Study Design
In the Copenhagen area (Copenhagen Municipality and the former Copenhagen County) all laboratory workup from 1.3 million inhabitants was performed at a single laboratory, serving both general practitioners (GPs – 85%) and practicing specialists (15%) from 2000 through 2015, the Copenhagen General Practitioners’ Laboratory, CGPL. The laboratory analyzed a broad range of blood, urine, semen, clinical physiological tests, cardiac tests, and lung function tests. The CopLab Database contains all numerical results (n = 112 million) of these tests from these individuals as described in detail previously.(4) Measurement of HbA1c, lipids in blood and estimated glomerular filtration rate based on plasma creatinine was performed as previously described.(5) Data from the CopLab Database have been merged with nationwide administrative registries by use of unique personal identification numbers assigned to all Danish residents.(6) Thus, by use of these registries we were able to collect information on hospitalizations, prescribed medications, education level, cohabitation, income, and vital status on an individual level. (7, 8)
Study population
Individuals were included at the time of their first HbA1c measurement in the CopLab database between 2007 through 2015 and followed for at least 3 years. Persons were excluded if they were younger than 30 years of age at study entry. Individuals with a history of AF or stroke prior to their study entry time were also excluded. The following comorbidities were identified by discharge diagnoses during a 10-year period before first measurement of HbA1c: ischemic heart disease, cancer, AF, renal disease, chronic obstructive pulmonary disease (COPD) hypertension and diabetes (Supplementary Table 1). Prior surgical procedures in the form of percutaneous coronary intervention (PCI) and coronary artery bypass grafting (CABG) were classified according to NOMESCO Classification of Surgical procedures. Baseline use of pharmacotherapy was defined by one or more redeemed prescriptions within 6 months of study entry (Supplementary Table 2). Use of the following drugs was assessed: angiotensin-converting enzyme inhibitors (ACE-I) or angiotensin-II receptor blockers (ARB), antiplatelets, mineralocorticoid receptor antagonists (MRAs), statins, β-blockers, diuretics and glucose lowering drugs. History of diabetes was identified by at least one filled prescription for glucose lowering drugs or one in- or outpatient contact.
Outcomes
Our primary outcome was new onset of AF. Stroke, cardiovascular mortality, and all-cause mortality were evaluated as secondary outcomes. Persons were followed until date of emigration, death or December 31st, 2018.
Exposures
At study entry persons were stratified according to diabetes status; Previous diabetes was defined by hospital discharge diagnosis for diabetes prior to study entry or filled in prescription for glucose-lowering drugs prior to study entry. We implemented four groups: 1) No diabetes and HbA1c < 42 mmol/mol, 2) no diabetes and HbA1c 42–47 mmol/mol (prediabetes), 3) New onset diabetes (HbA1C ≥ 48 mmol/mol and no known diabetes) and 4) Previous diabetes.
Statistical analyses
Baseline characteristics according to diabetes status were described by use of proportions for categorical variables and means with standard deviations or medians with quartiles for continuous variables. To investigate the effect of diabetes status on the outcomes we estimated the (cause-specific) hazard ratios (HR) using Cox proportional hazards models. The Cox models for AF, stroke, CVD and all-cause mortality were adjusted for age (spline), sex, year of baseline (spline), ischemic heart disease, cancer, renal disease, COPD and hypertension. Stratification on categorical covariates were used when adequate to meet model assumptions. Cumulative incidence curves were based on predictions from these Cox models. We further estimated the relative hazard of the outcomes across the range of HbA1c values by use of penalized splines.
In a sensitivity analysis, we further adjusted the models for creatinine. However, due to a sizeable amount of missing creatinine measurements, that we assumed were missing at random, we used multiple imputation by substantive model compatible fully conditional specification. (9, 10) For each outcome, we first fitted 11 imputed datasets based on the imputation models. Covariates were adjusted for as described above. We then fitted the Cox models for the primary outcomes on each imputed data and gathered the estimated HRs using Rubin’s rule.
All analyses were carried out using R (4.1.1).