The receiver operating characteristic (ROC) curve is the most widely used measure for evaluating the accuracy of diagnostic tests in terms of differentiating between two conditions. It is known that, in certain circumstances, the characteristics of the patient or the place where the diagnostic test is performed can modify the test’s accuracy. A new estimator for the conditional ROC curve, based on direct modelling, is proposed. In this approach, the effect of covariates and false positive fraction on the ROC curve is modelled non-parametrically using generalised additive models (GAM) combined with local polynomial kernel smoothers. The method allows for incorporation of more than one covariate in the regression model for the ROC curve and the possible interaction between them. The proposed model’s performance is examined in an in-depth simulation study. Finally, endocrine data are analysed with the aim of assessing the performance of several anthropometric measures in predicting clusters of cardiovascular risk factors in an adult population in Galicia (NW Spain), with adjustment for age and gender.