Serial endometrial thickness and risk of non‐endometrial hormone‐dependent cancers in postmenopausal women in UK Collaborative Trial of Ovarian Cancer Screening

ABSTRACT Objective Estrogen is a well‐established risk factor for various cancers. It causes endometrial proliferation, which is assessed routinely as endometrial thickness (ET) using transvaginal ultrasound (TVS). Only one previous study, restricted to endometrial and breast cancer, has considered ET and the risk of non‐endometrial cancer. The aim of this study was to explore the association between baseline and serial ET measurements and nine non‐endometrial hormone‐sensitive cancers, in postmenopausal women, using contemporary statistical methodology that attempts to minimize the biases typical of endogenous serial data. Methods This was a cohort study nested within the UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS). In the ultrasound arm of UKCTOCS, 50639 postmenopausal women, aged 50–74, underwent annual TVS examination, of whom 38 105 had a valid ET measurement, no prior hysterectomy and complete covariate data, and were included in this study. All women were followed up through linkage to national cancer registries. The effect of ET on the risk of six estrogen‐dependent cancers (breast, ovarian, colorectal, bladder, lung and pancreatic) was assessed using joint models for longitudinal biomarker and time‐to‐event data, and Cox models were used to assess the association between baseline ET measurement and these six cancers in addition to liver cancer, gastric cancer and non‐Hodgkin's lymphoma (NHL). All models were adjusted for current hormone‐replacement therapy (HRT) use, body mass index, age at last menstrual period, parity and oral contraceptive pill use. Results The 38 105 included women had a combined total of 267 567 (median, 8; interquartile range, 5–9) valid ET measurements. During a combined total of 407 838 (median, 10.9) years of follow‐up, 1398 breast, 351 endometrial, 381 lung, 495 colorectal, 222 ovarian, 94 pancreatic, 79 bladder, 62 gastric, 38 liver cancers and 52 NHLs were registered. Using joint models, a doubling of ET increased significantly the risk of breast (hazard ratio (HR), 1.21; 95% CI, 1.09–1.36; P = 0.001), ovarian (HR, 1.39; 95% CI, 1.06–1.82; P = 0.018) and lung (HR, 1.25; 95% CI, 1.02–1.54; P = 0.036) cancers. There were no statistically significant associations between ET and the remaining six cancers. Conclusion Postmenopausal women with high/increasing ET on TVS are at increased risk of breast, ovarian and lung cancer. It is important that clinicians are aware of these risks, as TVS is a common investigation. © 2019 The Authors. Ultrasound in Obstetrics & Gynecology published by John Wiley & Sons Ltd on behalf of the International Society of Ultrasound in Obstetrics and Gynecology.


INTRODUCTION
Cumulative estrogen exposure is linked to many types of cancers, particularly female cancers 1 . In breast 2,3 , endometrial 4,5 and, to a lesser extent, ovarian [6][7][8] cancer, this increase in risk associated with higher levels of estrogen is particularly well established. For lung cancer, there is a growing consensus that estrogens contribute to the increased rates seen in women compared with in men, and that smoking further augments the effect of estrogen [9][10][11][12] . For colorectal cancer, the evidence is unclear 13,14 , with an increased risk reported for estrone 15 and a decreased risk for estrogen 16,17 . The association between estrogen exposure and risk of gastric 18 , pancreatic 19 and bladder cancer 20 is less clear, but there is some evidence of a decreased risk associated with estrogen. For non-Hodgkin's lymphoma (NHL), definitive associations with estrogen have not been established 19,21 .
Estrogen receptors are expressed in many tissues, including the endometrium. Cumulative estrogen exposure, especially in postmenopausal women, leads to endometrial proliferation which could be easily assessed as endometrial thickness (ET) using transvaginal ultrasound (TVS). In postmenopausal women, the adipose tissue provides an additional source of estrogen, which in turn causes proliferation of the endometrium 22 . Whilst increased ET is well-established as an early detection marker for endometrial cancer 23 , ET could also be a potential risk marker for other hormone-sensitive cancers. A prospective study within the Prostate, Lung, Colorectal and Ovarian (PLCO) screening trial reported that an ET > 5 mm, compared with an ET < 3 mm, was associated with a 2-fold increase in the risk of breast cancer 24 . This study was limited by sample size (1272 women with 91 breast cancers) and the main analysis was restricted to only the baseline measurement. Furthermore, the statistical method used to incorporate the serial measurements was susceptible to bias.
Using a prospective cohort design in the ultrasound arm of the United Kingdom Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) 25 and a joint model of longitudinal and time-to-event data, the aim of this study was to explore whether ET measurement at baseline, or serial change over 11 years of screening, is associated with the risk of nine estrogen-dependent non-endometrial cancers in postmenopausal women.

