|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
© 2003 American Society for Clinical Oncology Calculation of the Risk of Ovarian Cancer From Serial CA-125 Values for Preclinical Detection in Postmenopausal Women
From the Massachusetts General Hospital and Harvard Medical School, Boston, MA; Gynaecology Cancer Research Unit, St Bartholomew’s and Royal London Hospitals Medical College, London, England; Division of Gynecologic Oncology, Cornell Medical School, New York, NY. Address reprint requests to Steven J. Skates, PhD, 50 Staniford St, Suite 560, Boston, MA 02114; email: sskates{at}partners.org.
Purpose: Previous studies of CA-125 levels from screening trials for ovarian cancer have indicated that serial CA-125 levels may identify cases better than a fixed CA-125 cutoff. We conducted a study to assess the screening performance of the risk of ovarian cancer calculation based on serial CA-125 levels from prospectively collected serum samples compared with a fixed CA-125 cutoff. Patients and Methods: The calculation was applied to data from a prospective trial of screening for ovarian cancer involving 22,000 postmenopausal women older than 45 years. The analysis was performed using 33,621 CA-125 results from 9,233 women for whom two or more serial samples were available. All serum samples from the patients with ovarian cancer were obtained before clinical detection. Sensitivity and specificity levels for preclinical detection of index cancers were calculated for various cutoffs for the risk and a single CA-125 measurement, and receiver operator curves were constructed. Results: The risk calculation significantly improved the area under the curve from 84% to 93% compared with a fixed cutoff for CA-125 (P = .01). For a target specificity of 98%, the risk achieved a sensitivity of 86% for preclinical detection of ovarian cancer, whereas CA-125 achieved a sensitivity of 62%. The estimates of performance are unbiased, because the risk calculation was derived independent of the data from this trial. Conclusion: These results provide the first evidence that preclinical detection of ovarian cancer using serial CA-125 levels interpreted with the risk calculation significantly improves screening performance compared with a fixed cutoff for CA-125. The results justify the incorporation of the risk calculation in a prospective, randomized, controlled trial.
OVARIAN CANCER leads to more deaths among women in the developed world than all other gynecologic malignancies combined.1 This poor prognosis has been attributed to the high proportion of cases that present at a late stage. Because there is a close correlation between stage at presentation and survival, detection of asymptomatic ovarian cancer in its early stages is an appealing approach to reducing mortality from this disease. Trials conducted in London and Stockholm during the last decade have yielded encouraging results about the performance of the serum tumor marker CA-125 in ovarian cancer screening.2,3 These studies have used a fixed cutoff to interpret CA-125 results, and there is concern that the sensitivity of this use of CA-125 for asymptomatic early-stage disease may be limited.3 To improve the performance of a CA-125–based screening program, an exploratory, graphical analysis of data from the London and Stockholm trials was undertaken. The analysis demonstrated that CA-125 levels increase rapidly over time in most women with preclinical ovarian cancer. In contrast, women without ovarian cancer have relatively stable CA-125 values over time even if the initial value is elevated.5 In general, the distinct behavior of serial CA-125 in women with and without ovarian cancer indicates that systematic analysis of sequential CA-125 levels may improve the performance of a CA-125–based screening program. Calculation of the risk of ovarian cancer from the CA-125 pattern was developed using data from the previous Stockholm study.4 The risk calculation is an estimate of the probability of having preclinical ovarian cancer based on the pattern of CA-125 values and age. The initial analysis was limited by the small number of cancers for which serial samples were available. We have now been able to analyze more than 33,000 serial CA-125 results taken from more than 9,000 women, with a median follow-up of 6.8 years (range, 4.67 to 9.13 years; 2.9% lost to postal follow-up) in the Royal London/St Bartholomew’s Hospital (SBH) study.5 This analysis of an independent data set of postmenopausal women at normal risk provides an unbiased estimate of the risk calculation’s operating characteristics for preclinical detection.
