The Hong Kong Diabetes Registry was established in 1995 at the Prince of Wales Hospital as part of a quality improvement program. This comprehensive registry has enabled us to examine the epidemiology and impact of treatments on clinical outcomes in real practice in Chinese diabetic patients. Hong Kong is a cosmopolitan city with 7 million people, the majority of whom are of southern Chinese ethnicity. The Prince of Wales Hospital is the teaching hospital of the Chinese University of Hong Kong and serves more than 1.2 million people in her catchment area. Since 1995 and on a weekly basis, we examined 30 - 50 diabetic patients referred from community-based primary care and hospital-based specialist clinics in an ambulatory setting. The 4-hour comprehensive assessment was performed according to a protocol modified from the European DiabCare protocol [
5]. Once patients underwent the assessment, their outcomes and clinical data are monitored until time of death. Written informed consent was obtained from all patients for research and publication purpose.
Between 1996 and 2005, 7387 diabetic patients were consecutively recruited in the Hong Kong Diabetes Registry. We excluded 2866 patients from this study due to having history of CVD at baseline (n = 1166), use of lipid-regulating drug at or before enrolment (n = 907), missing data for any variables included in the analysis (n = 522), having non-type 2 diabetes (n = 271). As a result, 4521 patients with type 2 diabetes without prior history of CVD and naïve to lipid-regulating drugs were included in the analysis.
All-cause death on or before July 2005 was recorded or otherwise censored on 30
th July 2005. We retrieved all discharge diagnoses, causes of death and drug dispensing data using the Hospital Authority Central Computer System which was used by all public hospitals in Hong Kong. A cardiovascular event was defined as coronary heart disease and/or stroke according to the International Classification of Disease, Ninth Revision [
3]. Coronary heart disease was defined as the first incidence of acute myocardial infarction (code 410) and coronary death (codes 410, 411-414, and 428), nonfatal ischemic heart disease (codes 411-414), nonfatal heart failure (code 428), and coronary revascularization (procedure code 36) and percutaneous transluminal coronary angioplasty or coronary atherectomy (procedure code 00.66). Stroke was defined as first incidence of stroke (codes 430 - 434 and 436) or death from stroke (codes 430 - 438) [
3]. The above data was matched by the Hong Kong identity card number, which is a unique identification number issued by the government for all the Hong Kong residents.
Statistical analysis
The SAS (release 9.10; SAS Institute, Cary, NC) program was used for all analysis. Follow-up time in years was estimated from enrolment to the earliest date of a CVD event, death or censoring whichever came first. We used yes/no coding for use of statins, fibrates and other drugs during the follow-up period. Cox proportional hazard regression was used to obtain hazard ratio (HR) with 95% confidence intervals (95% CI) of lipids (LDL-C, HDL-C, triglyceride) and lipid regulating drugs (i.e. statins and fibrates) for risk of CVD.
Firstly, the linearity of lipids for CVD risk was examined using restricted cubic spline Cox model analysis [
6]. Restricted cubic spline consists of piecewise cubic polynomials that are connected across different intervals of a continuous variable. It can fit sharply curving shapes [
6]. As in our previous analyses [
7,
8], we chose 4 knots at quartiles 0.05, 0.35, 0.65 and 0.95 which were suggested to offer adequate fit of the model with good compromise between flexibility and loss of precision due to overfitting of a small sample [
6]. Initially, spline terms of LDL-C, HDL-C and triglyceride were entered the spline Cox model. We then performed a forward stepwise algorithm with P = 0.10 for inclusion and removal to identify possible confounding factors including clinical and biochemical covariates at enrolment, drug use during follow-up period (See table
1 for a list of candidate covariates) and years of enrolment. Based on the HR curves of lipids, we identified threshold value for CVD risk and categorized patients using these cutoff values. Otherwise, we treated the covariates as linearly associated with CVD. We performed Cox model analysis to obtain the HR of lipids for CVD with adjustments for other lipid parameters and covariates.
