|
Abstract:
|
A key property to assess for a new diagnostic test is risk stratification: how well a test separates those at high risk of disease from those at low risk. We introduce the risk stratification distribution: the distribution of the changes in disease risk revealed by each test result. The mean risk stratification (MRS) is the average amount of extra disease that a test reveals for an individual patient. The MRS demonstrates that (1) big risk differences do not imply good risk stratification for markers that are rarely positive, (2) a large Youden's index, or AUC, do not imply good risk stratification if disease is too rare, and (3) risk stratification for rare diseases depends on neither sensitivity nor specificity, but on the difference of specificity and marker negativity. We provide decision-theoretic justification for MRS by demonstrating that the increase in expected benefit over the expected benefit of a random test is proportional to the MRS. We apply this framework to our experience incorporating HPV testing into cervical cancer screening guidelines.
|
