The Healthy Eating Jumpstart Program (hereafter, Jumpstart) was developed as a quality improvement initiative to address food access barriers among individuals with T2D and low income or food insecurity, recognizing that limited access to affordable, healthy foods may hinder implementation of recommended dietary changes. The program enrolled individuals with T2D and low income or food insecurity in a 3-month, nonrandomized grocery delivery and educational quality improvement program without a control group. Enrollment occurred between October 2022 and May 2023. The sample size was determined based on participant flow, budgetary constraints, and guidelines for feasibility pilot studies. Details of the program protocol are published elsewhere [14,15].

As a quality improvement evaluation, the program was assessed using a pragmatic pre-post design to evaluate changes in clinical outcomes. The evaluation is focused on within-participant changes rather than estimating causal effects. This manuscript follows the SQUIRE 2.0 (Standards for Quality Improvement Reporting Excellence) checklist [16] and is limited to the clinical metrics noted in the original protocol; other program measures will be published separately.

This program was developed for quality improvement purposes and was designated as not regulated by the University of Michigan Medical School Institutional Review Board (HUM00217985). After receiving an explanation of the program details, participants were asked to verbally confirm their interest in the program during enrollment and were prompted to sign a medical record release to allow access to their clinical data. Study IDs were assigned to participants to deidentify data for analysis. All data were collected and stored via secure university-approved iterations of Qualtrics (Qualtrics Inc) or REDCap (Vanderbilt University), phone calls, and printed surveys and were stored in the protected health information–compliant Michigan Medicine SharePoint [14]. Participants did not receive compensation for participating in the program beyond food credits (US $240 over 3 months) and a free 1-year Shipt membership.

Participants were recruited from 21 primary care practices engaged in the Michigan Collaborative for Type 2 Diabetes (MCT2D), a statewide collaborative quality initiative focused on improving care for T2D [15]. The existing infrastructure of primary care teams, including case managers, social workers, and other team members at each participating practice, identified and recruited patients meeting inclusion criteria. The physician or advanced practice provider responsible for the patient’s care attested that the patient was safe to enroll in a low-carbohydrate dietary program.

Individuals were eligible to enroll if they met the following criteria: age ≥18 years, a diagnosis of T2D, residence within a Shipt delivery zone, and at least one of the following indicators of low income—Medicaid insurance status, a positive screen for food insecurity, or self-reported household income <150% of the federal poverty level, based on household income and the number of people in the household. Diagnosis of type 1 diabetes, active pregnancy or breastfeeding, or use of a sodium-glucose cotransporter-2 inhibitor were reasons for exclusion.

Grocery delivery was selected for this intervention because transportation limitations, competing priorities, and affordability barriers may restrict access to healthy foods among individuals experiencing food insecurity. Participants received a 3-month Healthy Choice Allowance (US $80 per month) through the grocery delivery service Shipt to purchase healthy foods and have them delivered to their homes. Healthy Choice Allowance–eligible foods included fruits, vegetables, proteins, canned goods, and dairy and excluded items such as ice cream, pop, and candy. Participants also received low-carbohydrate educational materials during the program, including a comprehensive packet on nutrition skills, the basics of low-carbohydrate meal planning, and grocery shopping recommendations, as well as weekly e-newsletters highlighting recipes, nutrition information, and seasonal tips. All materials are publicly available online [14].

Participants’ clinical data were abstracted from medical records at the primary care practice from 6 months prior to enrollment through 12 months after enrollment. Hemoglobin A_1c (HbA_1c), weight, and glucose-lowering medications were reviewed and abstracted into a REDCap database for analysis. Because this program was implemented as a quality improvement initiative within routine primary care, participants were not expected to collect additional HbA_1c or weight measurements beyond routine care. Therefore, the data available and timing of measurements varied between individuals. Across all metrics, participants were excluded from the analysis if a baseline measurement was not available Figure 1. All participants were eligible for data collection, and none requested to be unenrolled from the program during the evaluation period.

Data management and preparation, including summary statistics, were conducted using Microsoft Excel for Microsoft 365 MSO (version 2606; Microsoft Corp). Statistical analyses were performed using SAS (version 9.4; SAS Institute Inc).

Analyses were designed to evaluate changes associated with participation in Jumpstart. As a nonrandomized quality improvement initiative without a comparison group, analyses focused on within-participant changes between baseline and follow-up. Findings are interpreted as associations rather than causal effects solely attributable to the intervention.