Table 1
Baseline clinical and biochemical characteristics of Chinese type 2 diabetic patients with no history of cardiovascular disease (CVD) divided according to the development of CVD during 4.9 (2.8-7.0) years of follow-up
Age (years) | 54 | 21 | 64 | 16 | <.0001‡ |
Male Gender | 45.7% (1898) | | 52.8% (196) | | .0086 |
Smoking status | | | | | <.0001† |
Ex smoker | 13.3% (553) | | 20.2% (75) | | |
Current smoker | 14.9% (620) | | 21.2% (78) | | |
Waist circumference (men, cm) | 87.5 | 12.5 | 88.0 | 11.0 | .2384 |
Waist circumference (women, cm) | 82.5 | 13.0 | 84.5 | 13.0 | .0165 |
Body mass index (kg/m2) | 24.6 | 5.0 | 24.9 | 4.3 | .4409 |
Duration of diabetes (Years) | 5 | 9 | 9 | 10 | <.0001‡ |
Systolic blood pressure (mmHg) | 132 | 25 | 141 | 25 | <.0001‡ |
Diastolic blood pressure (mmHg) | 75 | 14 | 75 | 13 | .1313‡ |
HbA1c (%) | 7.2 | 2.1 | 7.7 | 2.6 | <.0001‡ |
Spot urine albumin creatinine ratio (mg/mmol) | 1.55 | 5.75 | 7.13 | 54.8 | <.0001‡ |
eGFR (ml min-1 1.73 m-2) ξ | 109 | 39 | 93 | 41 | <.0001‡ |
LDL-C (mmol/L) | 3.1 | 1.2 | 3.4 | 1.3 | <.0001‡ |
≥3.0 mmol/L | 56.1%(2330) | | 69.0%(256) | | <.0001† |
HDL-C (mmol/L) | 1.28 | 0.45 | 1.18 | 0.45 | <.0001‡ |
HDL-C <1.0 in male or 1.3 in female | 23.8%(987) | | 25.9%(96) | | .3655† |
Triglyceride (mmol/L) | 1.27 | 0.97 | 1.40 | 1.01 | .0031‡ |
Drug use at enrolment
| | | | | |
Use of antihypertensive drugs at enrolment | 29.7%(1231) | | 42.3%(157) | | <.0001† |
Drug use during follow-up
| | | | | |
Use of statins during follow-up | 22.7%(942) | | 22.6%(84) | | .9799† |
Use of fibrates during follow-up | 6.41%(266) | | 4.58%(17) | | .1639† |
Use of other lipid lowering drugs during follow-up | 0.29% (12) | | 0.27% (1) | | .3832 †† |
Use of RAS inhibitors during follow-up | 48.9% (2029) | | 62.5% (232) | | <.0001† |
Use of gliclazide during follow-up | 43.7% (1814) | | 37.2% (138) | | .0152† |
Use of rosiglitazone during follow-up | 4.9% (204) | | 1.6% (6) | | .0038† |
Use of other oral antidiabetic drugs during follow-up | 79.0% (3280) | | 80.3% (298) | | .5587† |
Use of insulin during follow-up | 32.4%(1345) | | 45.0%(167) | | <.0001† |
Events during follow-up
| | | | | |
Death during follow-up | 5.42% (225) | | 27.5%(102) | | <.0001† |
We then examined the HRs of use of statins and fibrates for CVD defined as drug use from enrolment to the first CVD event, death or censoring date whichever came first. We used logistic regression procedures and a stepwise algorithm with P = 0.30 for entry and stay to select baseline covariates including age, sex, year of enrolment, waist circumference, LDL-C, HDL-C, triglyceride, smoking status, body mass index, HbA
1c, systolic blood pressure, Ln [(urine albumin-creatinine ratio (ACR)) + 1], estimated glomerular filtration rate (eGFR), duration of disease, retinopathy and neuropathy to estimate propensity scores for adjustment of probabilities of use of statins and fibrates [
9]. Age, waist circumference, LDL-C, HDL-C, triglyceride, HbA
1c, Ln(ACR + 1), eGFR, duration of disease, retinopathy and neuropathy were selected for the propensity score for statins (c-statistic = 0.79). Age, LDL-C, HDL-C, triglyceride, HbA
1c, systolic blood pressure, duration of disease, retinopathy and neuropathy were selected to predict use of fibrates during follow-up period (c-statistics = 0.79). Stratified models on deciles of both propensity scores were used to adjust for likelihood to use statins and fibrates during follow-up period. We further adjusted for covariates identified in the initial restricted cubic spline models to analyse the effects of statins and fibrates on CVD.
Correlations between pairs of baseline covariates were checked using Pearson's correlation test and none of the pairs were highly correlated (correlation coefficient < 0.60). Proportional hazards were checked as before. A two-sided P value < .05 was considered to be significant.