Ovarian Cancer Tumor Marker Behavior in Asymptomatic Healthy Women: Implications for Screening

Ovarian cancer is the leading cause of death due to gynecological malignancy in the United States. This year an estimated 23,100 new cases of ovarian cancer will be diagnosed in the U.S., and 14,000 women will die from this disease (1). Most ovarian cancers occur in postmenopausal women, with over half occurring in women over the age of 60 years. Therapeutic advances have had a modest effect on the overall survival rate during the past three decades. Currently only 46% of women diagnosed with ovarian cancer will survive longer than 5 years. If disease is detected before dissemination beyond the ovaries,however, the 5-year survival rate is 93%. Unfortunately, only 24% of ovarian cancers are detected at this early stage (1). Development of an effective screening program for early detection,especially in high-risk women, is a top priority for reducing mortality due to ovarian cancer.

Research into noninvasive first-line screening techniques has focused on serum tumor markers and TVS.² TVS can detect ovarian cancer with a sensitivity approaching 100%, but it is insufficiently specific and too expensive for use as a first-line screen. A multimodal strategy using TVS only in women with a positive serum tumor marker screen appears to be the most promising alternative to lowering error rates and improving cost-effectiveness. Such an approach requires that all positive first-line tumor marker screenings be evaluated by TVS before a definitive surgical diagnosis is sought.

Ovarian cancers evolve from different cell types, and therefore may require different tumor-associated antigens for monitoring or diagnostic screening. Using monoclonal antibody assays, a large number of markers have been identified that are associated with epithelial ovarian carcinomas. CA125, the most extensively studied of these, is a high molecular weight glycoprotein expressed by most epithelial ovarian cancers. Elevations of CA125 in the serum may precede stage III ovarian cancer by 10 to 12 months (2), but the percentage of women with stage I disease who have an elevated CA125 level (>35 units/ml)ranges from only 23 to 50% (3, 4). CA125 has an established role in the monitoring of patients with known ovarian cancer, but in its present use has been insufficiently sensitive and specific to be very effective as a diagnostic screening tool for early-stage disease.

Other candidate markers with a potential role in ovarian cancer diagnostic screening include UGP, LASA, DM/70K, and HER-2/neu. In some populations UGP has been a sensitive marker for gynecological malignancies, with a specificity exceeding 90% for ovarian cancer (5). LASA is a marker for malignant disease on the basis of elevated sialoglycoconjugate levels associated with tumor growth (6), but it is relatively nonspecific for ovarian cancer (7). The DM/70K marker is immunologically distinct from CA125 and potentially a good adjunct; it was extracted from epithelial ovarian tumor tissue and can be found in patients with epithelial ovarian tumors of all histological types,including mucinous tumors (8). HER-2/neuencodes a transmembrane glycoprotein shed into the sera of some patients with ovarian cancer. Amplification of the HER-2/neuoncogene has been associated with poor survival in patients with ovarian cancer whose tumors overexpress this oncogene (9). HER-2/neu is amplified or overexpressed in 25 to 30% of ovarian carcinomas, though its usefulness for screening has not yet been demonstrated (10, 11).

Recent efforts to improve ovarian cancer screening have focused on modeling longitudinal tumor marker measurements, and on jointly monitoring CA125 with other markers. Skates et al.(12) proposed a screening algorithm based on a linear regression model of serial CA125 measurements which yielded a positive predictive value of 16%, substantially greater than that based on a single assay (<2%). This work helped demonstrate the substantial potential benefit of using longitudinal biomarker measurements in ovarian cancer screening. Algorithms that include multiple markers have yielded mixed results. A combination of CA125, M-CSF, and OVX-1 detected more than 95% of stage I ovarian cancers in a retrospective postmenopausal cohort, with specificity approaching 90%(13). Other studies have suggested that the joint elevation of CA125 with adjunct markers may provide a more specific test than CA125 alone, but at the expense of decreased sensitivity (14).