Setting The design of the SBH trial has been described in a previous report.3 In summary, 22,000 postmenopausal (≥ 1 year of amenorrhea) women 45 years of age or older who were recruited between June 1, 1986, and May 1, 1990, underwent a prevalence screen and were then assigned, using a computer-generated, random number, to either annual interval screening for three incidence screens (n = 10,977) or follow-up without further screening (n = 10,958). Serum samples were collected prospectively, with the last collection on July 31, 1995. At each screen, venipuncture for CA-125 was performed, and women with levels exceeding 30 U/mL were referred for an ultrasound scan. Serum was frozen for subsequent analysis. Women with abnormal scan results were referred for a gynecologic opinion. All volunteers were traced for all cancers by annual postal questionnaires and, in July 1997, via the National Cancer Registry for cancers reported through to the time of the inquiry. For all index cancers, defined as primary invasive epithelial ovarian or fallopian tube cancer, surgical and histopathologic details were obtained from the hospital involved.
CA-125 Analysis
Definitions
Subjects and Samples
Statistical Analysis For comparison with a fixed cutoff, the last CA-125 level available for each subject from the prevalence or interval screens was chosen. All CA-125 values before detection of ovarian cancer were included in the risk calculation for each case. In constructing operating curves for preclinical detection, 200 cutoffs were evenly distributed on the logarithmic scale for CA-125 and the risk calculation, ranging from 6 to 500 U/mL and from 0.01% (10 in 100,000) to 100%, respectively. For each cutoff, the sensitivity and 1-specificity were calculated and graphed to form operating curves. Area under the curve (AUC) was calculated for both the risk and a fixed cutoff for CA-125. The AUC estimates the probability of correctly classifying an individual randomly chosen from the population. The AUC for the risk and a fixed cutoff for CA-125 were compared using the method of DeLong et al,7 and sensitivities for a fixed specificity were compared using McNemar’s test. The fixed specificity of 98% was chosen, because the SBH study attained this level of specificity for CA-125 as a first-line test.5 This corresponds to a cutoff of 41 U/mL for CA-125II and 2% for the risk calculation. Constructing operating curves for early-stage disease is not possible with the available data, because we do not know when a subsequently diagnosed late-stage index cancer was in early-stage disease.
Distribution of Risk and CA-125 Values In the SBH trial, results from 33,621 serial CA-125 estimations from 9,233 women with two or more samples (no index cancer, 9,212; index cancer, 21) were available for analysis. Figure 1  displays the distribution of CA-125 values and risk levels for women with and without an index cancer. There is a greater separation of women with an index cancer from women without an index cancer in the histograms of the risk compared with CA-125 measurement. This greater separation is summarized in Fig 2, which shows the relationship between sensitivity and specificity in operator curves for the risk and a single CA-125 level. The risk calculation significantly increases the AUC compared with CA-125; the AUC for CA-125 is 84.2%, whereas the AUC increases to 93.2% for the risk estimate (P = .01). The 95% confidence interval for the difference in AUC is 1.4% to 16%, which excludes a difference of zero.
Sensitivity and Specificity Figure 2  shows the relationship between specificity and sensitivity for the risk calculation and a fixed cutoff for CA-125. At levels of sensitivity greater than 70%, there was a steep decrease in specificity for thresholds based on a fixed cutoff for CA-125. In contrast, this decrease in specificity did not occur for the risk until sensitivity reached levels of more than 90%. Although the risk calculation achieved a PPV of 15.5% by itself for a sufficiently high cutoff level, we aimed to increase the sensitivity, because in clinical practice, an elevated risk calculation would be followed by ultrasound. In the SBH study, a specificity of 98% and a PPV of 2% achieved using a fixed cutoff for CA-125 as a first-line test made it possible to achieve an acceptable and high overall specificity and PPV after use of ultrasound as a secondary test. Because the intention is to use ultrasound as a secondary test after an elevated risk, we focused on the same specificity and PPV targets for the first-line test. At the target PPV of 2% and a specificity of 98%, the risk calculation using a 2% cutoff achieved a significantly higher sensitivity than a fixed cutoff of 41 U/mL for CA-125 (86% v 62%, respectively; P = .025). Ideally, sensitivity results should be subdivided by stage of disease; however, there were too few included cases to subdivide further and retain meaningful sensitivity estimates.
