plantable cardioverterdefibrillator [ICD], resuscitation from cardiac arrest, or hospitalization for unstable angina pectoris), d) the amount of hospitalizations for CV causes, and e) hospitalization for worsening HF (WHF). The definition and adjudication of all outcomes have already been described in detail previously, as have information on C-reactive protein (CRP) and N-terminal pro-B-type natriuretic peptide (NT-proBNP) [19,213].
sFRP3 was measured from blood samples taken immediately after an overnight rapid. All other blood samples were non-fasting and analyzed on fresh samples at a central laboratory (Medical Research Laboratories, Zaventem, Belgium). NT-proBNP was analyzed using commercially offered assay (Roche Diagnostics, Basel, Switzerland). An immunonephelometric high-sensitivity system was applied to measure CRP (Dade Behring, Atterbury, UK; sensitivity 0.04 mg/L). Serum sFRP3 was measured by enzyme immunoassay (R&D Systems, Minneapolis, MN) as validated previously [16].
For all baseline variables, differences between middle-tertile sFRP3 values and the combination of the highest- and lowest tertile had been tested with Student’s t-test for normally distributed variables, Fisher’s exact test for categorical data, and Wilcoxon rank-sum test for non-normally distributed variables. Trends over sFRP3 tertiles were tested with the Cuzick extension of the Wilcoxon rank-sum test, and all baseline 141136-83-6 variables with a p-value for trend 0.05 had been included in a multivariable analysis to identify degree of association with sFRP3. All survival analyses were conducted using the Cox proportional hazard regression 
model. A restricted cubic spline (RCS) analysis with three knots was undertaken on the outcome all-cause mortality to assess linearity of risk. 12147316 The RCS analysis revealed a U-shaped curve with lower risk for patients in the middle tertile of sFRP3 concentration corresponding approximately to the pattern seen in Kaplan-Meier plots for all-cause and CV mortality. Therefore, in multivariate analyses, sFRP3 was included as a categorical (by tertiles) variable to a version of the three stage model described elsewhere [23], which included mainly clinical variables at step one (LVEF, NYHA class, age, body mass index [BMI], diabetes mellitus [DM], sex, intermittent claudication, and heart rate [HR]). At step two, estimated glomerular filtration rate (eGFR) and apolipoprotein (Apo) B/ApoA-1 ratio had been included in the model, and finally, at stage 3, the logtransformed serum concentrations of NT-proBNP and CRP had been included. Harrel’s C-statistic was calculated for all endpoints applying the full model with and without sFRP3, and the difference between the C-statistics was estimated. We implemented a jack-knife cross-validation approach to correct for over-optimism associated with validating a model in the same material from which it is developed. In this approach predictions for each observation were obtained from models developed on the remaining observations. These cross-validated probabilities have been applied to calculate jack-knife C-statistics. Calculation of the net reclassification improvement (NRI) is increasingly being utilized to evaluate the prognostic usefulness of a biomarker [24]. When no established risk categories exist, the use of a category-free NRI has been advocated [25]. We therefore calculated the category-free NRI soon after adding sFRP3 to the full model. Confidence intervals and p-values for NRI had been determined by boot-strapping with 2000 repetitions. A two-sided p-value