Cane et al. (15) described trends between first and second measurements of CA125, LASA, DM/70K, UGP, and HER-2/neu in 425 premenopausal and 165 postmenopausal healthy women enrolled in the Gilda Radner Ovarian Cancer Detection Program. These data suggested that plateaus of elevated values may be more frequent in healthy women than was previously believed. Although 80% of elevated first-screen LASA, CA125, and DM/70K markers returned to normal levels on a second screening, over half of elevated HER-2/neu and UGP values remained elevated. These findings suggested that current definitions of normal tumor marker values should be reevaluated.

If the normal behavior of tumor markers is heterogeneous among healthy women, the use of a fixed screening cutoff level will not optimally detect elevations from each individual’s baseline. Most screening protocols to date have relied on a fixed cutoff for all individuals screened and thus have not taken advantage of information on heterogeneity of tumor marker levels among individuals [the algorithm by Skates et al. (12) is one exception]. To establish optimal screening rules for what is abnormal, a better recognition of the normal behavior of tumor markers in cancer-free women is clearly needed. The main objective of this analysis was to describe such behavior of five tumor markers in a large cohort of women at high risk of ovarian cancer.

The Gilda Radner Ovarian Cancer Detection Program began in 1991 at the Cedars-Sinai Medical Center in Los Angeles (7). In its first 6 years the program enrolled 1257 asymptomatic women volunteers from the Los Angeles area through physician- and self-referral. To be eligible, a woman had to be asymptomatic, available for follow-up for the next 5 years, and have a family history of either a first- or second-degree relative with ovarian cancer, or a first-degree relative with breast cancer, other gynecological cancer, or colon cancer. After obtaining informed consent, family histories were confirmed by pathology reports or death certificates. Screening with a battery of five serum tumor markers (CA125, HER-2/neu, UGP, LASA, and DM/70K), TVS, and color Doppler imaging was performed semiannually from July 1991 until July 1995, when the protocol was changed to annual screening. Tumor marker assays were performed in blinded fashion by Dianon Systems (Stratford, CT) using techniques described previously for CA125 (3), HER-2/neu(9), UGP (5), LASA (6), and DM/70K (8). In the first 6 years of follow-up, 10 ovarian cancer cases were detected (16). These cases were not the focus of and were not included in the present analysis.

To assess for heterogeneity of tumor marker levels among women, ANOVA was used to estimate between-person variance(τ²), within-person variance(ς²), and intraclass correlation(ICC = τ²/[τ² +ς ²]) separately for all five tumor markers. ICC is the proportion of total variability in tumor marker levels accounted for by variability among mean levels of different individuals. Heterogeneity of individual means was tested using the overall ANOVA F statistic, and heterogeneity of individual variances was tested using Bartlett’s χ²statistic.

Empirical Bayes analysis was used to estimate individual-specific tumor marker means and standard deviations (17, 18). For the estimation of multiple means, empirical Bayes analysis gives better(least mean squared error) unbiased estimates than any other known method. Individual means are estimated as an empirical posterior distribution, using all available data from the cohort to form a prior distribution. The individual means are estimated intermediate between the arithmetic within-person mean and the overall cohort mean. The amount of “shrinkage” toward the overall mean is 1 minus a shrinkage factor which resembles the ICC and is a function of the number and variability of measurements observed for the individual(shrinkage factor: B(n) =τ ²/[τ² +ς ²/n]). Empirical Bayes estimation is easily extended to multivariate models using two or more tumor markers. [See Efron and Morris (19) for a good nontechnical introduction to the empirical Bayes method.]

A hierarchical Bayes linear model (20) was used to characterize the bivariate behavior of CA125 and HER-2/neu. The hierarchical Bayes approach can be used to estimate individual-level means (μ_i) and standard deviations (ς_i) by combining multiple individuals’ tumor marker distributions and to estimate the variability of μ_i in the population (τ) and the variability of ς_i in the population. This methodology can incorporate random variation at different levels(within-person [ς²] and between-person[τ ²] variance) and the effect of specific factors (e.g., menopausal status) that may cause heterogeneity in tumor marker patterns among women.

