Table 1. Measures used to assess biomarkers for improvement in risk prediction.

Target and measure	Comments
Classification of subjects into risk categories
Proportion of subjects who are reclassified when adding the biomarker to an existing model [17]	Includes both correct and incorrect reclassifications
Net reclassification improvement (NCI) [31,32,26]	Rewards correct and penalizes incorrect reclassification; does not distinguish between up- and down-classification
Reclassification table: cross-tabulation of the proportion of subjects who are classified into each risk category by the models with and without biomarker, and comparison of the observed event rate with the assigned risk category. [28,19,32,17]	Straightforward, clinically relevant interpretation; no clear ordering of models; sensitive to choice of risk categories.
Discrimination accuracy
C-statistic (area under the ROC curve) [16,20,21]	Insensitive to moderate improvements in risk prediction; may not be clinically relevant
Integrated discrimination improvement (IDI) [31,26]	Extension of the NCI; does not depend on cut-points for risk categories
Graphical presentation
Plot of predicted risk against the risk percentile for models with and without the biomarker [19,18,15,16]	Intuitively, a stronger model should show a wider spread of predicted risks; does not measure the accuracy of the risk prediction [16]
Model calibration
Hosmer-Lemeshow statistic [32,33,23,17]	Compares estimated with observed proportions of subjects with events over several risk intervals; may be too sensitive in large samples. [32]
Global model fit
Akaike information criterion (AIC), and Bayesian information criterion (BIC) [32,23,17]	Widely used for variable selection; no direct clinical interpretation

Abbreviations: ROC, receiver operating characteristic, NCI, net reclassification index