Serial CA-125 Values in Cases and Noncases
Ovarian cancer screening has the potential to reduce mortality in a disease whose prognosis has improved minimally during the last three decades. However, to date, no randomized controlled trial of ovarian cancer screening with sufficient power to demonstrate a reduction in mortality has been reported.10 Preliminary studies evaluating the performance of screening have shown that ultrasound, although highly sensitive, lacks the specificity required for screening the general population.9 The interpretation of CA-125 using a single cutoff, developed for other purposes, has limited sensitivity in the context of early detection.3 Lowering the CA-125 cutoff may increase sensitivity but at the expense of reduced specificity. We advocate annual screening at the shortest interval that is logistically feasible to maximize sensitivity for early-stage disease, although more investigation of this topic is needed. The addition of a second-line test can improve the specificity and PPV of a screening strategy, but it cannot increase sensitivity. Given the substantial resources involved in a prospective screening trial,11 it is imperative to develop an optimized strategy that takes advantage of any additional information in serial CA-125 values. We have shown that the risk calculation from serial CA-125 levels is a significant improvement for preclinical detection compared with using a fixed cutoff for CA-125 measurement. In effect, this approach establishes an individual baseline CA-125 level for each woman and updates her risk of having the disease with each new CA-125 measurement. We have not discussed the exact cutoff levels for using the risk in clinical practice, because further research is required to define the optimal cut points. A computer is required to calculate the risk, but the same computer used to calculate the CA-125 value within the system performing the assay could also be used to calculate the risk. For a target PPV of 2% and a specificity of 98%, the risk calculation substantially increases the sensitivity for preclinical detection from 62% to 86%. It should be emphasized that the risk calculation alone for a PPV of 2% is not being advocated, but that the risk calculation should be followed by ultrasound for those at elevated risk. The PPV exceeded 25%,3 with a multimodal strategy incorporating a fixed cutoff for CA-125 measurement as a first-line test and ultrasound as a second-line test. Scanning a group identified by the more sensitive risk calculation test would likely increase the combined performance of a multimodal screening strategy. The high compliance rates with our previous screening study5 imply that following up women over time with serial CA-125 measurements will be acceptable to the screened population. Using the risk based on serial CA-125 values as a first-line test also has the advantage of being an objective test and is less expensive than transvaginal ultrasound. A recent model12 has indicated that a multimodal strategy incorporating CA-125 measurement as a first-line test and ultrasound as a second-line test is likely to be the most cost-effective strategy for ovarian cancer screening. In conclusion, we are sufficiently encouraged by the risk calculation’s performance for preclinical detection to begin at St Bartholomew’s Hospital a large, randomized, controlled trial incorporating the risk calculation as a first-line test followed by ultrasound for women at elevated risk (risk ≥ 1%). We believe that the improved operating characteristics for preclinical detection will translate into improved operating characteristics for early-stage detection.
DR. CANNISTRA: Through the algorithm that you’ve described, it seems that you are identifying individuals with normal CA-125s that start to rise, and distinguish them from those with elevated CA-125 levels that remain stable. DR. SKATES: You’re removing those women who get three monthly CA-125’s because they have levels that are fixed around 60 or 100, for example. So instead of having CA-125’s done every 3 months because they had this repeated elevated CA-125, you’re measuring those once a year because the risk is so low. On the converse, you’re increasing the sensitivity because you’re picking up women with CA-125’s that are low but start to rise. So both of those contribute to a low specificity and a high sensitivity or an increase in sensitivity at a fixed specificity. DR. CANNISTRA: It’s eliminating the group that has an elevated but stable CA-125 curve. DR. BAST: In some cases, there are physiologic mechanisms with CA-125 production. We know that adenomyoses or endometriosis in premenopausal women can elevate CA-125. Perhaps one woman in 50 will have fluctuations of CA-125 with normal menstrual cycles independent of known adenomyoses or endometriosis. Anything that inflames