The best magazine
Methods
Swedish National Diabetes Register
The Swedish National Diabetes Register was initiated in 1996 as a caregiver tool for local quality assurance and to provide feedback as part of diabetes care. Trained doctors and nurses report annually to the register, either online or through clinical record systems; no stringent criteria exist for how often patients visit an outpatient clinic. Information is collected during appointments at hospital outpatient clinics and primary healthcare centres nationwide. Several previous reports have been published concerning trends in risk factor control and risk prediction based on the register, including a more detailed description of the register and the Swedish healthcare system for patients with diabetes.
Patient Involvement
There was no patient involvement in this study. The work within the Swedish National Diabetes Register, as this article, is done in a continuous but informal dialogue with people with diabetes.
Participants
We included 18 168 people with type 1 diabetes entered in the Swedish National Diabetes Register for whom data were available about the use of insulin pump therapy or multiple daily injections. A total of 2441 people were being treated with insulin pump therapy during the study period from baseline to the final year, and 15 727 were treated with multiple daily injections during the whole study period to final year. Type 1 diabetes was epidemiologically defined as all patients who received insulin treatment only (for diabetes mellitus) and were aged under 30 at onset, almost all of whom had been reported by outpatient clinics from about 90 Swedish hospitals. Baseline appointments took place in 2005–07 with follow-up until 31 December 2012. Treatment with insulin pump has been documented in the register since the year 2004. All individuals were recruited from the Swedish National Diabetes Register with no exclusion criteria.
Examinations at Baseline and the End of the Study
Clinical characteristics at baseline included type of glucose lowering treatment, age, duration of diabetes, sex, HbA1c, systolic blood pressure, diastolic blood pressure, height, weight, waist circumference, physical activity, smoking habits, total cholesterol, high density lipoprotein cholesterol, triglycerides, microalbuminuria, plasma creatinine, use of antihypertensive drugs, lipid lowering drugs and acetylsalicylic acid (aspirin), atrial fibrillation, and histories of cardiovascular disease, heart failure and atrial fibrillation, Furthermore, baseline yearly income (in Swedish kroner), marital status (single, married, divorced, or widowed) and educational level (lower (up to school year 9), intermediate (years 10–12 of upper secondary school), and higher (college/university)) were obtained from the Longitudinal Integration Database for Health Insurance and Labour Market Studies, Statistics Sweden. Body mass index (BMI) was calculated as weight/height2. Waist circumference (cm) was measured at the height of the navel. Physical activity was graded as low (no activity or less than once a week) or higher (twice or more a week). Smoking was defined as one or more cigarettes a day, one pipe a day, or having quit within the past three months. The Swedish standard for recording blood pressure as used by the Swedish National Diabetes Register is the average (mm Hg) of two supine readings (Korotkoff sounds I-V) with a cuff of appropriate size after at least 5 minutes of rest. Analyses of HbA1c were quality assured nationwide by regular calibration with the HPLC Mono-S method and then converted to mmol/mol. Albuminuria was classed as urine albumin excretion >20 μg/min on two out of three consecutive tests (microalbuminuria or macroalbuminuria). A history of cardiovascular disease was defined the same way as for the outcome; ICD-10 (international classification of diseases, 10th revision) code I50 for heart failure; code I48 for atrial fibrillation; C00-C097 for all cancer; codes K70–74 for liver disease; and codes F20–29 and F30–39 for mental disorders.
We estimated updated mean HbA1c during the study period using all values from baseline until the year before an event occurred during the study or otherwise from baseline until 31 December 2012. Change in HbA1c during the study period was estimated as the difference between baseline and final measurements, the latter estimated as the value before the year of an event or otherwise the value in 2012. Hypoglycaemic attacks that required a hospital admission, with ICD-10 codes for hypoglycaemia and coma from the hospital discharge register, were entered during the study period from baseline until 31 December 2012.
