients represent logit coefficients.

^hStrength exercise was assessed using the strength subscale of the Rapid Assessment of Physical Activity (yes/no).

ⁱFlexibility exercise was assessed using the flexibility subscale of the Rapid Assessment of Physical Activity (yes/no).

^jHealthy diet behaviors were assessed using core and supplemental items from the Food Behavior Checklist (range 11-44; higher scores indicate healthier eating).

^kDiabetes distress was assessed using the Diabetes Distress Scale-17 (range 1-6, higher scores indicate more distress).

^lMedication adherence was assessed using the Adherence to Refills and Medications scale (range 12-48; higher scores indicate worse adherence).

Table 3.

| Within-group slopes for all outcomes, baseline to 6 months (top of the table), and baseline to 12 months (bottom of the table).^a

Variable	Total sample	DD B (95% CI) P	DD-Me-combined (Auto + Tel) B (95% CI) P	DD-Me-Auto B (95% CI) P	DD-Me-Tel B (95% CI) P
Clinical outcomesb	Baseline–6 months
HbA1c (%)	−0.17 (−0.20, −0.14) P < .001	−0.17 (−0.22, −0.10) P < .001	−0.17 (−0.21, −0.13) P < .001	−0.20 (−0.26, −0.14) P < .001	−0.15 (−0.20, −0.10) P = .001
LDL-C (mg/dL)	0.25 (−0.54, 1.04) P = .53	−0.14 (−1.48, 1.19) P = .83	0.50 (−0.46, 1.45) P = .31	0.53 (−0.73, 1.80) P = .41	0.41 (−0.96, 1.78) P = .56
SBP (mm Hg)	0.40 (0.01, 0.78) P = .04	0.72 (0.12, 1.33) P = .02	0.21 (−0.28, 0.70) P = .39	0.29 (−0.42, 1.04) P = .42	0.06 (−0.65, 0.77) P = .87
Patient-reported outcomesc
Patient–provider communicationd	0.00 (−0.02, 0.01) P = .98	0.00 (−0.03, 0.02) P = .77	0.00 (−0.02, 0.02) P = .88	0.00 (−0.02, 0.03) P = .86	−0.01 (−0.03, 0.02) P = .73
Blood glucose monitoring (d/wk)e	0.04 (0.004, 0.08) P = .02	0.05 (−0.01, 0.11) P = .08	0.03 (−0.01, 0.08) P = .16	0.02 (−0.04, 0.07) P = .52	0.04 (−0.03, 0.12) P = .28
Aerobic exercisef	0.01 (−0.01, 0.02) P = .31	0.01 (−0.02, 0.03) P = .48	0.01 (−0.01, 0.03) P = .35	0.01 (−0.02, 0.03) P = .53	0.01 (−0.02, 0.04) P = .53
Strength exerciseg,h	−0.01 (−0.06, 0.04) P = .20	−0.02 (−0.10, 0.04) P = .64	−0.09 (−0.30, 0.12) P = .40	−0.39 (−10.08, 0.31) P = .27	−0.03 (−0.23, 0.17) P = .79
Flexibility exerciseg,i	0.09 (0.03, 0.16) P = .01	0.05 (−0.04, 0.14) P = .25	0.12 (0.04, 0.20) P = .004	0.12 (0.02, 0.22) P = .02	0.10 (−0.04, 0.24) P = .16