peritoneal surfaces, pleural surfaces, or pericardial surfaces can elevate CA-125, as can liver disease with or without ascites. DR. ROSE: It’s true that CA-125 is elevated in 80% of ovarian cancers but only 20% of stage I ovarian cancers [J Natl Cancer Inst 80:208–209, 1988]. I’m surprised that we’re still putting such tremendous effort into screening such a relatively poor marker. DR. BAST: A more accurate estimate of CA-125 elevation in stage I disease is 50% to 60%. DR. SKATES: Again, those are the women who are going to be detected in stage I anyway, so how does CA-125 behave in those who are detected clinically in stage III and IV when they would have been detected in early-stage disease? DR. BAST: Ovarian cancer is neither a common nor a rare disease, but it’s prevalence in the postmenopausal population is one in 2,500. This raises the issue of whether you should screen for a disease that’s neither common nor rare. Screening for more prevalent diseases such as breast, prostate, or lung cancer has obvious advantages from the perspective of public health. Screening for ovarian cancer is a strategic decision [Control Clin Trials 18:251–270, 1997]. DR. CANNISTRA: We have to get away from the idea that CA-125 elevation above the conventional upper limit of normal should be the sole trigger for screening. We should be asking ourselves how often early-stage cancers produce CA-125. Based upon today’s discussion, it could be as high as 80%. If you look at it that way and have a means of capturing those 80%, then you don’t have to rely solely on an elevated level. You just have to identify a pattern of change in a given patient that is predictive of an underlying ovarian cancer, even if the change is occurring within the normal range of CA-125 values. DR. DePRIEST: It’s important that we discuss how many women-years are saved and how many it costs by giving second-, third-, fourth-, fifth-line salvage chemotherapy to patients with ovarian cancer. We need to be allocating resources not only in palliative chemotherapy but also in prevention and early detection. DR. RUSTIN: Dr. Skates, can you give us a ballpark figure, using your ROC analysis, of how much this will cost? DR. SKATES: It depends on whether you’re doing it in the UK or the US, but it’s between $20,000 and $40,000 per year of life saved.
We thank Nina Einhorn and Kerstin Sjövall for the data that formed the basis for the original development of the risk calculation.
This research was supported by grants from the Women’s Cancer Program at Dana-Farber/Partners Cancer Program, the National Cancer Institute (CA-57693), and Centocor Inc (S.J.S.) and Research into Ovarian Cancer and The Gynaecology Cancer Research Fund (I.J.; registered charities, London, England).
1. Landis SH, Murray T, Bolden S, et al: Cancer statistics, 1998. CA Cancer J Clin 48:6–29, 1998[Abstract]
2. Einhorn N, Sjovall K, Knapp RC, et al: Prospective evaluation of serum CA 125 levels for early detection of ovarian cancer. Obstet Gynecol 80:14–18, 1992 3. Jacobs I, Davies AP, Bridges J, et al: Prevalence screening for ovarian cancer in postmenopausal women by CA 125 measurement and ultrasonography. Br Med J 306:1030–1034, 1993 4. Skates SJ, Xu FJ, Yu YH, et al: Toward an optimal algorithm for ovarian cancer screening with longitudinal tumor markers. Cancer 76:2004–2010, 1995[CrossRef][Medline]
5. Jacobs IJ, Skates S, Davies AP, et al: Risk of diagnosis of ovarian cancer after raised serum CA 125 concentration: A prospective cohort study. Br Med J 313:1355–1358, 1996 6. Gilks WR, Richardson S, Spiegelhalter DJ: Markov Chain Monte Carlo in Practice. London, England, Chapman and Hall, 1996, p 486 7. DeLong ER, DeLong DM, Clarke-Pearson DL: Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 44:837–845, 1988[CrossRef][Medline] 8. Campbell S, Royston P, Bhan V, et al: Novel screening strategies for early ovarian cancer by transabdominal ultrasonography. Br J Obstet Gynaecol 97:304–311, 1990[Medline] 9. DePriest PD, van Nagell JR Jr, Gallion HH, et al: Ovarian cancer screening in asymptomatic postmenopausal women. Gynecol Oncol 51:205–209, 1993[CrossRef][Medline] 10. Rosenthal A, Jacobs I: Ovarian cancer screening. Semin Oncol 25:315–325, 1998[Medline] 11. Kramer BS, Gohagan J, Prorok PC, et al: A National Cancer Institute sponsored screening trial for prostatic, lung, colorectal, and ovarian cancers. Cancer 71:589–593, 1993[Medline] 12. Urban N, Drescher C, Etzioni R, et al: Use of a stochastic simulation model to identify an efficient protocol for ovarian cancer screening. Control Clin Trials 18:251–270, 1997[CrossRef][Medline] Submitted February 11, 2003; accepted March 21, 2003. This article has been cited by other articles:
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2003 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
TOP
ABSTRACT
INTRODUCTION