UKCTOCS
(ISRCTN22488978; ClinicalTrials.gov NTC00058032) is a multicenter randomized controlled trial of 202 638 postmenopausal women aged 50-74 years, from England, Wales and Northern Ireland, randomized to annual screening either using TVS (ultrasound arm; n = 50 639) or serum CA125 (multimodal arm; n = 50 640) or to no screening (control arm; n = 101 359) 25 . The eligibility criteria and trial details are described elsewhere 25,26 . UKCTOCS was approved by the UK North West MREC (00/8/34) on 23 rd June 2000. Participants provided written consent for use of their data in secondary studies. This analysis used data from the ultrasound arm, in which women were offered annual TVS screening between 17 th April 2001 and 31 st December 2011. We limited the analysis to those who had not undergone hysterectomy prior to recruitment, had complete covariate data and had at least one valid ET measurement.

TVS data
Women underwent annual TVS scans, with repeat scans performed when abnormal adnexal morphology was noted 25 . Between 2001 and 2008, the scans were performed using a Kretz/Medison SA9900 machine (Medison, Seoul, South Korea), and from 2008 they were performed using a Medison Accuvix XQ machine (Medison). In addition to ovarian morphology, ET measurements were collected routinely and entered in a central web-based trial management system 26 . ET was recorded at the thickest anteroposterior diameter of the endometrium in a sagittal plane of the uterus. Calipers were placed perpendicular to the outer edge of the endometrium. If there was fluid in the endometrial cavity, the cavity fluid and the double endometrial stripe were measured, and the fluid diameter at the same point was subtracted 23 .

Cancer notification
All women were flagged with the NHS Digital for England and Wales and Northern Ireland Cancer Registry, and cancer notifications (site, morphology and date of diagnosis) included in this analysis were diagnosed by 31 st December 2014. Nine types of cancer that have a possible association with estrogen were identified using cancer registry ICD-10 codes: ovarian (C56, C57, C48), breast (C50), lung (C34), colorectal and anal (C18-C21), pancreatic (C25), bladder (C67), liver (C22), gastric (C16) and NHL (C85).

Covariate data
Data on hysterectomy, height, weight, current or past use of the oral contraceptive pill (OCP), parity, current use of hormone-replacement therapy (HRT) and age at last Endometrial thickness and cancer risk 269 menstrual period (LMP) were captured at recruitment 25 . Approximately 77% of women also completed a postal follow-up questionnaire 3-5 years post-randomization, which included additional data on alcohol intake, smoking and hysterectomy. Data on hysterectomy were also obtained from Hospital Episodes Statistics (HES England) by identifying all women with Q07 (abdominal excision of the uterus) and Q08 (vaginal excision of the uterus) records and from the ultrasound scan form, which also recorded current HRT use.

Statistical analysis
Of nearly 270 000 ET measurements, 10 values > 50 mm (maximum of 503 mm) were excluded arbitrarily from the analysis. The remaining ET measurements were transformed to the binary logarithmic scale; the calculated hazard ratios (HR) therefore reflect a doubling of ET.
The association between ET and each cancer type was explored using joint models, that modeled simultaneously the two related processes of serial measurements over time (longitudinal submodel) and time to clinical outcome (survival submodel). Modeling either process in isolation may introduce bias 27 due to the endogenous nature of the biomarker. Furthermore, joint models use best unbiased linear predictions of the underlying biomarker process in the survival submodel, smoothing out some of the potential measurement error. This can increase the signal and further reduce bias (Appendix S1).
For the longitudinal submodel, current use of HRT at baseline and at each ET measurement, age at LMP, body mass index (BMI), parity and OCP use were included as fixed effects, in addition to a subject-specific random intercept. The survival submodel included the same adjustment covariates, as well as the association parameter that links the two submodels. We specified the current value link, which is the fitted value including random effects from the longitudinal submodel. The baseline hazard was modeled using a Weibull distribution for all cancers except breast cancer, which used a cubic spline model with one interior knot. Further details of the joint model are given in Appendix S1.
Both longitudinal and survival submodels used age as the time metric, implying delayed entry for the survival analysis. Analysis time was defined as from age at first ET measurement to age at cancer diagnosis of interest. Censorship was age at 31 st December 2014 28 , or earlier if there was poststudy-entry hysterectomy or death. There was no censorship for diagnosis of any other cancer prior to diagnosis of the cancer of interest, nor were they treated as a competing risk. A lung cancer-specific sensitivity analysis included alcohol and smoking as covariates for both submodels, reducing the sample size by 33%. The joint model failed to converge for the three rarest cancers (NHL, gastric and liver). For comparison, the association between ET and each cancer based on only the baseline measurement was assessed using a Cox model for all nine cancers. The proportional hazards assumption was assessed using the Schoenfeld residuals.
In addition, a standalone longitudinal mixed model, restricted to ET measurements in women with none of the nine investigated cancers or endometrial cancer, was used to construct age-dependent reference curves at various centiles for normal ET measurements. All statistical analyses were performed using Stata 15.1 (StataCorp. LLC, College Station, TX, USA) and the joint models were fitted using the user-written command 'stjm' 29 . Hazard functions with 95% CI were estimated for ET values of 2.5 mm and 5 mm (the ratio of which being equivalent to the HR for the log 2 ET measurements) and presented graphically for the age range, 50-80 years. One-, 5and 10-year absolute risk estimates were calculated for selected cancers, ET measurements and ages.

