Study sample

This prospective observational study of primary care treated patients is part of the ZODIAC (Zwolle Outpatient Diabetes project Integrating Available Care) study; the design and details of which have been presented elsewhere [11]. This project started in 1998 in Zwolle, The Netherlands, and is still ongoing. Briefly, the aim of the parent study was to study the effects of two different shared care interventions in patients with T2DM. The main findings were that structured shared care with task delegation to nurses appeared feasible and could positively affect quality of care for patients with T2DM [11]. The present study is a secondary analysis of the ZODIAC study to assess the predictive capability of PENK-A, and comprises two cohorts of the ZODIAC study: one cohort started at the beginning in 1998 and the other in 2001 [12]. The first cohort contained 1143 patients and the second cohort included 973 patients, of which 427 patients were already participating in the ZODIAC study since 1998. Therefore, the combined cohort consisted of 1688 patients.

Data collection

Baseline data consisting of a full medical history were collected in 1998 and 2001. Patients were considered to have macrovascular complications when they had a previous history of angina pectoris, myocardial infarction, percutaneous transluminal coronary angioplasty, coronary artery bypass grafting, stroke, or transient ischaemic attack. Physical and laboratory assessment data, such as blood pressure, body mass index, lipid profile, creatinine levels, HbA1c and urinary albumin-creatinine ratio, were collected annually. Blood pressure was measured twice with a Welch Allyn Sphygmomanometer in supine position after at least five minutes of rest. The mean blood pressure of two recordings was calculated for each visit.

Measurement of proenkephalin A

Since mature enkephalins that derive from the PENK-A precursor peptide, are unstable bioactive peptides, an immunoassay has been developed that measures a stable precursor fragment of PENK-A, termed PENK-A 119–159 [8]. The abbreviation PENK-A used in this manuscript refers to this precursor fragment. PENK-A was measured using a sandwich immunoassay, in 1204 out of the 1688 (71.3%) patients using serum collected at baseline and kept frozen at -80 degrees Celsius until analysis according to the instructions of the manufacturer (B.R.A.H.M.S. GmbH, Hennigsdorf/Berlin, Germany). The functional assay sensitivity of this assay is 18.5 pmol/L (20% CV) (unpublished data).

Clinical endpoints

There were two clinical endpoints: all-cause and cardiovascular mortality. The vital status and cause of death were retrieved from records maintained by the hospital and the general practitioners up to the end of 2012. Causes of death were coded according to The International Classification of Diseases, 9th revision (ICD-9). Cardiovascular mortality was defined as any death in which the principal cause of death was cardiovascular in nature (ICD-9 codes 390–459).

Statistical analyses

Analyses were performed with SPSS version 18.0 (SAS Institute, Cary, NC, USA) and STATA version 12 (StataCorp, College Station, Texas USA). Continuous variables are represented as mean (± standard deviation) for normally distributed values and as median [interquartile range] for the non-normally distributed variables. Normality of the variables was examined by inspecting Q-Q plots. Univariate linear regression analyses were used to investigate whether PENK-A concentrations were associated with several clinical parameters. Due to a skewed distribution, PENK-A and serum creatinine were logarithmically transformed.

Multivariate linear regression models were used to assess the association between PENK-A and baseline characteristics. Cox proportional hazards models were used to investigate the relationship between the serum concentration PENK-A, as a continuous and a categorical (PENK-A divided into tertiles) variable, and all-cause and cardiovascular mortality with and without adjustment for selected confounders. We used 1. an unadjusted model, 2. an age- and gender-adjusted model, and 3. a model in which we additionally adjusted for the following variables: duration of diabetes, smoking (dichotomous), macrovascular complications (dichotomous), body mass index (BMI), systolic blood pressure, HbA1c, serum creatinine, cholesterol-HDL ratio and albuminuria (dichotomous). Because of missing information regarding albuminuria, BMI, smoking and duration of diabetes in 33, 2, 6 and 6 patients respectively, all analyses were performed in 1157 out of 1204 patients (96%) whose PENK-A values were determined. STATA’s ph-tests were used to test the assumption of proportional hazards for baseline predictors. Survival curves were used to illustrate the relationship between PENK-A and mortality.

