The GRADE study was a prospective, multicenter, randomized trial conducted from 2013 to 2021 in the U.S. that compared the glycemic impact of the addition of one of four glucose-lowering medications to metformin: glargine U-100 (a long-acting insulin), glimepiride (a sulfonylurea), liraglutide (a glucagon-like peptide 1 (GLP-1) receptor agonist), and sitagliptin (a dipeptidyl peptidase 4 (DPP-4) inhibitor) in individuals with T2D (11). The protocol, statistical analysis plan, and full list of inclusion/exclusion criteria are available at ClinicalTrials.gov (NCT01794143). Briefly, GRADE enrolled 5047 participants with diagnosed T2D <10 years, HbA1c of 6.8–8.5% (50.8–69.4 mmol/mol), age ≥30 years (≥20 years for Native Americans), and treated with metformin alone for glucose-lowering treatment (minimum 1000 mg/day and up to 2000 mg/day). At the time of first occurrence of HbA1c >7.5% (confirmed, based on HbA1c measured quarterly) after the addition of randomized glucose-lowering medication, insulin glargine was added to the medication regimen for the non-glargine treatment groups and insulin aspart was added for the glargine group. At the next occurrence of HbA1c >7.5% (confirmed), insulin aspart was added and the randomized medication was discontinued for the non-glargine treatment groups. All clinical centers obtained institutional review board approval for the protocol and informed consent, and all study participants provided written informed consent.

Participants’ resting 12-lead ECG was obtained at baseline, year 2, year 4, and year 6 study visits using a GE/Marquette MAC1200 ECG machine (GE Healthcare, Miluakee, WI). All ECGs were centrally read and coded at the Epidemiological Cardiology Research Center at the Wake Forest School of Medicine (Winston-Salem, NC). After visual inspection for technical errors and inadequate quality, ECGs were automatically processed using Marquette 12-SL software running under GE MUSE (GE Healthcare). The Minnesota Coding system was applied to classify ECG tracings as either major or minor abnormalities (12). Mappings between Minnesota Codes and major or minor abnormalities are given in Supplementary Table 1.

Major ECG abnormalities included: major ventricular conduction defects, definite myocardial infarction (indicated by major Q wave abnormalities), possible myocardial infarction (indicated by minor Q/-QS wave plus major ST/-T abnormalities), major isolated ST/-T abnormalities, left ventricular hypertrophy, major atrioventricular conduction abnormalities, major QT prolongation, pacemaker use, and others.

Minor ECG abnormalities included: minor isolated Q/-QS waves, minor isolated ST/-T abnormalities, high R waves, ST-segment elevation, incomplete (left or right) bundle branch block, minor QT prolongation, short PR interval, left axis deviation, right axis deviation, frequent ventricular premature beats, and others.

In addition to the above ECG abnormalities, two time-domain HRV measures of ECG-derived CAN were calculated based on the average of three 10-second digital rhythm strips: the SD of all normal-to-normal R-R intervals (SDNN) (milliseconds), and the root mean square of successive differences between all normal-to-normal R-R intervals (RMSD) (milliseconds). The definitions of abnormal HRV were SDNN <17.13 ms and RMSD <24.94 ms, based on our previously validated ECG-derived CAN measure using the gold-standard cardiovascular reflex tests (13).

The following baseline factors were analyzed: diabetes duration (tertiles), HbA1c (tertiles), weight (tertiles), height (tertiles), waist circumference (tertiles), history of hyperlipidemia, medication use (statin, angiotensin-converting-enzyme (ACE) inhibitor, angiotensin receptor blocker (ARB), beta blocker, lipid-lowering medication), systolic (<120, 120–<140, or 140+ mmHg) and diastolic (<80, 80–<90, or 90+ mmHg) blood pressure, alcohol use (never, occasionally, or at least weekly), age (<45, 45–<60, or 60+ years), sex, race (Black, White, or all others), ethnicity (Hispanic vs. non-Hispanic), income (<$20k, $20k-<$35k, $35k-<$50k, $50k-<$75k, $75k+), and education (high school/GED or less, some college, college, or graduate school). Body mass index was calculated as weight (kg) divided by height (m) squared (BMI, tertiles). Diabetic peripheral neuropathy was assessed using the Michigan Neuropathy Screening Instrument (MNSI) (14–15); an abnormal score was defined as an MNSI clinical score of 2.5 or more or an MNSI instrument score of 7 or more (15). Estimated glomerular filtration rate (eGFR) was calculated using the CKD-EPI Creatinine Equation (16) and categorized as <90 or 90+ mL/min/1.73m².

