Toward an optimal algorithm for ovarian cancer screening with longitudinal tumor markers

Stored samples from women in the Stockholm screening study were reassayed for CA125II (Centocor, Malvern, PA) and OVX1. The postmenopausal women older than age 50 without ovarian cancer were randomly split into a training set to develop a screening test based on longitudinal marker levels and a second set to validate the test. The CA125II data from each woman is summarized by the slope and intercept from a linear regression of log(CA125II) on time since first sample. The slope versus the intercept for the training set and the ovarian cancer cases formed a bivariate scatter plot. A curve was drawn on the scatter plot that separated most of the women with ovarian cancer from all other women; it delineated a screening test. The specificity of this test was examined on the validation set with a specificity of 99.8%. Bayes' theorem was used to calculate the risk of ovarian cancer (ROC) based on the intercept, slope, and assay variability. It is important to account for assay variability because it can produce large slopes over short periods of time. The maximum risk, which identified 83% (5 of 6) of the ovarian cancers detected within a year of last assay, was applied as a test to the training set and confirmed a high specificity of 99.7%. With this specificity and sensitivity, the ROC algorithm using the CA125II assay has an estimated positive predictive value of 16%, substantially greater than the positive predictive value based on a single assay. Further study is planned to confirm the sensitivity of this approach.