Follow-up and Definition of Endpoints
All individuals were monitored from the baseline examination until death or the first incident or until 31 December 2012. The mean follow-up period was 6.8 years, with a total of 114 135 person years. The major primary endpoints were fatal or non-fatal coronary heart disease, fatal or non-fatal cardiovascular disease, fatal cardiovascular disease, and total mortality. Non-fatal coronary heart disease was defined as non-fatal myocardial infarction (ICD-10 code I21), unstable angina (ICD-10 code I20.0), percutaneous coronary intervention, and/or coronary artery bypass grafting. Fatal coronary heart disease was defined as ICD-10 codes I20-I25. Stroke was defined as fatal or non-fatal cerebral infarction, intracerebral haemorrhage, or unspecified stroke (ICD-10 codes I61, I63, I64). Cardiovascular disease was defined as the composite of coronary heart disease or stroke, whichever came first. A secondary endpoint was mortality from non-cardiovascular disease.
A history of heart failure was defined as ICD-10 code I50, and atrial fibrillation before the study start was defined as ICD-10 code I48. All events were obtained by linking to the Swedish cause of death and hospital discharge registers, a reliable validated alternative to revised hospital discharge and death certificates.
Statistical Analysis
We applied five imputations using the Markov chain Monte Carlo technique for missing data in the sample of 18 168 individuals, using the SAS MI and MIANALYSE procedures. We recorded baseline clinical features as mean values (SD) or frequencies (%) of each multiple imputed variable in the two treatment groups (insulin pump therapy or multiple daily injections) and calculated significance for crude differences between the two groups with Student's t test or χ test. We used crude Kaplan-Meier curves for all outcomes to compare the two groups with log rank test and for observed hypoglycaemic episodes during study follow-up.
We estimated a propensity score for treatment with pump with logistic regression as the conditional probability of being treated with pump given the baseline characteristics, including the covariates age, sex, duration of diabetes, history of cardiovascular disease, heart failure, atrial fibrillation, baseline HbA1c, systolic and diastolic blood pressure, BMI, total and high density lipoprotein cholesterol, triglycerides, cumulative microalbuminuria, creatinine, renal insufficiency, smoking, physical activity, antihypertensive drugs, lipid lowering drugs and aspirin, educational levels, yearly income, marital status and baseline years. We calculated P values for differences between the two treatment groups after adjustment with the propensity score, including all 36 variables, estimated by generalized linear models (link id for continuous data and link logit for dichotomous data). We also computed standardised differences between the two groups; a difference of less than 10% was considered to be satisfactory. The distribution of the propensity score stratified in fifths was calculated for the two treatment groups, as well as the number of outcomes by each fifth of the score (Appendix Table A).
We used Cox regression analysis to estimate hazard ratios with 95% confidence intervals for outcomes comparing insulin pump treatment with multiple daily injections. Covariate adjustment was performed by stratification with fifths of the propensity score.
The proportional hazard assumption of the Cox regression analyses was tested by adding an interaction term of the predictor and log time and by analysing Schoenfeld residuals—both were found to be non-significant and satisfied the proportional hazard assumption. Interactions between the two treatment groups and all covariates included in the propensity score were analysed by means of maximum likelihood estimation; no interactions were found between any covariates.
Unmeasured confounders can affect the results if they are unrelated to, or not fully accounted for by, measured confounders or if they affect the decision to use insulin pump treatment and not multiple daily injections (treatment allocation bias). We therefore performed a sensitivity analysis by quantifying the effects of a hypothetical unmeasured confounder when comparing individuals treated with insulin pump therapy and multiple daily injections with an algorithm as presented by Lin and colleagues, yielding progressive changes of the hazard ratios observed for the analysed outcome with presence of an unmeasured covariate of 1.3 or 1.4 and present more frequently in one of the groups (injection) than the other.
All statistical analyses were performed with SAS version 9.3 (SAS Institute, Cary, NC). A two sided P<0.05 was considered significant.
Source: ...