GRADE was a randomized, unmasked, parallel-arm, prospective multicenter clinical trial comparing the effectiveness on glycemic control over time of four glucose-lowering medications (basal insulin glargine U100, sulfonylurea glimepiride, GLP-1 receptor agonist liraglutide, and DPP-4 inhibitor sitagliptin), when added to metformin (12). The study was approved by each center’s institutional review board, and all participants provided written informed consent.

Details of the GRADE design have been published (12), and the full protocol is available at https://grade.bsc.gwu.edu. Briefly, eligible participants were diagnosed with type 2 diabetes for <10 years at the time of screening, and were age ≥30 years (≥20 years if American Indian) at the time of diagnosis. Further, at baseline, they had an HbA1c 6.8%–8.5% (51–69 mmol/mol) and were taking 1,000 to 2,000 mg metformin daily.

After randomization, HbA1c was measured quarterly. The primary metabolic outcome was HbA1c ≥7% (53.0 mmol/mol). The secondary metabolic outcome was HbA1c >7.5% (58.5 mmol/mol). Both outcomes required confirmation. At the time of the secondary metabolic outcome, glargine therapy was to be added in the three non-insulin treatment groups and short-acting insulin in the glargine treatment group (12).

During July 2013 – August 2017, the main GRADE Study recruited 5,047 participants who were followed to April 2021. The glucagon substudy was launched in 2020, 3 years after completion of enrollment. (Appendix S1).

A sample size of at least 100 participants with the primary metabolic outcome (HbA1c ≥7%) in each treatment group would provide >80% power to detect a 40% difference in risk per SD increase in glucagon between two treatment groups (e.g. Hazard Ratio/SD = 1.82 in group A vs. Hazard Ratio/SD=1.3 in group B; A/B ratio = 1.82/1.3 = 1.4 or 40% increase) (13). A sample size of 724 participants who had completed the baseline and year 1 oral glucose tolerance test (OGTT) ensured that there would be at least 100 participants in each treatment group with the primary metabolic outcome.

A non-random convenience sample of 724 participants was selected on the basis that they had provided both 0 and 30-minute collections at both baseline and year 1 of the OGTTs done over the five years. Owing to the post-hoc selection of the sample, there was no a priori allowance for possible imbalances among the treatment groups that might lead to differences between the selected 724 versus the 4,323 not selected (Appendix S1).

An OGTT was performed at baseline and again at years 1, 3, and 5. After an overnight fast of at least 8 hours, participants drank a 75-g glucose solution within 5 minutes, with blood samples drawn at 0, 30, 60, 90, and 120 minutes. Plasma glucose and C-peptide were measured on all samples, while glucagon was measured on the fasting and 30-minute samples. Participants held their morning dose of glucose-lowering medication(s) on the day of the OGTT. Participants in the glargine group took their usual dose of medication the night before the OGTT.

HbA1c was measured in EDTA whole blood on the Automated Glycohemoglobin Analyzer HLC-723G8 (Tosoh Medics, San Francisco, CA) with an automated high-performance liquid chromatography method. Calibration of this method was evaluated using standard values derived by the NGSP. Glucose was measured in EDTA plasma by a hexokinase method on a cobas c501 chemistry analyzer (Roche Diagnostics, Indianapolis, IN). C-peptide was measured in EDTA plasma by immunoassay per the manufacturer’s instructions on a Roche cobas 6000 system (Roche Diagnostics). Glucagon was measured in aprotinin-treated EDTA plasma using a glucagon ELISA kit (Mercodia, Winston-Salem, NC).

The early glucagon and C-peptide responses were calculated as the decrement or increment, respectively, in peptide concentration from 0 to 30 minutes divided by the change in glucose concentration over the same time interval (glucagon index [GGI] and C-peptide index [CPI], respectively). The change in the GGI and fasting glucagon reflect the α-cell response to treatments, while the change in the CPI and fasting C-peptide reflect the β-cell response to treatment. A surrogate measure of insulin sensitivity, HOMA2-%S, was calculated from the fasting glucose and C-peptide concentrations using the HOMA2 calculator (https://www.rdm.ox.ac.uk/about/our-clinical-facilities-and-units/DTU/software/homa).

Cox proportional hazards models assessed the joint association of baseline measures of α-cell function (i.e. fasting glucagon and GGI) with the incidence of HbA1c ≥7.0% using a 2 degrees of freedom (df) likelihood ratio test. If significant, then each factor was tested separately using a Wald test. Like models assessed the joint association of baseline measures of β-cell function (i.e., fasting C-peptide and CPI). Additional models assessed and tested the association of the four measures of α- and β-cell function with the outcome, and to test the association of baseline α-cell function when adjusted for β-cell function and vice versa. Similar models were used to assess the association of time-varying measures of α- and β-cell function at 1, 3, and 5 years of follow-up, with estimates and tests of the joint and individual associations of α- and β-cell function with the outcomes. Unadjusted Spearman (rank) correlations described associations without adjustment for other factors.

