A plea for AUC confidence intervals in diagnosis models used in gynecology

Over the last decade many studies in the gynecology literature have been investigating the performance of diagnosis models such as Univariate, Risk of Malignancy Index (RMI) and Logistic Regression (LR). Typical performance results are claimed in terms of sensitivity (SEN), specificity (SPE), accuracy (ACC), Positive Predictive Value (PPV), Negative Predictive Value (NPV), with some studies als including Receiver Operating Characteristic (ROC) curve and its Area Under the Curve (AUC). It remains, however, that all these measures do not reflect any sample size and thus making it sometimes difficult to assess with confidence the true performance of these diagnosis models, in particular for small sample size. In this paper, we propose to use systematically, a ROC-based methodology that makes possible to calculate the Confidence Interval (CI) at each ROC point. The methodology is generic and robust to sample size, and based on Probability Density Function (PDF) without any assumption on the distribution. We illustrate its use on 6 recent studies and show that results with the additional AUC 95% CI contour is more adequate to compare the performance of these diagnosis models, especially with studies using different sample size.