Data on patients with type 2 diabetes in the Louisiana Experiment Assessing Diabetes outcomes (LEAD) cohort study were obtained through the Research Action for Health Network (REACHnet) [6,7]. REACHnet is a partnership of health systems, academic centers, and public health organizations that constitute an innovative data network for conducting efficient, multi-site research. Data from three REACHnet partner health systems were used in the current study including Ochsner Health System, Tulane Medical Center, and Access Health Louisiana. All patients’ data were extracted from EMRs in these three health systems. The dataset included electronic health record data for the study cohort between January 1, 2013 and July 31, 2018. A unique global identifier was used to link records across the three health systems to avoid duplication of individual patients in the pooled dataset. In total, 18,706 patients (out of 203,701 records) were identified as duplicates across the three partner health systems.

The definition of type 2 diabetes in the present study was formulated according to the Surveillance, PREvention, and Management of Diabetes Mellitus (SUPREME-DM) criteria [8] as follows: (1) one or more of the International Classification of Disease, Ninth Revision, Clinical Modification (ICD-9-CM) codes and Tenth Revision, Clinical Modification (ICD-10-CM) codes for type 2 diabetes associated with in-patient encounters; (2) two or more ICD codes associated with out-patient encounters on different days within 2 years; (3) combination of two or more of the following associated with out-patient encounters on different days within 2 years: 1) ICD codes; 2) fasting glucose level ≥126 mg/dL; 3) 2-hour glucose level ≥200 mg/dL; 4) random glucose ≥200 mg/dL; 5) HbA1c ≥6.5%; and 6) prescription for an antidiabetic medication. A total of 107,562 patients between the ages of 30 and 94 years were identified. After the exclusion of patients with prior CVDs and incomplete data, the present study included 67,544 patients with diabetes (40,431 whites and 27,113 African Americans). Compared with patients with diabetes excluded from the present study, the patients included had similar ages (66.5±12.1 years of age vs. 66.3±12.5 years of age) with more African Americans (40.1% vs. 36.2%) and slightly fewer men (47.5% vs. 49.1%).

The study and analysis plan were approved by the Pennington Biomedical Research Center, Tulane University, and Ochsner Health System Institutional Review Boards. Registrations of this study at clinicaltrials.gov (https://clinicaltrials.gov) or somewhere else were not applicable. We used an electronic dataset compiled from medical records but not containing personally identifiable information except for the date of birth; thus, we did not obtain written informed consent from patients in the study cohort.

The National Patient-Centered Clinical Research Network (PCORnet) common data model is a specification that defines a standard organization and representation of data for the PCORnet distributed research network [9]. Patients’ data extracted from this common data model for the present study included date of birth, race/ethnicity, sex, encounter dates, weight, height, body mass index (BMI), blood pressure, tobacco use, diagnoses of diabetes, atrial fibrillation and stroke and dates of diagnoses, laboratory test dates, total cholesterol, triglycerides, high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), HbA1c, estimated glomerular filtration rate (eGFR), and medication prescriptions such as antihypertensive drugs, glucose-lowering drugs, lipid-lowering drugs, and antiplatelet or anticoagulant drugs. These data elements were collected starting from the date of diabetes diagnosis (baseline). Using smoking status reported at each clinical visit, we classified the patients into three groups: current smokers, ever smokers, and never smokers. The eGFR was estimated using the Modification of Diet in Renal Disease (MDRD) [10].

We created the follow-up database in electronic form by using the number assigned to each patient who visited the health system with a unique patient identifier. The updated mean value of HbA1c was calculated for each participant from baseline to each year of follow-up. For example, after 1 year, the updated mean was the average of the baseline and 1-year values, and after 3 years it was the average of baseline, 1-, 2-, and 3-year values. In the case of an event occurring during followup, the period for estimating the updated mean value was from baseline to the year before the event occurred. The average number of HbA1c measurements during the follow-up period was 6.64. Stroke (ischemic or hemorrhagic) was the primary outcome in the present analysis. ICD-9-CM and ICD-10-CM codes were used to identify hemorrhagic stroke (ICD-9-CM codes 430-432 and ICD-10-CM codes I60-I62), ischemic stroke (ICD-9-CM codes 433-436 and ICD-10-CM codes I63-I64), and any stroke (ICD-9-CM codes 430-436 and ICD-10-CM codes I60-I64) events. The distributions of all ICD-9 and ICD-10 codes were: 430 (1.8%), 431 (3.4%), 432 (1.7%), 433 (39.4%), 434 (23.1%), 435 (5.5%), 436 (0.6%), I60 (1.9%), I61 (2.4%), I62 (1.5%), and I63 (18.7%). These diagnoses were recorded in the course of routine patient care by the patients’ treating clinicians. Incident stroke events were defined only when they occurred after the first record of diabetes diagnosis. The duration of follow-up for each cohort member (personyears) was tabulated from the date of the first documented diabetes diagnosis to the date of diagnosis of the outcome, death of inpatients or July 31, 2018. Either the admission date, the visiting date of emergency room or the visiting date of outpatient clinic was used when the ICD codes for stroke were first recorded. Diagnosis of stroke events could be made in either outpatient, inpatient or emergency encounters. Encounter types including ambulatory visit and other ambulatory visit were considered as outpatient encounters, while encounter types including inpatient, emergency department, emergency admission to inpatient, institutional stay, observation stay and institutional consult were considered as either inpatient or emergency encounters.

Cox proportional hazards regression was used to estimate hazard ratios (HRs) for incident stroke according to levels of HbA1c. HbA1c was evaluated in the following two ways: (1) as categories (<6.0%, 6.0% to 6.9% [reference group], 7.0% to 7.9%, 8.0% to 8.9%, 9.0% to 9.9%, and ≥10%); and (2) as a continuous variable when restricted cubic spline curves were performed. HbA1c levels were included in the models as dummy variables, and the significance of the trend across categories of HbA1c was tested in the same models by giving an ordinal numeric value for each dummy variable. The proportional hazards assumption in the Cox model was assessed with graphical methods and with models including time-by-covariate interactions. In general, all proportionality assumptions were appropriate. All analyses were first carried out adjusting for age and sex, and further for race, BMI, systolic blood pressure, LDL-C, HDL-C, triglycerides, eGFR, smoking, insurance type, use of antihypertensive drugs, use of glucose lowering drugs, use of lipid-lowering drugs, use of antiplatelet or anticoagulant drugs, and presence of atrial fibrillation. Sensitivity analyses were performed among patients who were diagnosed with stroke only in inpatient or emergency encounters. Statistical significance was considered to be P<0.05. P for trend indicated the P-value for an overall linear trend. All statistical analyses were performed by using IBM SPSS Statistics for Windows version 24.0 (IBM Corp., Armonk, NY, USA) and SAS for Windows version 9.3 (SAS Institute Inc., Cary, NC, USA).

The data that support the findings of this study are available from REACHnet but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of REACHnet.