Generalized Estimating Equations (GEE) models provided longitudinal estimates and tests of outcomes in each treatment group at each visit (averaged over treatment groups), and in each treatment group (averaged over visits), with adjustment for the baseline outcome, age, BMI, sex, race, and HOMA2-%S. These models included a treatment-by-visit interaction term to allow for different longitudinal patterns of the means by treatment group.

The p-values for the six pairwise comparisons among the treatment groups of the mean outcomes were adjusted using the Holm step-down correction (13; 14). For other analyses, simple nominal p-values are presented with no further correction for multiple tests of significance.

Both intention-to-treat (ITT) and on-treatment (OT) analyses were conducted. The ITT analyzed each participant according to their randomized treatment assignment, regardless of adherence. The OT analysis truncated participant follow-up data at the time of treatment modification, defined by stopping study metformin or randomized medication; starting protocol-defined rescue insulin upon reaching HbA1c >7.5%; or starting a non-study glucose-lowering medication. Thus, OT results are based only on follow-up data when participants were taking their prescribed study medications.

Table S1 describes participant characteristics by treatment assignment. Overall, participants were middle-aged and obese, with about two-thirds male and two-thirds self-identifying as white. On average, baseline glycemia (HbA1c and fasting glucose) was consistent with inadequately controlled diabetes. Aside from a minimal, albeit significant, difference in HbA1c (p=0.049), the baseline characteristics were similar for those selected for the glucagon substudy to the participants in the GRADE study not included in the current analyses (Table S2).

Figure 1 illustrates the cumulative incidence of the primary and secondary outcomes by treatment group using both ITT and OT analyses. The curves among groups are similar, with fewer events in the OT analysis. At Year 5, the primary outcome cumulative incidence was greatest with glimepiride and sitagliptin and least with glargine, while that for the secondary outcome was greatest with glimepiride and least with glargine. These orderings among groups differed somewhat from participants not included in the glucagon study (Figure S1) (10).

Since our goal was to understand the effect of the different medications on islet function and glucose metabolism, all additional analyses employ the OT approach.

Figure 2 illustrates the changes in several metabolic parameters at baseline and Years 1, 3, and 5 among the groups, with detailed data provided in Table S3. All assigned medications reduced the concentrations of the glycemic measures (HbA1c, fasting glucose, and 2-hour glucose). All glycemic values were significantly lower on treatment at 1, 3, and 5 years compared to baseline, except for HbA1c and 2-hour glucose at 5 years with glimepiride, and fasting glucose at 5 years with sitagliptin.

Among α-cell function measures, fasting glucagon remained unchanged from baseline with glargine and glimepiride, and decreased nominally over time with liraglutide, although this was not statistically significant. In contrast, the fasting glucagon increased with sitagliptin at 1 and 3 years relative to baseline. Glucagon at 30 minutes was nominally lower with glargine and liraglutide at 1 and 3 years relative to baseline, while with sitagliptin it was greater at 1 year than baseline. GGI did not change significantly from baseline at any time point, except in the glargine group at 3 years, and at all time points with sitagliptin, with a larger decrement compared to baseline at each year.

Among β-cell function measures, with glargine the fasting and 30-minute C-peptide were significantly reduced at all years, while CPI only differed from baseline at year 1. With glimepiride, fasting and 30-minute C-peptide were significantly greater at year 1, while CPI did not differ significantly from baseline at any year. With liraglutide, fasting C-peptide was not different from baseline at any year, 30-minute C-peptide was greater at year 1, while the CPI was numerically greater at all years, but only the increase at year 1 was statistically different. Lastly, with sitagliptin, compared to baseline, fasting C-peptide decreased significantly at year 3, 30-minute C-peptide was greater at year 1 and lower at year 5, and CPI significantly increased at years 1 and 3.

Table 1 describes the association of baseline α- and β-cell measures, and their time-dependent changes, with the risk of the primary metabolic outcome (HbA1c ≥7%) in the combined treatment groups. Tables S4 and S5 present detailed analyses by treatment group of the primary and secondary outcomes, respectively.

Baseline fasting glucagon and GGI were not associated with the primary metabolic outcome (p=0.55 on 2df). There was, however, nominal treatment group heterogeneity of the association of baseline GGI with the primary outcome (p=0.07, Table S4), with increased risk of the outcome in the liraglutide group (HR=1.27; p>0.05, Table S4). In contrast, baseline CPI, but not fasting C-peptide, was significantly associated with progression to the primary metabolic outcome (HR=0.88 p<0.05). A one standard deviation (1SD) increase in CPI was associated with a 12% reduction in risk of the primary metabolic outcome. In a model adjusted for both α- and β-cell measures, the α-cell measures remained unassociated with primary metabolic failure (p=0.55 on 2df), whereas the β-cell measures were associated (p=0.04) with risk of the primary outcome.