Healthy diet behaviorsj	0.13 (0.05, 0.21) P = .01	0.12 (−0.01, 0.24) P = .08	0.14 (0.04, 0.24) P = .005	0.09 (−0.04, 0.23) P = .17	0.22 (0.08, 0.36) P = .002
Diabetes distressk	−0.02 (−0.03, −0.01) P = .001	−0.03 (−0.05, −0.005) P = .03	−0.02 (−0.03, −0.002) P = .02	0.00 (−0.03, 0.02) P = .67	−0.03 (−0.06, −0.01) P = .007
Medication adherencel	−0.15 (−0.21, −0.09) P < .001	−0.03 (−0.14, 0.09) P = .65	−0.22 (−0.29, 0.15) P ≤ .001	−0.20 (−0.35, −0.05) P = .008	−0.24 (−0.33, −0.15) P = .001
Clinical outcomesb	Baseline–6–12 months
HbA1c (%)	−0.07 (−0.09, −0.05) P < .001	−0.08 (−0.12, −0.05) P < .001	−0.07 (−0.09, −0.04) P < .001	−0.08 (−0.11, −0.05) P < .001	−0.05 (−0.08, −0.02) P = .001
LDL-C (mg/dL)	0.14 (−0.25, 0.54) P = .47	0.23 (−0.85, 1.31) P = .67	0.12 (−0.38, 0.61) P = .65	0.24 (−0.41, 0.88) P = .48	0.03 (−0.71, 0.76) P = .95
SBP (mm Hg)	0.09 (−0.09, 0.28) P = .33	0.23 (−0.09, 0.55) P = .15	0.02 (−0.20, 0.25) P = .83	−0.04 (−0.34, 0.27) P = .81	0.08 (−0.23, 0.39) P = .61
Patient-reported outcomesc
Patient–provider communicationd	0.00 (−0.01, 0.02) P = .64	0.00 (−0.02, 0.02) P = .93	0.00 (−0.10, 0.02) P = .58	0.01 (−0.01, 0.03) P = .51	0.00 (−0.02, 0.02) P = .94
Blood glucose monitoring (d/wk)e	0.04 (0.01, 0.07) P = .01	0.06 (0.01, 0.11) P = .03	0.03 (−0.01, 0.08) P = .12	0.02 (−0.04, 0.08) P = .47	0.05 (−0.01, 0.11) P = .13
Aerobic exercisef	0.01 (−0.01, 0.02) P = .64	0.01 (−0.01, 0.03) P = .50	0.00 (−0.02, 0.02) P = .88	0.00 (−0.02, 0.02) P = .87	0.00 (−0.02, 0.03) P = .73
Strength exerciseg,h	−0.02 (−0.07, 0.02) P = .30	−0.04 (−0.12, 0.04) P = .29	−0.02 (−0.12, 0.08) P = .65	−0.07 (−0.27, 0.13) P = .48	−0.02 (−0.16, 0.11) P = .74
Flexibility exerciseg,i	0.06 (0.02, 0.10) P = .001	0.02 (−0.04, 0.08) P = .45	0.08 (0.04, 0.13) P = .001	0.13 (0.05, 0.22) P = .003	0.03 (−0.04, 0.10) P = .40
Healthy diet behaviorsj	0.14 (0.08, 0.20) P = .01	0.06 (−0.03, 0.15) P = .19	0.19 (0.11, 0.27) P ≤ .001	0.13 (0.02, 0.24) P = .02	0.26 (0.14, 0.38) P < .001
Diabetes distressk	−0.02 (−0.03, −0.01) P = .001	−0.02 (−0.03, 0.002) P = .09	−0.02 (−0.03, −0.01) P = .001	−0.01 (−0.03, 0.00) P = .16	−0.03 (−0.05, −0.02) P = .001
Medication adherencel	−0.14 (−0.18, −0.09) P < .001	−0.05 (−0.14, 0.03) P = .24	−0.18 (−0.24, −0.13) P < .001	−0.16 (−0.25, −0.07) P ≤ .001	−0.21 (−0.29, −0.13) P < .001