Participant glycemic control was evaluated using HbA_1c measurements. HbA_1c data were only included in the analysis if the participant had a baseline HbA_1c as well as 1 additional measurement after enrollment (N=74). Participants were categorized as having controlled diabetes at baseline if their HbA_1c was <7% and uncontrolled diabetes if their HbA_1c was ≥7%. Glycemic control cutoffs were updated from 8% in the original protocol design to 7% to align with clinical metrics recommended by the ADA [17].

To evaluate changes in HbA_1c, individual participant data were assigned to time points based on their enrollment date. Baseline HbA_1c was the closest data point to their enrollment date within 21 days after enrollment. Values for 3, 6, 9, and 12 months were selected as the latest laboratory value within each time frame. Longitudinal change in HbA_1c was calculated as the absolute percentage-point difference between each time point and baseline at 3, 6, 9, and 12 months. Change in HbA_1c was averaged for each time point across the cohort. These results were then stratified by baseline HbA_1c control status as an exploratory analysis because participants with HbA_1c ≥7% have greater potential for glycemic improvement.

An overall change in HbA_1c was also assessed. Each participant’s change in HbA_1c was calculated using their latest HbA_1c value, at or after 6 months, minus their baseline value. This approach was selected to reflect the pragmatic nature of the program evaluation, in which follow-up measurements were obtained through routine clinical care rather than at prespecified research visits. Statistical significance was then assessed using a paired t test (α=.05, 2-tailed). This test was repeated with stratification by baseline HbA_1c control status. A negative value for change in HbA_1c indicates a reduction in HbA_1c. Normality of the difference scores was assessed using descriptive indices. Although data demonstrated some deviation from normality (skewness=−1.6; kurtosis=4.1), paired t tests were used due to their general robustness to moderate departures from normality given our sample size (N=68). Nonparametric methods (Wilcoxon signed-rank tests) were additionally conducted as sensitivity analyses.

Participant weight and BMI were evaluated using height and weight measurements. Weight data were only included in the analysis if the participant had a baseline weight as well as 1 additional measurement after enrollment (n=73).

To evaluate changes in weight, individual participant data were assigned to time points based on their enrollment date. Baseline weight was the closest data point to their enrollment date within 21 days after enrollment. Values for 3, 6, 9, and 12 months were selected as the latest value within each time frame. Longitudinal change in weight was calculated as the percent change at 3, 6, 9, and 12 months, using the time point value minus the baseline value, divided by the baseline value. Percent change in weight was averaged for each time point across the cohort. These results were then stratified by glucagon-like peptide-1 receptor agonist (GLP-1 RA) status as an exploratory analysis due to their highly effective use for weight management. Participants were considered to be on a GLP-1 RA if they had any GLP-1 RA use documented during the 12-month program follow-up.

An overall percent change in weight was also assessed. Each participant’s percent change in weight was calculated using their latest weight value, at or after 6 months, as the latest value minus their baseline value, divided by baseline value. As with HbA_1c, follow-up measurements reflected routine clinical care and therefore were not collected at uniform research-defined time points. Statistical significance was then assessed using a paired t test (α=.05, 2-tailed). This test was repeated with stratification by GLP-1 RA use. The percentage difference scores were approximately normally distributed (skewness=–0.2; kurtosis=0.6) for the full dataset. Consequently, paired t tests were used for the primary analyses, with nonparametric methods (Wilcoxon signed-rank tests) conducted as sensitivity analyses.

Medications were categorized as active or inactive at baseline and consolidated into their medication classes. For example, basal and bolus insulin were consolidated into a category of insulin therapy, with total daily units calculated. Medication dosing at 12 months was compared to baseline, and the changes were categorized as increased, decreased, eliminated, or unchanged [18]. The addition of a new medication was considered an increased dose. Changes within the GLP-1 RA class were assessed using dose equivalencies between medications [19]. Participants whose medical records included no specified medication use were assumed to be taking no glucose-lowering medications. The percentage of medications with increased, decreased, eliminated, or unchanged dose over the evaluation period was then assessed across the cohort and by medication category.

We enrolled a total of 83 participants between October 2022 and May 2023, with the program concluding at the end of July 2023. Participants primarily identified as female, White, and having a high school degree or less. At baseline, the mean HbA_1c was 7.6% (SD 2%), and the average weight was 105 (SD 25.8) kg. Participants primarily had a BMI ≥30.0 kg/m², with a mean BMI of 38.2 (SD 8.8) kg/m². Demographic characteristics of participants are reported in Table 1. All participants successfully placed at least 1 grocery order during the program and received mailed educational materials at their home.

