Title: Optimal cutpoints for classification in medical diagnostic tests
Authors: Monica Lopez Raton - Conselleria de Educacion- Xunta de Galicia (Spain) [presenting]
Carmen Cadarso Suarez - Universidad de Santiago de Compostela (Spain)
Elisa Maria Molanes Lopez - Universidad Complutense de Madrid (Spain)
Abstract: Continuous diagnostic tests (biomarkers or risk markers) are often used to discriminate between healthy and diseased populations. For their clinical application, the key aspect is how to select an appropriate cutpoint or discrimination value $c$ that defines positive and negative test results. In general, individuals with a test value smaller than $c$ are classified as healthy and otherwise as diseased. In the literature, different optimality criteria there exist to select $c$. We consider an interesting cost-based generalization of the Symmetry point (the optimal $c$ that maximizes simultaneously both types of correct classifications) incorporating the misclassification costs, and we propose confidence intervals for this optimal cutpoint and its sensitivity and specificity measures using two approaches: a parametric approach based on the Generalized Pivotal Quantity (GPQ) under normality and a nonparametric approach based on the Empirical Likelihood (EL). In addition, we develop two R packages, OptimalCutpoints and GsymPoint, to facilitate clinicians selecting optimal cutpoints in their daily practice. A new classification rule is also proposed by logistic generalized additive regression models (GAMs), that provides an improved discriminatory capacity where traditional Receiver Operating Characteristic (ROC) analysis is not valid, being necessary more than one optimal cutpoint on which to base the classification.