The GRADE study randomized 5047 participants who underwent quarterly assessments for an average follow-up of 5.0 years (17). This analysis included 4769 participants (94%) with an ECG available at baseline, and participants were included for each follow-up study visit at which an ECG was performed. Among the 4769 participants with baseline ECGs, there were 4437 participants (93%) with ECGs at year 2 and 3519 participants (74%) with ECGs at year 4. A CONSORT flow diagram is shown in Supplementary Figure 1. Most of the analysis sample had less than 6 years of follow-up (3647 participants, 76%), and the COVID-19 pandemic caused logistical challenges at year 6. There was thus a relatively low number of participants with an ECG at year 6 (992 participants, 21%) and year 6 data were excluded from the analyses.

Descriptive analyses summarized baseline characteristics of participants with versus without any ECG abnormalities and with versus without ECG evidence of CAN across the four treatment groups (these p-values were unadjusted for multiplicity).

Prevalence of ECG abnormalities (minor, major, and any) and ECG-derived CAN at baseline were summarized by treatment group.

Incidence of new ECG abnormalities by treatment group was analyzed using separate generalized estimating equation (GEE) logistic regression models for minor, major, and any ECG abnormalities. The models included treatment group, time-point (i.e., year 2 or 4), and the interaction between treatment and time. Because participants with more baseline ECG abnormalities were ‘eligible’ for fewer new ECG abnormalities at future time points, the models adjusted for the number of relevant baseline ECG abnormalities. The models also adjusted for baseline heart rate, blood pressure (systolic and diastolic), lipids (total cholesterol, HDL, triglycerides), and use of beta blockers. Dependence of the outcomes over time was incorporated into each model by modeling the log odds ratio between year 2 and 4 outcome variables (18). For each model, a Wald test assessed an overall treatment effect over time. Analyses compared the liraglutide treatment group to the three non-liraglutide groups combined, and compared all four treatment groups. The liraglutide vs. three non-liraglutide groups comparison was a prespecified analysis included in the Statistical Analysis Plan. For analyses comparing all four groups, a gatekeeping strategy was used. Separately for each outcome of minor, major, and any ECG abnormalities, first an overall Wald test of treatment group differences was conducted. If the overall test was significant, then Wald tests for all 6 pairwise treatment comparisons were conducted, with these 6 tests adjusted for multiple comparisons using the Holm procedure (19).

Incidence of ECG-derived CAN by treatment group was analyzed using a similar procedure to that used for incidence of new ECG abnormalities, restricted to participants who did not have ECG-derived CAN at baseline. Models were adjusted for baseline heart rate, blood pressure (systolic and diastolic), lipids (total cholesterol, HDL, triglycerides), and use of beta blockers. Indices of HRV indicative of CAN (SDNN and RMSD) were analyzed over time using GEE linear regression models. The distributions of SDNN and RMSD were right-skewed, hence the indices were log-transformed. The models included treatment group, time-point, and the interaction between treatment and time, and were adjusted for the baseline HRV index, heart rate, blood pressure, lipids (total cholesterol, HDL, triglycerides), and use of beta blockers. An autoregressive AR(1) working correlation matrix was used. A procedure similar to the one used to test for treatment differences in incidence of new ECG abnormalities was used to test for treatment differences in the log-transformed HRV indices over time.