The association of changes in islet measures with risk of primary metabolic outcome was examined in Cox proportional hazard models adjusted for baseline age, race, sex, BMI, HbA1c, and HOMA2-%S (Table 1; Follow-up models). Changes in both fasting glucagon and GGI were not associated with risk of the primary metabolic outcome in models either unadjusted or adjusted for β-cell function (p=0.28 on 2df; p=0.35 on 2df, respectively). Likewise, changes in fasting C-peptide were not associated with risk of the primary metabolic outcome in either models unadjusted or adjusted for α-cell function (p>0.05 for both). However, changes in CPI were associated with risk of the primary outcome in both models unadjusted (p<0.05) and adjusted for α-cell function (p<0.001).

In the adjusted Cox proportional hazard models a 1SD increase in CPI was associated with a 17% reduction in risk of the primary metabolic outcome. As with the assessment of the baseline measures, there was no evidence of treatment group heterogeneity for either fasting C-peptide (p=0.16, Table S4 follow-up models) or CPI (0=0.33, Table S4 follow-up models). When the model was adjusted for the same baseline variables, excluding HOMA2-%S, the findings were the same (data not shown).

While the previous analyses addressed relationships of α- and β-cell measures with the primary metabolic outcome, we also examined relationships with the glycemic parameters as continuous variables. At baseline, both fasting glucagon and C-peptide were positively correlated with fasting glucose (r=0.144, p<0.001 and r=0.181, p<0.001, respectively), indicating that higher fasting concentrations of glucagon and C-peptide were associated with higher levels of fasting glucose before initiation of GRADE treatments.

Table 2 summarizes the associations of α-cell and β-cell measures at each year and overall with changes from baseline in glycemic measures (fasting glucose, 2-hour glucose, and HbA1c), adjusted for baseline factors. Models included interactions of treatment, visit, and α- and β-cell measures to test differences in associations by treatment group and visit.

During follow-up (Table 2), overall associations of α- and β-cell measures with the glycemic variables were significant, more so than at each specific visit due to decreased power within the visit subgroups. Associations of β-cell measures with the glycemic measures were much stronger than those of the α-cell measures. Changes in fasting C-peptide were also positively associated with changes in fasting glucose within treatment groups and years (Table S6).

As the early change in islet hormone concentrations during an OGTT has its predominant effect on post-prandial glucose responses, we examined the association of the early glucagon and C-peptide responses with the 2-hour glucose concentration (Table 2). A 1SD increase in GGI was associated with a mean 0.32 mmol/L increase in 2-hour glucose (p<0.001). Further, there was no heterogeneity in this association by visit (p=0.64) or treatment (p=0.08). Also, a 1SD increase in the CPI was associated with a 0.84 mmol/L reduction in the 2-hour glucose concentration. Again, there was no heterogeneity by treatment (p=0.38), but there was by visit (p=0.01). There was no association of CPI with the 2-hour glucose at year 5, but significant decreases of 0.8 and 1.12 mmol/L at years 1 and 3, respectively.

As both fasting and post-prandial glucose contribute to overall glycemia, we also assessed the contribution of these islet cell measures to HbA1c. At baseline, fasting glucagon was positively correlated with HbA1c (r=0.078, p<0.05). Further, change in fasting glucagon was associated with change in HbA1c overall and at year 1 (0.07 and 0.09 increase in HbA1c %-points per 1SD increase in overall fasting glucagon; p<0.01 and p<0.05, respectively). There was no evidence of heterogeneity of the associations by visit (p=0.76) or treatment group (p=0.91).

While GGI was not correlated with HbA1c at baseline (r=0.038, p=0.31), changes in GGI were associated with changes in HbA1c overall (0.07 increase in HbA1c %-points per 1SD increase in GGI; p<0.01) as well as at years 1 and 3 (0.08 and 0.09 increase in HbA1c %-points per 1SD increase in GGI (p<0.05, p<0.05, respectively). There was no evidence of heterogeneity of the associations by visit (p=0.79) or treatment group (p=0.27).

Fasting C-peptide was weakly correlated with HbA1c (r=0.048, p=0.002) at baseline. However, change in fasting C-peptide was associated with a change in HbA1c overall (0.12 increase in HbA1c %-points per 1SD increase in fasting C-peptide; p<0.001) and also years 1, 3, and 5 (Table 2). There was no heterogeneity by visit (p=0.13), but there was by treatment group (p<0.001). There was no change (per 1SD increase in fasting C-peptide) in HbA1c in the liraglutide group (−0.01, p>0.05), and increases of 0.15 (p<0.01) in glargine and 0.13 (p<0.05) in sitagliptin, and a larger increase in the glimepiride group (0.34, p<0.01).

CPI was negatively correlated with HbA1c (r=−0.247, p<0.001) at baseline. Changes in CPI were associated with changes in HbA1c when evaluated overall (0.18 decrease in HbA1c %-points per 1SD increase in CPI; p<0.001) as well as at years 1 and 3 (0.15 and 0.24 decrease in HbA1c %-points per 1SD increase in CPI (p<0.001, p<0.001, respectively). There was no heterogeneity by visit (p=0.12) or treatment group (p=0.99).