| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
ORIGINAL RESEARCH |
From the Departments of Gynecology and Obstetrics and Biostatistics and Epidemiology, Cleveland Clinic Foundation, Cleveland, Ohio.
Address reprint requests to: Anne M. Weber, MD, Department of Gynecology and Obstetrics, Cleveland Clinic Foundation, 9500 Euclid Avenue, A81, Cleveland, OH 44195, E-mail: webera{at}cesmtp.ccf.org
 |
Abstract |
|---|
Top
 Abstract Materials and MethodsResultsDiscussionReferences |
|---|
Methods: We conducted a case-control study of women with abnormal perimenopausal or postmenopausal bleeding who had endometrial samplings; cases had endometrial cancer or complex hyperplasia and controls had benign endometrial histologies. Multivariate logistic regression models identified factors associated with risks of endometrial neoplasia. The predictive abilities of our models and a published model were assessed using the area under receiver operating characteristic (ROC) curves, for which an area of 1.0 indicated perfect positive predictive ability and an area of 0.5 was expected by chance.
Results: There were 57 cases of endometrial hyperplasia or cancer and 137 controls. Parity was related inversely (odds ratio [OR] 0.70; 95% confidence interval [CI] 0.56, 0.88; P = .002) and weight directly (OR 1.02 per kg; 95% CI 1.01, 1.04; P = .018) to the risk of endometrial neoplasia. Age (OR 1.04 per year; 95% CI 1.00, 1.08; P = .06) and diabetes (OR 3.50; 95% CI 0.99, 12.33; P = .052) were significant marginally. The area under the ROC curve for our model was 0.75, indicating moderate predictive ability; the area under the ROC curve for the published model was lower at 0.66.
Conclusion: Current clinical predictive models based on case-control studies do not have sufficient predictive ability to determine if women with abnormal perimenopausal or postmenopausal bleeding should have diagnostic testing.
Endometrial cancer is the most common gynecologic malignancy in the United States, with about 36,000 new cases diagnosed each year.1 Atypical endometrial hyperplasia is a precursor of endometrial cancer and can progress to invasive disease in up to 29% of cases.2 The clinical significance of complex hyperplasia without atypia still is debated, but endometrial cancer was reported in up to 9% of patients over time.2,3 Abnormal bleeding is the most common symptom of endometrial neoplasia, and the standard clinical practice is diagnostic testing of all women with abnormal perimenopausal or postmenopausal bleeding.
The yield of such testing is relatively low with prevalence of endometrial cancer in women with postmenopausal bleeding of about 10%.4 If an individual’s risk of endometrial neoplasia can be predicted on the basis of known demographic and clinical factors, it might be possible to avoid discomfort, risk, and expense of testing in women at low risk, while targeting high-risk women.
Clinical prediction rules or models use clinical factors to make predictions about clinical decisions such as who requires diagnostic testing.5 Factors significantly associated with endometrial neoplasia are widely known, and in theory, those factors could be used in a clinical model predictive of the risk of endometrial neoplasia in women with abnormal bleeding. Feldman et al6 developed a model based on four factors (age greater than 70 years, diabetes, nulliparity, and postmenopausal status). They estimated an 87% risk of endometrial neoplasia for women with postmenopausal bleeding with all four factors and a risk of 2.6% in women with none.
Models based on case-control data might not work well as a predictor in other populations. Before clinical prediction rule is adopted, it should be validated independently in different populations. The objective of our study was to develop a model based on risk factors for endometrial neoplasia in our population and then assess its predictive ability, and that of the published model, using our data set.
|
Materials and Methods |
|---|
|
Top
AbstractMaterials and MethodsResultsDiscussionReferences |
|---|
We included patients with cancer (defined as any invasive malignancy of the endometrium), adenocarcinoma in situ, and complex endometrial hyperplasia with and without atypia. For ease of reference, cases were described as patients having endometrial neoplasia. Controls were defined by benign histologies on endometrial samplings (including simple hyperplasia). Specimens with endometrial tissue insufficient for definite pathologic diagnoses were included as controls because many clinicians interpreted that result as indicative of endometrial atrophy. Diagnoses were assigned by standard pathologic criteria. Charts were abstracted for demographic and clinical data. Postmenopausal bleeding was defined as bleeding after at least 1 year of amenorrhea or abnormal bleeding on menopausal hormone therapy. All other women were considered perimenopausal. Complete information was available on 57 (38.8%) of 147 cases. With that number of cases, a 2:1 ratio of controls to cases was chosen to provide 90% power to detect odds ratios (ORs) of 2.5 or higher, given all but extremely low or high control-group exposure rates.7 To ensure enough controls, identification numbers for 140 patients were selected from potential controls using a computer-generated, random-selection algorithm. Complete data were available on 137 patients, or 5% of all eligible controls.