Medical records were obtained for 81 (98%) of the 83 participants (Figure 1), including data from 1228 clinical encounters (excluding emergency medicine visits), 205 diabetes medications, 278 HbA_1c measurements, and 590 weight measurements. Encounter data included social work, care management, nutrition group classes, dietitian visits, diabetes education, podiatry, endocrinology, and primary care, among others.

HbA_1c follow-up data were available for 74 program participants; 55% (n=41) were categorized as having uncontrolled diabetes at baseline (HbA_1c ≥7%), and the mean HbA_1c was 7.7% (SD 2%) at baseline for this subset. At the end of follow-up, 54% (37/68) of individuals were categorized as having uncontrolled diabetes. Figure 2 (Multimedia Appendix 1) shows the average change in HbA_1c over time for the full cohort at each time point, as well as changes stratified by baseline HbA_1c control status.

The average overall reduction in HbA_1c from baseline was 0.4% (SD 1.4%; mean 7.7%, SD 2% to mean 7.3%, SD 1.6%; t₆₇=–2.4; P=.02; d=–0.3; n=68). The mean interval from baseline to the latest HbA_1c measurement was 271 (SD 58, range 164-357) days. Individuals who were categorized as uncontrolled at baseline (n=38) had an average HbA_1c reduction of 0.7% (SD 1.8%; mean 8.9%, SD 1.9% to mean 8.2%, SD 1.5%; t₃₇=–2.4; P=.02; d=–0.4). Individuals categorized as controlled at baseline (n=30) had a limited HbA_1c reduction of 0.1% (SD 0.8%; mean 6.2%, SD 0.5% to mean 6.1%, SD 0.9%; t₂₉=–0.7; P=.51; d=–0.1). These results were similar in sensitivity analyses with nonparametric testing (signed-rank P values of .02, .04, and .20, respectively).

Weight follow-up data were available for 73 program participants; 78% (n=57) had a BMI ≥30.0 kg/m² (defined as obesity) at baseline, and 60% (n=44) were prescribed a GLP-1 RA medication at some point during the evaluation period. Figure 3 (Multimedia Appendix 2) shows the change in weight over time for the full cohort, as well as changes stratified by GLP-1 RA use.

The overall reduction in weight from baseline was 1.7 (SD 4.3, range –11.6 to 6) kg, which corresponded to a 1.7% (SD 4.4%) reduction (mean 105, SD 24.6 kg to mean 103.3, SD 24.8 kg; t₆₅=–3.2; P=.002; d=–0.4; n=66). The mean interval from baseline to the latest weight measurement was 301 (SD 49, range 177-363) days. Individuals who had no GLP-1 RA use during the evaluation period had no detectable weight reduction during the evaluation, with a 1.1% (SD 4.4%) reduction (mean 99, SD 23.7 kg to mean 97.9, SD 23.5 kg; t₂₈=–1.3; P=.19; d=–0.2; n=29), while those with GLP-1 RA use experienced statistically significant weight loss of 2.2% (SD 4.3%; mean 109.7, SD 24.6 kg to mean 107.5, SD 25.3 kg; t₃₆=–2.2; P=.004; d=–0.5; n=37). These results were similar in sensitivity analyses with nonparametric testing (signed-rank P values of .005, .28, and .004, respectively).

At program enrollment, 72 (89%) of the 81 participants with medical record data were taking ≥1 glucose-lowering medication. At the end of the follow-up, 71 (88%) participants were taking glucose-lowering medications. Across the 74 participants who took 1 or more glucose-lowering medications during the program period, no change in dosing was most frequently observed (77/166, 46%), and slightly more medications were increased or added (n=52, 31%) than decreased or discontinued (n=37, 22%). Separating medication use by class, Figure 4 shows medication numbers and modifications during the evaluation period across these 74 patients (Multimedia Appendix 3).

Overall, several evaluated clinical metrics improved over the course of the program in line with our hypothesis. These changes were observed through up to 9 months of follow-up after the intervention period. The most significant clinical improvement was observed in glycemic control, with a small but statistically significant reduction in weight, and medications largely remained unchanged. These data suggest that the Jumpstart program supported participants in improving health outcomes through the provision of healthy foods and low-carbohydrate education.