From 1991 to 1997, serial measurements of CA125, LASA, DM/70K, UGP, and HER-2/neu were obtained from 1257 women who made a total of 5358 screening visits. On average, the number of visits/woman was 4.3(ranging from 1 to 14), and the average interval between screenings was 7.8 months. Characteristics of the cohort are given in Table 1. Sixty-three percent of women were premenopausal (defined as <40 days since last menstrual period at all observations), 27.5% were postmenopausal (≥180 days since last menstrual period at all observations), and 8.4% were perimenopausal (≥40 and <180 days since last menstrual period for at least one observation or changed menopausal status during follow-up). Average ages were 56 years for postmenopausal women and 41 years for premenopausal women. Descriptive statistics for all five tumor markers by menopausal status are shown in Table 2.

Tumor marker levels were compared by menopausal status and evaluated for heterogeneity among women. CA125 and HER-2/neu mean levels were significantly lower among postmenopausal women than among premenopausal women, whereas LASA, DM/70K, and UGP were significantly higher among postmenopausal women (Kruskal-Wallis test, P = 0.0001 for all comparisons). CA125, DM/70K, UGP,and HER-2/neu distributions were highly skewed, and for subsequent analysis were converted to an approximately normal distribution using the natural logarithm transformation. Two percent of postmenopausal women and 15% of premenopausal women had first-screen CA125 values exceeding 35 units/ml, a frequently used cutoff level. Overall, CA125 levels were somewhat higher than observed in other cohorts, but the variance decomposition (the focus of this analysis)was similar to other cohort data on the logarithmic scale, which appears more appropriate for longitudinal modeling (12).

ICC was estimated separately for each tumor marker. High ICC implies that a large proportion of data variability occurs among individual mean levels, and thus is suggestive of true heterogeneity among women. When ICC is substantial in magnitude, one can expect substantial improvement when using information from longitudinal screening compared with a single threshold rule (12). Estimates of ICC were approximately 0.6 for log CA125 and log UGP, 0.4 for LASA, 0.3 for log HER-2/neu, and 0.2 for log DM/70K (Table 3). These estimates suggest considerable heterogeneity in individual tumor marker means, particularly for log CA125 and log UGP.

For all markers, individual means were significantly heterogeneous,regardless of menopausal status (overall F test, P <0.001). Within-person variances were also significantly heterogeneous for all five markers (Bartlett’s test, P < 0.001). These findings corroborated the ICC evidence for substantial heterogeneity in tumor marker patterns in this cohort.

The five tumor markers we studied behaved approximately independently as indicated by small pairwise correlations. Pearson correlation coefficients based on the total cohort data ranged from −0.13 to 0.26(Table 4). When examined by menopausal status, the magnitudes of these correlations were not substantially changed (data not shown).

Development of individual-specific screening rules relies on the estimation of tumor marker parameters for each individual undergoing screening. To illustrate this in the Gilda Radner cohort, the hierarchical Bayes model was used to estimate the mean, variance, and correlation of log CA125 and log HER-2/neu for all postmenopausal women who had at least three measurements of these markers (n = 323). Table 5 shows these estimates: the cohort mean (μ₀) and between-person SD (τ) for log CA125 and log HER-2/neu; and the between-person ρ, an estimate of correlation between individual-specific log CA125 and log HER-2/neu means. For two representative individuals, estimates of within-person mean(μ_i), SD (ς_i), and correlation (ρ_i) of log CA125 and log HER-2/neu are presented also (Table 5). Individual-specific parameters estimated in this way may be used to identify more optimally any tumor marker increase from an individual’s naturally occurring mean level. For example, in Table 5, person 1 has a log CA125 mean of 2.63 and SD of 0.34; person 2 has a mean and SD of 2.85 and 0.43,respectively. If these values were known to be reliable, an upper bound for normal could be fixed for person 1 at 2.63 + 2 × 0.34 =3.31, and for person 2 at 2.85 + 2 × 0.43 = 3.71. Because a standard cutoff is log (35) = 3.56, this implies that the first woman should have a lower cutoff than the standard, and the second should have a higher cutoff than the standard. The statistical details are complicated for how to implement this rule when individual-level tumor marker data are incomplete or uncertain. Skates et al.(12) have proposed an algorithm for CA125 that can be adapted to other single markers, but specific algorithms that can capture information from multiple markers in this manner have yet to be proposed.