The association between various factors and likelihood of endometrial neoplasia was first assessed by univariate analysis with all available data. Factors with significance levels up to 0.20 (at least ten patients with that factor) and minimal missing data were identified as candidates for multivariate logistic regression modeling. For highly correlated factors, the factor with fewer missing data was chosen for the model; thus, parity was included, but gravidity was excluded, hypertension included but antihypertensive medication excluded, and weight included but body mass index (BMI) excluded. We included menopausal status as a potential factor, although it did not meet our criterion of P 
.20 because it was associated previously with endometrial cancer.
Multivariate models used various selection techniques, including the use of bootstrap samples for forward selection of factors.8 Bootstrapping involves creating a large number of artificial data sets by random sampling with replacement (ie, the same patient can be selected more than once) from the original data set, doing analyses, and pooling the results. In this case, 1000 data sets were generated, and for each data set, a forward selection algorithm was used to find the most significant factor. The factor chosen most frequently among the 1000 analyses was added to the model, and further factors were sequentially added using the same process. This approach simulates identifying factors that might accurately predict outcome in populations similar to ours. The original data then are fit to a statistical model containing those factors, and the results are compared with other models, the theoretical and mathematical details of which were validated by Efron.8
The fit of the bootstrap and other competing models was assessed and compared using Hosmer-Lemeshow goodness-of-fit tests,9 comparisons of the area under the receiver operating characteristic (ROC) curves,10,11 and, in the case of nested models, likelihood ratio tests. A nonsignificant result on the Hosmer-Lemeshow goodness-of-fit test indicates a good fit of the model. The area under the ROC curve is a summary measure of the model’s predictive value and is based on sensitivity and specificity. Sensitivity and specificity are preferred, particularly in case-control studies, because they are not dependent upon the prevalence rate of disease in the sample, in contrast to positive and negative predictive values. The area under the ROC curve can range from zero to one, with values close to one indicating better predictability and values close to 0.5 indicating what would be expected purely by chance with no predictive ability. P values for the area under ROC curves refer to whether the area is significantly greater than 0.5, whereas P values for comparing two areas under ROC curves refer to whether the difference between the two curves is zero.
Descriptive data on continuous variables are presented as mean ± standard deviation (SD). Results of logistic regression analyses are presented as ORs with 95% confidence intervals (CIs). All statistical tests were two-tailed, and P values .05 were statistically significant.
|
Results |
|---|
|
Top
AbstractMaterials and MethodsResultsDiscussionReferences |
|---|
shows demographic and clinical characteristics and results of univariate analyses for those factors selected as possible candidates for multivariate models.
|
); model 2, developed with bootstrapping, included only age, parity, diabetes, and weight (Table 3). Both models fit the data well according to the Hosmer-Lemeshow statistic and had moderate predictive ability (area under the ROC curve = 0.75 for model 1 and 0.74 for model 2; P < .001 for both, indicating a significant difference from area under the ROC curve of 0.5). The areas under the ROC curves of the two models were not significantly different (P = .514). The P-value for the likelihood ratio test, P = .424, indicated that menopausal status, hypertension, and hormone use in model 1 contributed little additional predictive information beyond model 2.
|
|
). These factors included age greater than 70, nulliparity, menopausal status, and diabetes; only diabetes was significant in our sample. The Hosmer-Lemeshow test showed an adequate fit of the model (P = .560). The area under the ROC curve for model 3 was 0.66, which was significantly better than 0.5 (P = .001) but not significantly different from model 2 (P = .167).
|
|
Discussion |
|---|
|
Top
AbstractMaterials and MethodsResultsDiscussionReferences |
|---|
Feldman et al6 proposed a scoring system for predicting the risk of endometrial neoplasia based on case-control data. Significant factors in their study were similar to ours, but the predictive ability of their factors on our patients was relatively poor, most likely because case-control studies are drawn from contrived samples in which all cases but only a portion of controls are studied. In those samples, the prevalence of disease will not be the same as in a sample drawn from an entire cohort of individuals.
Endometrial neoplasia encompasses a spectrum of disease with vastly different biologic behavior. It is not known if the different types of endometrial disease are associated with different demographic and clinical risk factors; our sample of cases was not large enough to address this question. Analyzing different endometrial diseases in one group might explain the model’s limited ability to identify endometrial neoplasia. The association between endometrial hyperplasia and cancer is not understood fully; it is still debated whether hyperplasia actually progresses to invasive cancer (analagous to cervical dysplasia and cancer) or whether hyperplasia is just a marker of underlying increased risk for endometrial cancer. We chose to include complex hyperplasia with and without atypia in our model because both are associated with an increased risk of endometrial cancer. The model might have had stronger predictive ability if cases were restricted to endometrial cancer and complex atypical hyperplasia.
Although a prospective cohort study would be preferable to a case-control design, a prospective study is not practical because of the relatively low prevalence of endometrial cancer in postmenopausal women with bleeding (10%)4 and the extremely low incidence in asymptomatic women (estimated at 1.7 cases per 1000 women per year).12 Data were collected retrospectively, so some information was missing, and it was not possible to include all cases in the sample. Data on many potentially important factors were excluded because of small numbers of patients with those factors. Case-control studies are particularly vulnerable to selection bias, so cases and controls might not represent a sample of the population of interest. To address this in our study, we excluded referred cases and collected data only on cases arising from the population of women with abnormal perimenopausal and postmenopausal bleeding at our center. To avoid selection bias in identifying controls, we randomly selected them from the same population as the cases.