Abbreviations: Auto, automated; DD, Dulce Digital; DD-Me, Dulce Digital-Me; Tel, telephonic; HbA1c, glycosylated hemoglobin; LDL-C, low-density lipoprotein-cholesterol; SBP, systolic blood pressure.

^aUnstandardized regression coefficients are presented and indicated the rate of change within groups shown in column headings.

^bAnalyses control for age, gender, employment, and language.

^cAnalyses control for income.

^dPatient–provider communication was assessed using the healthcare team subscale of the Chronic Illness Resource Survey (range, 1-5).

^eBlood glucose monitoring (days/week) was assessed using the blood glucose question from the Summary of Diabetes Self-Care Activities measure (range, 0-7).

^fAerobic exercise was assessed using the aerobic subscale of the Rapid Assessment of Physical Activity (5-point scale, 1-5).

^gBinary logistic regression analyses were conducted for this outcome. Unstandardized regression coefficients represent logit coefficients.

^hStrength exercise was assessed using the strength subscale of the Rapid Assessment of Physical Activity (yes/no).

ⁱFlexibility exercise was assessed using the flexibility subscale of the Rapid Assessment of Physical Activity (yes/no).

^jHealthy diet behaviors were assessed using core and supplemental items from the Food Behavior Checklist (range 11-44; higher scores indicate healthier eating).

^kDiabetes distress was assessed using the Diabetes Distress Scale-17 (range 1-6, higher scores indicate more distress).

^lMedication adherence was assessed using the Adherence to Refills and Medications scale (range 12-48; higher scores indicate worse adherence).

Figure 1.

HbA1c Over Time by Group, Abbreviations: Auto, automated; DD, Dulce Digital; DD-Me, Dulce Digital-Me; Tel, telephonic.

Per-protocol analyses of changes in HbA1c revealed no group-by-time interaction effects for any timepoint or group combination (all Ps > .05). In analyses of the total sample, significant improvement in HbA1c was observed across all groups for baseline–6 months (mean∆ per month −0.18%, 95% CI −0.21, −0.15; P < .001) and baseline–6–12 months (mean∆ per month −0.07%, 95% CI −0.09, −0.05; P < .001).

Patient-reported outcomes

Time-by-group interaction effects were found for medication adherence in the baseline–6- and baseline–6–12-month models, when comparing DD to DD-Me-combined, and to DD-Me-Auto and DD-Me-Tel, respectively (Table 2). Within-group analyses (Table 3) showed significant improvement in self-reported medication adherence over time for all DD-Me groups, but not for DD, across both timeframes. Time-by-group interaction effects were also found for flexibility exercise and a healthy diet, as well as for the baseline–6–12-month timeframe (Table 2). For flexibility exercise, the DD-Me-Auto group showed more improvement than the DD and DD-Me-Tel groups (Table 3). For a healthy diet, the DD-Me-combined and DD-Me-Tel groups showed more improvement than the DD group (Table 3). The total sample showed significant improvements in glucose self-monitoring and diabetes distress in the baseline–6- and baseline–6–12-month models (Table 3).

Proportion achieving target

For descriptive purposes, the proportion of the total sample and each intervention group that met HbA1c targets of <7%, <7.5%, and <8% at the follow-up time points are shown in Supplementary Table S2.

Discussion

This study compared the effectiveness of a previously tested digital texting intervention for type 2 diabetes, DD, to a novel, more personalized feedback approach, DD-Me, in Latine, mostly Spanish-speaking individuals with low insurance, income, and education levels. For clinical outcomes, the adaptive DD-Me intervention did not show better results than the static DD approach. However, across all groups, HbA1c decreased by >1% at 6 months when the digital interventions concluded and by 0.84% at 12 months. This level of improvement is comparable to effects seen with studies conducted using medication, continuous glucose monitoring, and intensive lifestyle interventions,^29-31 which are resource-intensive relative to this simple texting intervention. Although medication changes were not captured, we would anticipate these to occur equally across the randomized groups. Furthermore, DD and DD-Me were feasible and acceptable in the population studied as indicated in our previously published process evaluation.³² HbA1c improvements of the magnitude observed can have immediate beneficial effects of lower emergency department visits,³³ improved diabetes distress,³⁴ and if sustained, longer-term effects of lower micro- and macrovascular disease, and mortality rates.^35,36 At 6 months, 40% of participants achieved HbA1c below 7.5%, and 50% below 8%; although proportions meeting the target reduced slightly at 12 months, most were maintained. HbA1c is monitored as a care quality measure through the Healthcare Effectiveness Data and Information Set and other reporting frameworks, and health systems commonly use diabetes registries and HbA1c to identify populations for interventions. The digital interventions tested in this study can be easily integrated to enhance primary care interventions disseminated via EMR portals, during office visits, or at hospital discharge to help improve care and meet population health goals.

