We retrospectively analysed the data of a prospective observational study of type 2 diabetic patients enrolled into the Hong Kong Diabetes Registry. Upon enrolment, diabetic patients undergo comprehensive assessments which follow a structured protocol whose methodology has been described previously [11–14]. Briefly, the Hong Kong Diabetes Registry was established in 1995 at the Prince of Wales Hospital, which serves a population of over 1.2 million. Since 1995, diabetic patients attending medical clinics at the Prince of Wales Hospital can be referred to the Diabetes Centre for comprehensive assessment based on the European DIABCARE protocol . Hong Kong has a heavily subsidized healthcare system, so the vast majority of patients with chronic illnesses, including diabetes, are managed in public hospitals governed by the Hospital Authority (HA), which provide for 95% of the total hospital bed-days in Hong Kong . Once the participants are entered into the Registry, their outcomes including hospital admissions will be monitored until the death of the patient [11–14]. From 1995 to the 31st December 2007, 10,129 patients have been enrolled into the cohort. After excluding 417 patients with type 1 diabetes (including missing data on classification of diabetes type), and 945 with missing variables used in the analysis, 8,767 patients were included in this analysis.
The enrolled patients periodically underwent a comprehensive 4-hour assessment for quality assurance including interview by diabetes nurses, anthropometric measurements, blood and urine tests, fundus examination and podiatry assessment. After 8 hours of overnight fast, blood was sampled for assay of fasting plasma lipids [total cholesterol (TC), HDL-C, triglyceride (TG) and calculated LDL-C], glucose, HbA1c, renal and liver function tests. We used spot urinary albumin: creatinine ratio (ACR) to define albuminuria (ACR ≥ 2.5 mg/mmol in men and ≥3.5 mg/mmol in women). The abbreviated Modification of Diet in Renal Disease Study (MDRD) formula recalibrated for Chinese  was used to define CKD as eGFR < 60 ml/min/1.73 m2. All laboratory analyses were performed using standard methods in the Department of Chemical Pathology of the Prince of Wales Hospital. The laboratory is accredited by the Australian National Association of Testing Authorities. Informed written consent was obtained from all participants and the study was approved by the Chinese University of Hong Kong Clinical Research Ethics Committee prior to its initiation.
Severe hypoglycaemia was defined as one or more hospitalisations due to hypoglycaemia 12 months prior to enrolment (‘at enrollment’), in order not to miss events and subjects with characteristics relevant to this present analysis, or during the follow-up period , as defined from enrolment to death or 31st January, 2009. Using the same definintion for severe hypoglycaemia, we had previously reported that severe hypoglycaemia identified vulnerable type 2 diabetic patients who were at risk for premature death and were associated with cancer subphenotypes . We ascertained all clinical outcomes using the HA Central Computer Management System (CMS), which records diagnoses of all hospital discharges, including mortality based on the International Classification of Diseases, Ninth Revision (ICD-9). The mortality data was cross-checked with the Hong Kong Death Registry and the cause of death was defined by the principal discharge diagnosis.
The Statistical Analysis System (Release 9.30) was used to perform all analyses (SAS Institute Inc., Cary, NC, USA) unless specified. Follow-up time was calculated as the period in years from the first enrolment to the date of death or 31st January 2009, whichever came first. All data were expressed as mean ± SD or median (interquartile range, IQR). Cox proportional hazard regression was used to obtain hazard ratio (HR) and 95% confidence interval (CI) of variables of interest. We used the Yes/No coding scheme for all major drug use at enrolment.
Immortal time is defined as the period of time without exposure to hypoglycemia from the point of enrolment into the study to the date of first hospitalization due to hypoglycemia during the follow-up period in this study. This however, may introduce immortal time bias [18, 19]. In this regard, we performed a validation study of various methods to cope with immortal time bias and found that removal of immortal time led to the least inflated hazard ratio . Thus, in this study, we excluded 896.8 person-years of immortal time from the analysis by moving the commencement point of follow-up from date of enrolment to the time of first hospitalization due to hypoglycemia during follow-up. For patients with severe hypoglycemia at enrolment, the immortal period was considered ‘0’ and their clinical profile at enrolment was used for analysis. For patients who developed severe hypoglycemia during follow-up, the enrolment was moved to the time of severe hypoglycemia during follow-up. As the metabolic profile may deteriorate over time, as in our validation study , we used multivariable linear regression to obtain partial regression coefficients of age (βa) and duration of diabetes (βb) from all other covariables at enrolment and used the estimated values of HbA1c, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP), LDL-C, HDL-C, TG, ACR, and eGFR derived from the following formulae: Xt = Xb + βa Ti + βb Ti, where Xt is the value at hypoglycemia during follow-up, Xb is the value at baseline, and Ti is the immortal time (Additional file 1: Table S1 for βa and βb). Prior cardiovascular diseases (CVD) and cancer were re-estimated taking into consideration whether these events occurred at enrolment, or during the immortal time period.
We further used relative excess risk due to interaction (RERI) and attributable proportion due to interaction (AP) [18, 21] to estimate additive interaction between hypoglycemia and CKD on all-cause death. The RERI is the excess risk due to interaction relative to the risk without exposure. AP refers to the attributable proportion of disease due to interaction in persons with both exposures. RERI >0 or AP > 0 indicates significant additive interaction.
A three-step adjustment scheme was used to control for covariables. First, we obtained the HR in univariable analysis, followed by further adjustment for age, sex, use of tobacco and alcohol, BMI, duration of diabetes, HbA1c, systolic BP, LDL-C, HDL-C, TG, natural log-transformed (spot urinary ACR + 1), eGFR, prior CVD and/or cancer at enrolment and drug use at baseline.
The plots of LOG [-LOG (Survival function)] versus LOG (follow-up time in years) were used to check proportional hazards assumption for categorical variables, while the Supremum test was used to check the assumption for continuous variables . In case of violation of the proportional hazard assumption, a stratified Cox model analysis on the variable concerned was used to adjust for its confounding effect. To avoid co-linearity, SBP but not DBP, and BMI but not waist circumference, were used in the model fitting. Pearson correlation was used to exclude highly correlated variables from the models (correlation coefficient > 0.60) . A p value <0.05 (two-sided) was considered statistically significant.
The Statistical Package for Social Sciences version 16 (SPSS, Chicago, US) was used to obtain the adjusted plot of cumulative mortality stratified by hypoglycemia and CKD and their combination over time.