For assessing the prognostic capabilities of PENK-A, we performed the following analyses stepwise. First, calibration was investigated using the Grønnesby and Borgan test assessing the goodness of fit; a nonsignificant result means an acceptable calibration [13]. Calibration is a measure of how well predicted probabilities agree with actual observed risk. Second, the Harrell’s C statistic was used to investigate the capability of each model to predict mortality and to compare how well the presence of PENK-A, in the different models used, predicted mortality [14]. The Harrell’s C value is a rank-based measure (more or less comparable with the area under the receiver operating characteristic curve). The higher the value, the better the model predicts mortality. Third, the integrated discrimination improvement (IDI) was calculated [15]. The IDI can be interpreted as the difference between model-based probabilities for events and non-events for the models with and without PENK-A. Finally, the proportion of variation explained by the survival model was estimated using R2 (D-method) [16].

In addition to investigate whether the relationship between PENK-A and mortality was affected by other clinical variables, interaction terms were added to model 3. Interaction was tested when variables were associated with PENK-A in the multivariate linear regression analyses. Interaction was evaluated as significant at a p-value of 0.10 [17].

Ethics statement

Both cohorts of the ZODIAC study (1998 and 2001) and the informed consent procedure were approved by the local medical ethics committee of the Isala hospital, Zwolle, The Netherlands. Verbal informed consent was obtained for all patients by the participating diabetes specialist nurses and the consent was documented in the patients records. According to Dutch law, written informed consent was not necessary for this type of study in 1998 and 2001. All data were analysed anonymously.

Previous studies

Notably, the 2 previous studies that both claimed that PENK-A was promising as a biomarker, did not systematically adjust for the same important risk factors as we did in our study [9][10]. Although a relevant association between PENK-A as a continuous variable and mortality (given the width of the 95% CI, model 3) could not be excluded in this study, the associations were attenuated after adding other confounders. In our post-hoc analyses without serum creatinine, PENK-A was significantly and independently associated with all-cause and cardiovascular mortality. These results could suggest that PENK-A is a biomarker for renal function. However, Ng et al. adjusted for the estimated glomerular filtration rate (eGFR), and therefore it seems unlikely that serum creatinine is the only missing confounder. Doehner et al. only adjusted for age, stroke severity and brain lesion size, and Ng et al., although correcting for eGFR, did not include smoking, BMI, albuminuria and cholesterol in the multivariate analyses [9]. Therefore, the results of both previous could have been influenced by unmeasured confounders for which we were able to adjust for.

Risk prediction

Although the IDI showed a significant improvement in discrimination for cardiovascular mortality prediction for the age- and gender-adjusted PENK-A levels, care must be taken to rely on these results because measures like IDI are not developed in the context of censored data. In addition, serious criticism has been published against the use of IDI and NRI indices [18][19]. Simulation results show that these measures can show predictive improvement even when non-informative predictors are added to a prediction model [20]. The addition of PENK-A to the age- and gender-adjusted models did not increase the C values. Although we cannot exclude that PENK-A has a role in risk prediction when PENK-A is studied as a single marker, there was no additional value when all traditional cardiovascular risk factors were available, at least not for mortality prediction in patients with T2DM. This study confirmed previous studies that it is difficult to achieve improvements in risk prediction by adding a biomarker to models with all conventional risk factors [21].

Strengths and limitations

Strengths of this study were the prospective design, the high event rate (cause of death was available for 501 patients out of 525 deceased patients) and the long follow-up period. This study also had several important limitations. Firstly, the observational design prevented us from drawing conclusions on causality. Secondly, the serum concentration of PENK-A was assessed in 71% of the total study population, which could have led to selection bias. However, no relevant differences were observed in baseline variables between the final study group and those subjects without PENK-A measurements (n = 285). Furthermore, missing of PENK-A data was related to mortality and this could have led to an underestimation of the effects of PENK-A on mortality (HR all-cause mortality 1.20 (95% confidence interval 1.00–1.44)). Furthermore, PENK-A was measured only once, therefore potential daily variability in concentration or measurement errors could not be accounted for.