A previous randomized trial demonstrated the effectiveness of DD versus usual care.¹⁶ Although the original intervention was positively received, provider and patient input indicated that personalized messaging, behavioral goal setting, and feedback were desirable. In the current study, adaptive messaging was provided via 2 methods to differentiate if live interaction versus minimal touch automated approaches might have a preferential effect on engagement and outcomes. However, only small differences in clinical outcomes were observed. Greater improvements in self-reported diabetes self-management behaviors were observed in DD-Me versus DD in some cases, but these did not translate to improved clinical management in the adaptive groups. Furthermore, across groups, participants self-reported improved adherence to glucose testing, and reduced diabetes distress. Diabetes distress is common among individuals with type 2 diabetes³⁷ and predicts poorer disease management and worse health and quality of life outcomes.^38,39 Changes in behaviors and distress may have contributed to improved glycemic regulation in the current study, although these potential mediating pathways were not tested due to insufficient statistical power. Nonetheless, the DD interventions may provide a culturally appropriate approach to address diabetes distress commonly found in minoritized racial and ethnic groups.¹⁸ Interestingly, patient–provider communication did not improve following the intervention, and additional attention to effective and efficient methods of integrating digital data into the usual care process is needed. A statistically significant, although clinically small increase in SBP over time was identified, which had dissipated by 12 months. The reason for this trend is unclear.

Several limitations should be noted. First, as a comparative effectiveness trial between a known effective intervention-DD, and a new adaptive approach-DD-Me, there was no usual care control group. Dulce Digital had previously demonstrated superiority to a usual care group and therefore this study sought to establish the superiority of a strategically personalized, adaptive approach-DD-Me. This study failed to demonstrate superiority although it did demonstrate similar and significant glucose lowering among all groups, including the fully automated version-DD-Me-Auto. This may indicate that fully automated methods of delivering education and feedback via digital tools in Latine populations can be effective and less labor intensive. Second, both a previous process evaluation³² and the current report identified differences in participant engagement that influenced the degree to which the DD-Me interventions were received as intended. Specifically, in the DD-Me-Auto group, participants who responded to fewer EMA items and/or monitored glucose less frequently received less personalized feedback, resulting in an intervention more similar to the static DD condition than was intended. In the DD-Me-Tel group, participants with fewer completed calls received less personalized feedback from the health coach, and coaching feedback reports may have lacked specificity if participants provided inconsistent EMA and/or glucose data. A per-protocol analysis for HbA1c produced nearly identical findings as intent-to-treat analyses, but a more additional nuanced analysis of how intervention dosage (represented as a continuous metric) and each unique intervention component related to clinical and behavioral changes will be examined for a future report. Recent literature has documented that personalization can be beneficial in better engaging patients to use digital tools in self-management for chronic illness; however, barriers to uptake can interfere with receiving the full intervention.^40-42 We encountered this in our study and future research may need to adapt methods and further include patients in the development of digital tools to overcome engagement barriers. Third, enrollment was adversely affected by the COVID-19 pandemic. During the start of the pandemic, research staff were able to pivot to collect blood samples with minimal contact and conduct surveys telephonically to gather the greatest amount of data for participants already enrolled in the study. Despite these significant efforts, many participants remained fearful of coming to the laboratory. New enrollments that required face-to-face visits to the health center were not possible. Regular medical visits were being prioritized with distancing requirements, and research visits were eliminated. Importantly, given the small effect size of between-group differences in HbA1c change over time, even a larger sample size would have been unlikely to detect a statistically or clinically meaningful difference.

Nonetheless, given the small effect size of between-group differences in HbA1c change over time, even a larger sample size would have been unlikely to detect a statistical or clinical difference. In the event of future public health emergencies, we have learned that much of the initial outreach, consenting process, survey, and interview process can be done electronically and telephonically. Training on the digital tools may be feasible via video telehealth visits, although this may still pose some difficulty with select patients who require more hands-on support, and laboratory visits can be conducted through mobile van outreach when patients cannot attend clinic appointments. Through these experiences, we have learned how to adapt and potentially more efficiently conduct future study visits during both routine times and public health emergencies. Fourth, we did not have complete data on medication changes, which could have influenced results. Finally, diabetes self-management behaviors were self-reported, which could bias the findings. Analyses are underway examining intervention effects on changes in glucose and behaviors assessed by connected devices.

