Visit-to-visit variability of fasting plasma glucose as predictor of ischemic stroke: competing risk analysis in a national cohort of Taiwan Diabetes Study

This retrospective cohort, encompassing all enrollees in a National Diabetes Care Management Program (NDCMP) in Taiwan, is a population-based study of 63,084 ethnic Chinese type 2 diabetic patients enrolled in the NDCMP in Taiwan during 2002 to 2004. Date of entry into NDCMP was defined as the index date. NDCMP is a case management program established by the National Health Insurance (NHI) Bureau in 2002. Those with clinically diagnosed diabetes based on ADA criteria (International Classification of Disease, 9th Revision, Clinical Modification (ICD-9- CM) diagnosis code 250) were recruited without restriction for anti-diabetes medication. Type 2 diabetic patients treated with diverse insulin sensitizers, insulin secretagogues and insulin regimens were included. For diagnosis of a new patient, an individual with FPG >126 mg/dl (or 7.0 mmol/L) or plasma glucose ≥200 mg/dl (or 11.1 mmol/L) during an oral glucose tolerance test (OGTT) repeats the test on a different day to increase the validity of the diabetes diagnosis. Exclusion criteria were type 1 diabetes (ICD-9-CM code 250.x1/x3), gestational diabetes (ICD-9-CM code 648.83) and stroke (ICD-9-CM code 430-438), as well as being younger than 30 years of age. We enrolled patients with more than two recorded follow-ups of at least one year to rate FPG variability; 31,689 were eligible. Excluding those with missing data left 28,354 for analysis (Figure 1). We compared baseline characteristics between the patients included and those excluded using standardized mean differences. All standardized mean differences were less than 0.1 standard deviations (SD), indicating a negligible difference in means or proportions between groups. The study was approved by the China Medical University Hospital Ethical Review Board.

As of March 1995, Taiwan’s government launched an NHI program that covered approximately 99% of its population of 23.74 million in 1999 [16]. By the end of 2010, NHI contracted with 100% of hospitals and 92% of clinics island-wide [17]. The database includes patient demographics, diagnoses, and prescriptions in hospital and outpatient claims; claims data are randomly audited by the NHI Bureau. Expert reviews on a random sample for every 50 to 100 ambulatory and inpatient claims in each hospital and clinic are conducted quarterly to enhance the validity of claim data, and a severe penalty is imposed for a false diagnostic report by the NHI Bureau. This study used datasets for inpatient care by admission and outpatient visits during 2002 to 2010. Individuals in Taiwan carry unique personal identification numbers (PIN). For security and privacy purposes, patient identity data is scrambled cryptographically by the National Health Insurance Research Database (NHIRD). All NHI datasets can be interlinked with the PIN of each individual. Data comprise information for all insured subjects with regard to demographic data, date and source of diagnosis, ambulatory care, inpatient admission and outpatient/inpatient treatment. ICD-9-CM codes were used to identify individual health status. Due to comprehensive coverage of the NHI program, the proportion of enrollees withdrawing from NHI is very low; thus, bias due to being lost to follow-up is negligible. Enrollees underwent comprehensive assessment of their disease status and complications as well as a series of blood tests, urine tests and body measurements upon entering NDCMP. They completed standardized, computerized questionnaires administered by a case management nurse to record the previous or current status of their disease, medications and lifestyles. After a twelve-hour overnight fast, blood was drawn from an antecubital vein and sent for analysis within four hours post-collection. All patients were followed up regularly every three to six months.

The primary outcome measure, determined by inpatient/outpatient claims, was based on the major ICD-9-CM discharge diagnosis code, ischemic stroke (ICD-9-CM codes 433-434). Accuracy of ischemic stroke diagnoses in NHIRD had been validated in an earlier study [18]. Accuracy of ischemic stroke diagnosis in our inpatient/outpatient claim dataset was 94%, indicating NHIRD as a valid resource for population research identifying ischemic stroke. We searched NHIRD data for ischemic stroke with at least one inpatient or three outpatient claims one year after the index date, excluding incident cases within one year, to rule out cause-and-effect. Linking the unique PIN with computerized files identified 2,250 cases, followed from one year after the index date until ischemic stroke, death or withdrawal from NHI. Other chronic conditions were tabulated for 12 months prior to enrollment, using outpatient and inpatient claims data: coronary artery disease (ICD-9-CM codes 410 to 413, 414.01 to 414.05, 414.8 and 414.9), congestive heart failure (ICD-9-CM codes 428, 398.91 and 402.x1), cancer (ICD-9-CM codes 140 to 149, 150 to 159, 160 to 165, 170 to 175, 179 to 189, 190 to 199, 200, 202, 203, 210 to 213, 215 to 229, 235 to 239, 654.1, 654.10, 654.11, 654.12, 654.13 and 654.14), atrial fibrillation (ICD-9-CM code 427.31), hyperlipidemia (ICD-9-CM code 272), hypertension (ICD-9-CM codes 401 to 405), chronic hepatitis (ICD-9-CM codes 571, 572.2, 572.3, 572.8, 573.1, 573.2, 573.3, 573.8 and 573.9), chronic obstructive pulmonary disease (ICD-9-CM codes 490 to 496) and hypoglycemia (ICD-9-CM code 251).

The CV of FPG measurements from outpatient visits within the first year of the index date for each patient was calculated. FPG-CV was calculated only when more than two FPG measurements were performed in the first year. To adjust for the possibility that the number of visits may affect variation, the CV value was divided by the square root of the ratio of total visits divided by total visits minus 1 [19]. Patients were grouped into quartiles according to FPG-CV. To rule out a possible specific threshold of FPG-CV impacting the significance of the findings, sensitivity analysis classified patients into 10 subgroups according to deciles of FPG-CV. Kaplan-Meier cumulative incidence plots were derived. To weigh competing risk of death, an extended Cox proportional hazards model of competing risk data on stroke incidence fitted a proportional subdistribution hazards regression model with weights for subjects who underwent competing risk event of death, according to an extension of the Lunn and McNeil method [20]. Extended Cox proportional hazard models with competing risks served to evaluate the association between FPG-CV categories and incident ischemic stroke. Hazard ratios (HRs) and 95% confidence intervals (CI) adjusted for age, gender and multiple variables. Multivariate models (1) adjusted for age (continuous) and gender; (2) additionally adjusted for tobacco (yes/no), alcohol (yes/no), duration of diabetes, type of hypoglycemic drug, antihypertensive treatment (yes/no) and obesity (body mass index ≥27 kg/m²); and (3) additionally adjusted for coronary artery disease, congestive heart failure, cancer, hyperlipidemia, hypertension, atrial fibrillation, chronic hepatitis, chronic obstructive pulmonary disease and estimated glomerular filtration rate (eGFR). Interaction of FPG-CV and HbA₁c was probed by adding their product terms into the full model using the likelihood ratio test for significance (set at two-tailed P ≤0.05). All analyses were performed with the SAS statistical package for Windows (Version 9.3, SAS; Cary, NC, USA).