Page 26 - Biomarkers for risk stratification and guidance in heart failure

Multimarker risk score in emergency department dyspnea
Statistical analysis
Data are presented as frequencies, mean±SD or median (interquartile range, IQR). Comparisons between groups were performed using chi-square for categorical
data and 1- way analysis of variance or Kruskal-Wallis H test for continuous data,
as appropriate. Receiver operating characteristic (ROC) curve analysis was used 2 to assess prognostic accuracy of biomarkers and to determine optimum cut-off
points (i.e. maximizing both sensitivity and 1-specificity) of continuous variables for predicting 90-day mortality. Cut off points were rounded off to make them clinically meaningful. Spearman’s rank correlations were used to test correlations between biomarkers.
Logistic regression analysis was performed to test the association between biomarkers and 90-day mortality. Multivariable analysis was performed for clustered variables (i.e. for clinical covariates) and laboratory findings separately. We included variables that were univariably associated with 90-day mortality (stepwise with P<0.1 as the cut off for entry). Thus, in a first step, a final clinical model and a separate final biomarker panel were established from multivariable analysis. In a second step, the final biomarker panel was added in a stepwise fashion to the final clinical model, which resulted in the final prediction model. We checked for collinearity and interactions among covariates and found none of significance. Model accuracy, calibration and discrimination were evaluated as recently suggested 21 by (i) c-statistic, a measure of the area under the curve (AUC), (ii) the Hosmer-Lemeshow statistic, (iii) integrated discrimination improvement (IDI), and (iv) net reclassification index (NRI). Risk categories of <2%, 2-15% and >15% were used for calculation of the NRI 22.
Independent predictive variables in the final prediction model formed the basis for our risk score. When simplifying the score, a loss in AUC of ≥1% was not accepted. The risk score was internally validated by cross-validation (90% of original sample, 10 replications) and by non-parametric bootstrapping (1,000 resamples using random sampling with replacement), as proposed 23.
The PRIDE (ProBNP Investigation of Dyspnea in the ED) mortality score was calculated as proposed 6. The additional predictive value of the final biomarker panel on top of the PRIDE mortality score 6 was investigated for both 90-day and 1-year mortality in a multivariable logistic regression model and tested by C-statistic, NRI and IDI.
25