Conclusion

Static and adaptive digital interventions for Latine adults with type 2 diabetes had similar and clinically significant effects on HbA1c and improved diabetes distress and self-management behaviors. This study builds on evidence from our previous trial demonstrating the effectiveness of the culturally tailored, low-cost DD intervention versus usual care, in Latine adults.¹⁶ DD and DD-Me have the potential for integration into chronic care management of higher-risk patients needing additional education. Program enrollment can be encouraged electronically through EMR portals via mobile devices. Furthermore, dissemination by community-based peer education groups can be considered in partnership with local health centers with limited resources for supplemental DSME/S. Additional research is needed given the limitations of this trial, but the promising results to date suggest the relevance of a large-scale implementation study of DD in the future.

Supplementary material

Supplementary material is available at Annals of Behavioral Medicine online.

Author contributions

Athena Philis-Tsimikas (Conceptualization [equal], Data curation [equal], Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Resources [equal], Supervision [equal], Validation [equal], Writing—original draft [lead], Writing—review & editing [lead]), Addie L. Fortmann (Conceptualization [supporting], Data curation [supporting], Funding acquisition [supporting], Investigation [equal], Methodology [equal], Project administration [supporting], Supervision [equal], Writing—original draft [supporting], Writing—review & editing [supporting]), Taylor Clark (Formal analysis [supporting], Investigation [supporting], Methodology [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Samantha Spierling Bagsic (Data curation [supporting], Formal analysis [supporting], Methodology [supporting], Validation [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Emilia Farcas (Conceptualization [supporting], Data curation [supporting], Investigation [supporting], Methodology [supporting], Validation [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Scott C. Roesch (Data curation [lead], Formal analysis [lead], Methodology [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), James Schultz (Conceptualization [supporting], Investigation [supporting], Methodology [supporting], Project administration [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Todd P. Gilmer (Investigation [supporting], Methodology [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Job G. Godino (Conceptualization [supporting], Investigation [supporting], Methodology [supporting], Project administration [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Kimberly L. Savin (Data curation [supporting], Methodology [supporting], Project administration [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Mariya Chichmarenko (Data curation [supporting], Investigation [supporting], Writing—review & editing [supporting]), Jennifer A. Jones (Methodology [supporting], Project administration [supporting], Validation [supporting], Writing—original draft [supporting], Writing—review & editing [supporting]), Haley Sandoval (Data curation [supporting], Project administration [supporting], Supervision [supporting], Writing—review & editing [supporting]), and Linda C. Gallo (Conceptualization [equal], Data curation [supporting], Formal analysis [equal], Funding acquisition [equal], Investigation [equal], Methodology [equal], Project administration [equal], Validation [equal], Writing—original draft [equal], Writing—review & editing [equal])

Funding

Research reported in this publication was supported by the National Institute of Diabetes and Digestive and Kidney Diseases of the National Institutes of Health under Award Number R01DK112322 (Philis-Tsimikas and Gallo, Multiple Principal Investigators), the National Center for Advancing Translational Sciences UL1TR002550, and the New York Center for Diabetes Translational Research, P30 DK111022-03. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funding body had no role in the study design, collection, analysis, or interpretation, or in preparing the manuscript.

Conflicts of interest

None declared.

Prior presentation

Prior presentation in abstract form at the American Diabetes Association Scientific Sessions 2023, San Diego, California.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from all individual participants included in the study.

Study registration and analytic plan pre-registration

ClinicalTrials.gov registration: NCT03130699, Initial Release 04/24/2017.

Analytic code availability

The analytic code used to conduct the analyses presented in this study is not available in a public archive. They may be available by emailing the corresponding author.

Materials availability

Materials used to conduct the study are not publicly available. They may be available by emailing the corresponding author.

All authors have followed Annals of Behavioral Medicine’s Author Guidelines and that all authors have read and approved the paper.

Data availability

Deidentified data from this study are not available in a public archive. Deidentified data from this study will be made available (as allowable according to institutional IRB standards) by emailing the corresponding author.