Evaluating prediction models in reproductive medicine

Abstract

Prediction models are used in reproductive medicine to calculate the probability of pregnancy without treatment, as well as the probability of pregnancy after ovulation induction, intrauterine insemination or in vitro fertilization. The performance of such prediction models is often evaluated with a receiver operating characteristic (ROC) curve. The area under the ROC curve, also known as c-statistic, is then used as a measure of model performance. The value of this c-statistic is low for most prediction models in reproductive medicine. Here, we demonstrate that low values of the c-statistic are to be expected in these prediction models, but we also show that this does not imply that these models are of limited use in clinical practice. The calibration of the model (the correspondence between model-based probabilities and observed pregnancy rates) as well as the availability of a clinically useful distribution of probabilities and the ability to correctly identify the appropriate form of management are more meaningful concepts for model evaluation.