The HbA_1c reduction of 0.4% at 6 to 12 months observed in our program is comparable to prior clinical trial data on LCDs, which reported reductions ranging from 0.09% to 0.61% at 3 to 6 months compared to controls [20,21]. However, prior meta-analyses indicate that these improvements are not sustained at 12 and 24 months [20,21], with HbA_1c values often returning toward baseline. In contrast, we did not observe evidence of an HbA_1c rebound after 6 months among participants with available data. The worsening of glycemic control in historical studies may partly reflect the highly restrictive clinical trial environment, resulting in low adherence to the intervention in the longer term [21-23]. The pragmatic design of the Jumpstart quality improvement program, emphasizing education to guide a shift toward a lower-carbohydrate dietary pattern rather than immediate adoption of a strict very low–carbohydrate diet, may have contributed to the absence of an observed HbA_1c rebound in our population. Participants with uncontrolled baseline HbA_1c had a larger observed reduction of 0.67%, suggesting that patients with higher baseline HbA_1c may experience greater improvement.

The change in weight after the Jumpstart program is less clear. Although a significant but small reduction in weight was noted across the full cohort, this result was attenuated when participants were stratified by GLP-1 RA use. This could indicate that GLP-1 RA use is a major contributor to weight loss rather than dietary changes. However, this stratification does not account for the duration of GLP-1 RA use, and individuals in the GLP-1 RA group could have been on the medication for more than 12 months (weight loss may have plateaued [24]) or may have taken only 2 doses during the evaluation period (which may not have been long enough to reach a therapeutic dose and influence weight [25]). Prior research has also noted suboptimal dietary intake among individuals taking GLP-1 RAs, suggesting that dietary interventions such as Jumpstart may complement pharmacotherapy and support healthier long-term eating patterns [26-28]. These findings do not rule out weight loss as a possible benefit of this program, but a more rigorous evaluation with a larger cohort would provide more information to isolate medication effects from dietary effects.

Beyond clinical impact, this program has several important strengths to note. The Healthy Eating Jumpstart provides insights into the impact and implementation of a pragmatic intervention in a high-risk population with low income and/or food insecurity that is often underrepresented in dietary intervention research. This program, administered through provider referrals at primary care clinics across the state of Michigan, suggests a feasible model for scaling combined food-is-medicine programs (food delivery+behavioral education) in the clinical setting. Successful implementation of food-is-medicine initiatives in the clinical setting is limited [12], despite the growing movement toward implementation. Pragmatic programs such as Jumpstart are needed. Health systems and community programs could consider evaluating similar dual-component interventions as a strategy to support clinical outcomes in highly vulnerable populations with diet-sensitive conditions. Additionally, this program was designed to be very low burden for patients, who simply receive access to food and educational content in their homes, without the need for additional clinic visits or counseling sessions.

The limitations of this program and its analysis are largely related to the observational design, limited sample size, and a lack of a control group in the cohort-based implementation. Because of these limitations, causal inference and generalizability are limited, and we were unable to more rigorously assess and control for medication changes or time-varying factors in our analyses. No adjustment was made for multiple testing, and findings with statistical significance should also be interpreted cautiously given the potential for type I error. The timing of the follow-up data points varied across participants because data were collected during routine clinical care. Consequently, the paired analyses are heterogeneous and reflect change from baseline to each participant’s latest available measurement rather than change at a standardized follow-up interval.

Clinic-specific variation in staffing, referral processes, and patient support may also influence program engagement and participant characteristics. Detailed food purchasing data and dietary intake were not captured in this manuscript, which limits a more direct connection between these factors and clinical metrics; however, all participants did use grocery credits during the program. No formal assessment of unintended consequences, such as participant burden or adverse dietary effects, was conducted. Additionally, although medical records included data from more than 1000 encounters, data from outside the primary care setting were limited, and records may have been incomplete, which could have resulted in missed medication changes, weight measurements, or laboratory values that would influence results. A larger-scale implementation with clinical evaluation beyond 12 months would help provide additional context on the sustainability of this program.

The results from this clinical evaluation suggest that the Jumpstart quality improvement program was associated with improvements in HbA_1c beyond 6 months and small reductions in weight among participants with available follow-up data. Because this evaluation was observational and lacked a control group, causal effects of the program cannot be determined, and medication-related influences on weight and glycemic outcomes could not be fully isolated. A low-burden program such as Jumpstart may be a feasible approach to support long-term dietary change in patients with T2D and food insecurity or low income. Similar programs may also be beneficial in conjunction with medications for glycemic control and weight loss to support sustainable lifestyle changes and long-term health goals.