Screening protocols using serum tumor markers generally have relied on a fixed cutoff level to determine abnormal values. Women at risk for ovarian cancer whose CA125 levels exceed 35 units/ml, for example, will undergo more vigilant monitoring with imaging modalities or further clinical evaluation. Such protocols have not adequately used information on the heterogeneity of tumor markers in a screening population. If tumor marker patterns are naturally heterogeneous among women, the use of a fixed cutoff level is a crude, suboptimal method for detecting increases from each individual’s baseline. The goals of this analysis were to characterize the normal behavior of five candidate tumor markers in a large ovarian cancer screening cohort and to highlight the implications for screening.

The CA125 mean level was significantly lower and the UGP mean level significantly higher among postmenopausal compared with premenopausal women. These findings have been reported previously (15, 21, 22). We also observed that DM/70K and LASA mean levels were higher, and the HER-2/neu mean level lower, among postmenopausal compared with premenopausal women. For LASA and HER-2/neu these differences were slight, but all comparisons were statistically significant.

Individual-specific tumor marker characteristics were also estimated for this cohort of 1257 women. These women had very heterogeneous patterns of all five markers studied, particularly CA125. The overall CA125 mean was 25 units/ml with wide 95% reference centiles (6–75 units/ml), indicating substantial heterogeneity of individual mean levels. The average CA125 SD was 12 units/ml, also with wide 95%reference centiles (0.7–56 units/ml), suggesting substantial heterogeneity of individual-specific variability. Fifteen percent of premenopausal women and 2% of postmenopausal women had a first-screen CA125 level exceeding 35 units/ml, and 68% and 57% of these women,respectively, had a recurrent elevation on the second screen without development of ovarian cancer. ICC, the best summary indicator of heterogeneity among individuals, was nearly 0.6 for log CA125 and log UGP, and less than 0.4 for LASA, log DM/70K, and log HER-2/neu. These findings suggest that tumor marker patterns are substantially heterogeneous even among healthy, cancer-free women.

Approximate independence of the tumor markers that we studied gives a clinically useful result. The combined false-positive rate from screening with multiple markers is well estimated by the sum of individual false-positive rates of the markers, provided that the specificity of markers is reasonably high (as is usually the case for reasonable screening candidates). For instance, in postmenopausal women that we studied, a CA125 cutoff level of 35 units/ml gave a specificity of 0.98 (the probability of a negative test in the absence of disease),and the HER-2/neu standard cutoff level of 20 units/ml gave a specificity of 0.95. Because these markers were approximately independent, the combined specificity of both tests used jointly is 0.98 × 0.95 = 0.931, and the probability of at least one false-positive test is 1 − 0.931 = 0.069. The latter is very closely estimated by the sum of individual false-positive rates (1 −specificity) for these markers, 0.02 + 0.05 = 0.07.

The heterogeneity of tumor marker patterns observed among women in this cohort underscores the need for incorporating individual-specific decision rules in screening protocols. To date, most diagnostic tests using tumor markers have not accounted for a woman’s screening history in the evaluation of tumor marker levels. In screening, a fixed cutoff level is suboptimal to a degree that depends strongly on the ICC. It is because of this phenomenon that the algorithm by Skates et al. (12) performed better than the conventional use of CA125. Using serial measurements of tumor markers,individualspecific screening rules may be developed that use all available information on the individual level as well as on the screening population level. This approach to screening, extended to multiple markers, will require information on ICCs of markers such as are presented in this study.