Endometrial biopsy essentially has replaced D&C as the more cost-effective test in initial evaluation of women with postmenopausal bleeding.13 Vaginal ultrasound was suggested as a less invasive and less costly alternative to endometrial biopsy, with a high negative predictive value when endometrial thickness is less than 4 mm.4 However, ultrasound has low positive predictive value and further testing to exclude endometrial neoplasia must be done in more than half the women with postmenopausal bleeding.14 It does not seem likely that a clinical predictive model can be developed that would estimate such a low risk of endometrial neoplasia so accurately in women with abnormal perimenopausal or postmenopausal bleeding that diagnostic testing would be unnecessary. A model could be used to triage women to different diagnostic tests. Women at low predicted risk could undergo initial screening with ultrasound, using its high negative predictive value and its reduced discomfort and lower cost. To avoid ultrasound’s low positive predictive value and frequent need for further testing, women at high risk could undergo immediate endometrial biopsies for definitive tissue diagnosis.
Further research is needed to develop and test a clinical predictive model that assigns women with abnormal perimenopausal or postmenopausal bleeding to different diagnostic protocols based on their estimated risk of endometrial neoplasia. It might be possible to extend such a model to screening asymptomatic women at high risk for endometrial neoplasia. In one study that screened asymptomatic women with diabetes, hypertension, or both, preinvasive endometrial neoplasia was found in 6.3%, 1.7%, and 1.3%, respectively.15 It might be premature to recommend screening in groups thought to be at high risk. More study is required to differentiate women whose endometrial disease represents a potentially life-threatening condition from those in whom it would remain indolent, and diagnosis and treatment, particularly surgery, would incur greater risk of morbidity and mortality. Endometrial cancer and its precursors, like many other neoplasms, frequently exist in asymptomatic women and remain undiagnosed during life. In one autopsy study of women who died from other causes, previously unsuspected endometrial cancer was detected at a five times higher rate (26 per 10,000) than a comparable incidence rate during life (five per 10,000).16
|
Footnotes |
|---|
Received June 23, 1998. Received in revised form September 28, 1998. Accepted October 15, 1998.
|
References |
|---|
|
Top
AbstractMaterials and MethodsResultsDiscussionReferences |
|---|
2. Kurman RJ, Kaminski PF, Norris HJ. The behavior of endometrial hyperplasia: A long-term study of "untreated" hyperplasia in 170 patients. Cancer 1985;56:403–12.[Medline]
3. Huang SJ, Amparo EG, Fu YS. Endometrial hyperplasia: Histologic classification and behavior. Surg Pathol 1988;1:215–29.
4. Karlsson B, Granberg S, Wikland M, Ylostalo P, Torvid K, Marsal K, et al. Transvaginal ultrasonography of the endometrium in women with postmenopausal bleeding—A Nordic multicenter study. Am J Obstet Gynecol 1995;172:1488–94.[Medline]
5. Wasson JH, Sox HC, Neff RK, Goldman L. Clinical prediction rules: Applications and methodological standards. N Engl J Med 1985; 313:793–9.[Abstract]
6. Feldman S, Cook EF, Harlow BL, Berkowitz RS. Predicting endometrial cancer among older women who present with abnormal vaginal bleeding. Gynecol Oncol 1995;56:376–81.[Medline]
7. Schlesselman JJ. Case-control studies: Design, conduct, analysis. New York: Oxford University Press, 1982.
8. Efron B. Bootstrap methods: Another look at the jackknife. Ann Stat 1979;7:1–26.
9. Hosmer DW, Lemeshow S. Applied logistic regression. New York: John Wiley and Sons, 1989.
10. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143: 29–36.[Abstract]
11. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics 1988;44:837–45.[Medline]
12. Koss LG, Schreiber K, Oberlander SG, Moussouris HF, Lesser M. Detection of endometrial carcinoma and hyperplasia in asymptomatic women. Obstet Gynecol 1984;64:1–11.
13. Feldman S, Berkowitz RS, Tosteson AN. Cost-effectiveness of strategies to evaluate postmenopausal bleeding. Obstet Gynecol 1993;81:968–75.
14. Weber AM, Belinson JL, Bradley LD, Piedmonte MR. Vaginal ultrasonography versus endometrial biopsy in women with postmenopausal bleeding. Am J Obstet Gynecol 1997;177:924–9.[Medline]
15. Gronroos M, Salmi TA, Vuento MH, Jalava EA, Tyrkko JE, Maatela JI, et al. Mass screening for endometrial cancer directed in risk groups of patients with diabetes and patients with hypertension. Cancer 1993;71:1279–82.[Medline]
16. Horwitz RI, Feinstein AR, Horwitz SM, Robboy SJ. Necropsy diagnosis of endometrial cancer and detection-bias in case/control studies. Lancet 1981;2:66–8.[Medline]
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
