Cancer patients and matched control subjects from CARET were all postmenopausal women who were current or former smokers and who had a mean age of 59 years (SD = 5.7 years). All were white, except for one matched case–control pair of African American women. The available information indicated that this population was at average risk for ovarian cancer. The distributions of most risk factors are generally comparable between cancer patients and control subjects (). Mean standardized baseline serum levels of all six markers were slightly higher in cancer patients than in control subjects, but the only marker with a difference that approached statistical significance was spondin-2. The mean (±SD) CA125 level at baseline was 16.7 U/mL (±36.6 U/mL) among control subjects and 22.4 U/mL (±52.3 U/mL) among cancer patients. Most women provided blood specimens on two or more occasions (range = 1–11 occasions) between 0 and 18 years before diagnosis among cancer patients () and during a comparable time interval before the reference date among control subjects.
Baseline characteristics of ovarian cancer patients and matched control subjects in Carotene and Retinol Efficacy Trial (CARET)
Times of blood collections and ovarian cancer diagnosis among women who developed ovarian cancer during the Carotene and Retinol Efficacy Trial. Open circles = times of blood collections; solid circles = time of ovarian cancer diagnosis.
Among control subjects, modest pairwise Pearson correlations were found between levels of spondin-2 and B7-H4 (r
= .41, P
= .003), DcR3 (r
= .44, P
< .001), and CA125 (r
= .37, P
= .002). Among cancer patients, the only correlation that reached statistical significance at the P
≤ .008 level was between spondin-2 and DcR3 (r
= .59, P
< .001) (Supplementary Table 1
, available online).
Available tumor characteristics indicate that 16 of the 34 cancer patients were diagnosed with serous carcinoma, including 15 known to be at advanced stage (). The other histologies observed included mucinous (n = 5), adenocarcinoma not otherwise specified (n = 4), and endometrioid (n = 3). Early-stage tumors were primarily mucinous. Of the 23 cancers with available tumor grade information, 13 (57%) were anaplastic, seven (30%) were poorly differentiated, and three (13%) were moderately differentiated.
Distribution of tumor characteristics*
Similar CA125 protein levels were observed between cancer patients and control subjects until approximately 3 years before diagnosis of ovarian cancer, at which point (by visual inspection), the mean marker level among cancer patients began to rise (, A). A similar pattern, though less pronounced, was observed for HE4 protein levels and to a lesser extent, for mesothelin protein levels (, B and C). Levels of B7-H4 and DcR3 in cancer patients and control subjects were indistinguishable throughout the study (, D and E). Spondin-2 levels showed a slight increase over time among cancer patients resulting in a small separation during the final year before diagnosis (, F).
Figure 2 Lowess curves of standardized marker levels by time before diagnosis or reference date. Standardized marker levels were rescaled to have a mean of 0 and a SD of 1 among control subjects. A) CA125. One cancer patient and one control subject had standardized (more ...)
In descriptive analyses, the discriminatory power of the individual markers, as assessed by ROC methods, was limited (with AUC statistics that ranged from 0.56 to 0.75) but showed increasing accuracy with time approaching diagnosis (, A–F). For CA125, the AUC statistics that were based on 68, 18, and 14 samples, respectively, were 0.57, 0.68, and 0.74 for the intervals of 4 or more years, 2–4 years, and less than 2 years before diagnosis. A finer division of the time axis gave a stronger gradient in AUC statistics, with an AUC of 0.89 for the final year before diagnosis (Supplementary Figure 1, A
, available online). ROC curves for HE4, mesothelin, B7-H4, DcR3, and spondin-2 provide a similar pattern of generally improving classification as the time to diagnosis decreased (, B–F and Supplementary Figure 1
, B–F, available online).
Figure 3 Receiver operating characteristics curves by time before diagnosis. Standardized biomarker levels were used for this analysis and were rescaled to have a mean of 0 and a SD of 1 among control subjects. A) CA125. B) Human epididymis protein 4. C) Mesothelin. (more ...)
Lowess curves of mean levels over time before diagnosis of composite marker 1, defined for each observation on each woman as the sum of her standardized levels of CA125, HE4, and mesothelin, indicated that the level of this composite marker began to rise 4–5 years before diagnosis in cancer patients (, A). Summing all six markers (ie, composite marker 2) did not alter this pattern substantially (, B). Composite marker 3, defined for each observation on each woman as the maximum of her standardized biomarker levels of CA125, HE4, and mesothelin, and composite marker 4, similarly defined as the maximum of all six standardized levels at each time point, also indicated a change in the levels of these composite markers among cancer patients at approximately 3 years before diagnosis (, C and D). The corresponding ROC curves and AUC statistics for composite markers 1–4 (, E–H) indicated only small improvements in classification performance over individual markers.
Figure 4 Lowess curves of standardized marker levels by time before diagnosis or reference date (A–D) and corresponding receiver operating characteristic curves by time before diagnosis (E–H) for composite markers. A and E) Composite marker 1 (defined (more ...)
The above analyses use marker data in a retrospective fashion, evaluating their performance against a known time of diagnosis. In practice, however, a marker would be assessed and decisions made on the basis of a woman's estimated probability of being diagnosed with cancer, conditional on currently available marker levels and other risk factors or symptoms, but without any information regarding time to diagnosis.
We used Cox regression models to assess the value of CA125, HE4, mesothelin, B7-H4, DcR3, and spondin-2 individually and in combinations (ie, composite makers 1–4) in this prospective setting. In the model evaluating CA125 alone, an elevation in CA125 level of 1 SD was associated with an increased risk of ovarian cancer (hazard ratio [HR] = 1.42, 95% confidence interval = 1.18 to 1.70; P < .001) (), implying that women in this population with a CA125 level of 53 U/mL would have an incidence rate that was approximately 1.4 times higher than that of comparable women with a CA125 level of 16 U/mL. In separate models, HE4, mesothelin, and spondin-2 were associated with statistically significantly elevated risks of ovarian cancer (with HRs ranging from 1.35 to 1.58) (). Composite markers 1–4 were also associated with statistically significantly increased risks of ovarian cancer; however, use of all six markers (ie, composite markers 2 and 4) did not provide stronger results than those that were based solely on CA125, HE4, and mesothelin levels (composite markers 1 and 3).
Associations between biomarkers and ovarian cancer risk*
Ovarian cancer is a heterogeneous disease, with serous histology being the most prevalent and one of the most lethal subtypes in postmenopausal women. Serous tumors are almost always detected at a late stage. Biomarkers that have shown promise for ovarian cancer have been chosen primarily for their ability to identify patients with late-stage serous disease. We hypothesized that any signal observed in the overall group might be stronger in a more homogeneous group that contained only serous tumors. We repeated the Cox regression models in the small subgroup of 16 patients with serous ovarian cancer and found somewhat higher risks associated with all of the individual markers and composite markers, except for B7-H4 (), although only the hazard ratios for CA125, HE4, or spondin-2 and the four composite markers’ risk reached statistical significance.
To examine these markers jointly, we used a forward stepwise procedure to select the most predictive set of individual markers within the same general regression model. Only CA125 and mesothelin entered this model. In analyses limiting cancer patients to those with serous tumors, only CA125 and HE4 were found to be predictive of ovarian cancer ().