Heterogeneity in the treatment effect on ECG abnormalities (minor, major, and any) and ECG-derived CAN was assessed based on the baseline factors listed in the Other Baseline Measures section above. For each baseline factor, separate GEE logistic regression models were fit for minor, major, and any ECG abnormalities and for ECG-derived CAN, with covariates for treatment group, the baseline factor, time-point. Models for and all interactions between treatment, the baseline factor, and time-point. Models for new ECG abnormalities (minor, major, and any) also adjusted for the number of baseline ECG abnormalities as a covariate. These models were unadjusted for other baseline adjustment factors due to concerns about collinearity with the baseline stratification factor. Treatment effect heterogeneity by the baseline factor with respect to the longitudinal progression of ECG abnormalities or ECG-derived CAN was tested with a Wald test of the interaction of treatment with the baseline factor for all time-points, separately for each outcome of minor/major/overall ECG abnormalities and CAN. Separately for each outcome, the 24 tests (corresponding to 24 baseline factors) were adjusted for multiplicity using a 1% false discovery rate (FDR) (20). For each baseline factor for which treatment effect heterogeneity was not significant at the 1% level, the marginal association with longitudinal progression of ECG abnormalities or ECG-derived CAN was tested by fitting a model without treatment and performing a Wald test of the effect of the baseline factor across all time-points, adjusted using a 1% FDR among baseline factors for which the test of treatment effect heterogeneity was not significant (separately for each outcome).

All statistical models used available case analysis (i.e., all available ECGs from baseline, year 2, and year 4 were included in each model, even for participants who were missing a year 2 and/or year 4 ECG).

The Glycemia Reduction Approaches in Diabetes: A Comparative Effectiveness Study (GRADE) is funded by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK). This manuscript is based on follow-up data and outcome assessments from the 5047 participants enrolled into the study. This database is available in the NIDDK Central Repository at https://repository.niddk.nih.gov/study/151.

There were no differences in baseline characteristics across treatment groups (Supplementary Table 2). Table 1 presents baseline characteristics of participants with versus without any ECG abnormalities, and with versus without ECG-derived CAN at baseline. Participants with any ECG abnormality were older; more likely to have longer duration of diabetes, be male, and White; have higher height, weight, BMI, and waist circumference; have higher systolic blood pressure; and more likely to have a history of myocardial infarction or stroke, be a current or past smoker, and use guidelines-directed therapies (lipid-lowering medications, ACE inhibitor or ARB, beta-blockers) (all nominally p<0.05).

Participants with ECG-derived CAN were also older, with longer diabetes duration and higher HbA1c, more likely to be male and White, have higher waist circumference, higher systolic blood pressure, a lower eGFR, and were more likely to have a history of myocardial infarction, diabetic peripheral neuropathy (defined by an abnormal MNSI), history of hyperlipidemia, statin and/or other lipid-lowering medication use, ACE inhibitor or ARB use, and beta blocker use, compared to participants who did not have ECG-derived CAN at baseline.

Due to attrition at year 4, baseline characteristics were compared between participants with and without an ECG at year 4 (Supplementary Table 3). Although there were a few statistically significant differences at baseline between those with and without a year 4 ECG for sex, education, diabetes duration, weight, and LDL, these differences were small and not clinically meaningful. In addition, there were no differences found between those with and without a year 4 ECG in terms of treatment group assignment or cardiovascular disease history (i.e., history of myocardial infarction and stroke) at baseline, and so attrition bias due to missing year 4 ECGs is expected to be minimal.

Table 2 displays descriptive prevalence statistics of ECG abnormalities and of the HRV indices used to define CAN at baseline, overall and across treatment groups. The baseline prevalence of ECG abnormalities and of ECG-derived CAN was high, at 57% for any abnormality and 53% for CAN. The baseline distributions of ECG abnormalities, CAN, and HRV indices were similar across the four treatment groups, as expected. Baseline prevalence statistics for specific ECG abnormalities are presented in Supplementary Table 4.

The incidence analyses are restricted to participants with non-missing data, including for all of the baseline adjustment factors (7956 post-baseline ECGs, 99.2% of sample).

Figure 1 displays the percentages of all participants with minor, major, and overall ECG abnormalities at the baseline, year 2, and year 4 visits. At baseline, 51% and 15% of participants had a minor or a major abnormality, respectively (57% for any abnormality). These percentages are relatively similar at year 2 and rise to 53% and 18% of participants at the year 4 visit (61% have any abnormality).

Figures 2a–2c (Supplementary Table 5) display the estimated incidence of ECG abnormalities at years 2 and 4 post-baseline by the 4 treatment groups, for any minor abnormality, any major abnormality, and any minor or major abnormality, after adjustment for baseline risk factors. There is no significant evidence that the incidence of ECG abnormalities varies across the four treatment groups (p=0.90 for any minor abnormality, p=0.13 for any major abnormality, p=0.81 for any minor or major abnormality).

Figures 2d–2f (Supplementary Table 5) display the model-estimated probabilities of ECG abnormalities by the liraglutide and non-liraglutide treatment groups, respectively for any minor abnormality, any major abnormality, and any minor or major abnormality, after adjustment for baseline risk factors. There is evidence that incidence of major ECG abnormalities varied between the liraglutide and non-liraglutide groups (p=0.03), with higher incidence for the non-liraglutide groups over time. In particular, the model-estimated probability of any new major ECG abnormality was 6.9% (95% CI: 5.4%, 8.6%) for the liraglutide group vs. 8.0% (7.0%, 9.2%) for the non-liraglutide groups at year 2, and 9.3% (7.5%, 11.5%) for the liraglutide group vs. 12.7% (11.3%, 14.3%) for the non-liraglutide groups at year 4.

For CAN incidence, the analyses include only the participants who did not have ECG-derived CAN at baseline and were not missing any data, including all the baseline adjustment factors (1909 participants, 78.6% of sample with no ECG-derived CAN at baseline; 3230 ECGs, 91.9% of post-baseline ECGs among those with no ECG-derived CAN at baseline).

Figure 2g (Supplementary Table 5) displays the estimated incidence of ECG-derived CAN at years 2 and 4 for the liraglutide and non-liraglutide groups, after adjustment for baseline risk factors. There was no evidence that the incidence of ECG-derived CAN was different when comparing the liraglutide and non-liraglutide groups overall across the year 2 and 4 visits (p=0.42).

Figures 2h and 2i (Supplementary Table 5) display the model-estimated HRV indices used to define CAN analyzed as continuous variables for the liraglutide and non-liraglutide groups at years 2 and 4. SDNN was higher at year 2 in the liraglutide group compared with the non-liraglutide group (p=0.02 for overall test across years 2 and 4); in particular, the model-estimated mean logarithm of SDNN was 2.77 (95% CI: 2.74, 2.81) for the liraglutide group vs. 2.69 (2.66, 2.72) for the non-liraglutide groups at year 2, and 2.69 (2.63, 2.75) for the liraglutide group vs. 2.67 (2.64, 2.71) for the non-liraglutide groups at year 4. However, no significant treatment group differences were observed for RMSD (p=0.48).

We assessed whether the effect of liraglutide vs. non-liraglutide treatment on ECG abnormalities (any minor, any major, and any minor or major) and CAN differed by categories of the baseline variables described in the Other Baseline Measures section above. There were no significant treatment effect interactions between baseline factors and rates of ECG abnormalities or ECG-derived CAN during follow-up (see Supplementary Table 6).

Based on separate regression models, each of the following baseline factors were significantly associated with higher incidence of minor, major, and overall (i.e., any minor or major) ECG abnormalities at years 2 and 4 (see Supplementary Table 6): higher systolic blood pressure, higher weight (minor and overall ECG abnormalities only), higher height (minor ECG abnormalities only), higher waist circumference (minor and overall ECG abnormalities only), diabetic peripheral neuropathy (defined as an abnormal MNSI), lower eGFR, use of beta blockers, use of lipid-lowering medications (overall ECG abnormalities only), older age, and non-Hispanic ethnicity compared to Hispanic ethnicity (minor and overall ECG abnormalities only) (all p<0.01). Participant age at study entry was significantly associated with rates of ECG-derived CAN during follow-up (p<0.0001), with older age being associated with higher incidence of ECG-derived CAN at years 2 and 4.

We conducted a sensitivity analysis where the dataset was truncated to study visits prior to the COVID-19 pandemic, i.e., all visits up to and including March 15, 2020. All analyses were repeated and the overall results, shown in Supplementary Figure 2, mirrored those from the main analyses (shown in Figure 2).

To address concerns that some models may suffer from multicollinearity, and consequently a potential loss in precision in the treatment effect estimates, we repeated all analyses without adjusting for any covariates in regression models. The results are shown in Supplementary Figure 3 and were similar to those from the main analyses (shown in Figure 2). As expected, the significant treatment effects between liraglutide and the non-liraglutide groups for any new major ECG abnormality and for the log of SDNN were also found in the unadjusted analysis (Supplementary Figure 3e and 3i), but with slightly larger p-values than in the covariate-adjusted analysis, likely due to increased precision from covariate adjustment.