RESULTS
Of 50 639 women randomized to the ultrasound arm in UKCTOCS, 9651 had undergone hysterectomy prior to recruitment and 237 underwent hysterectomy between recruitment and the first scan. A further 1893 women had no ET measurements and 753 (1.9%) women had missing data on at least one of the included covariates. The analysis therefore included 38 105 women who had at least one ET measurement and complete data on covariates ( Figure 1). Overall, there were 267 567 (36 169 baseline, 231 398 serial) ET measurements, with a median of 8 (interquartile range (IQR), 5-9; maximum, 19) per participant.
At recruitment, the median age of the women was 60.6 (IQR, 56.2-66.2) years, median BMI was 25.6 (IQR, 23.2-28.9) kg/m 2 , median age at LMP was 50.6 (IQR, 47.9-53.1) years and 17.4% of women were currently using HRT. Of the women, 88.0% were parous and 60.4% had used OCP. Table 1 shows the baseline and follow-up characteristics in the 36 168 women with an ET measurement at baseline, overall and according to baseline ET. Baseline ET was < 3 mm in 18 429 (51.0%) women, ≥ 3 to < 5 mm in 11 694 (32.3%) women and ≥ 5 mm in 6045 (16.7%) women. All factors were strongly associated (P < 0.0001) with ET, except for smoking (P = 0.254) and OCP use (P = 0.021). Age and alcohol use were associated negatively with ET, whereas age at LMP, BMI, parity, HRT use and OCP use were associated positively with ET. Figure 2 depicts the age-dependent reference curves for estimated ET in women who did not develop any of the described cancers (n = 35 058), based on a standalone mixed model. Median ET was estimated to be 2.91 mm in women at age 50 years, with the endometrium thinning slightly over time to 2.46 mm at age 80.
During a combined total of 407 838 (median, 10.9; IQR, 9.8-11.8) years of follow-up, 1398 breast, 381 lung, 495 colorectal, 222 ovarian, 94 pancreatic, 79 bladder, 62 gastric and 38 liver cancers and 52 NHLs were registered. For the joint models, there was a slight difference in sample size between the analyses for each cancer type, which was due to the cancer of interest occurring between recruitment and first ET measurement in some cases. The sample size in the six successfully fitted joint models varied from 38 078 (breast cancer) to 38 105 (ovarian, lung and pancreatic cancer) women (Table 2). In the Cox models, in which only the baseline ET measurement was used, the sample size was, at most, 1936 lower than that in the joint model due to missing baseline measurements. The frequency of cancer was also slightly reduced ( Table 2). In the joint models, there was statistical evidence at the 5% level of a positive association between ET and the risk of three of the six cancer types ( Table 2). The strongest association was with breast cancer (HR, 1.21; 95% CI, 1.09-1.36; P = 0.001). This model was the only instance in which a Weibull baseline hazard was not suitable (P = 0.0002). The 1-year absolute risk of breast cancer for a woman aged 50 with an ET of 5 mm was 165 in 100 000 vs 136 in 100 000 for an ET of 2.5 mm. For ovarian cancer, the HR was 1.39 (95% CI, 1.06-1.82; P = 0.018) and for lung cancer the HR was 1.25 (95% CI, 1.02-1.54; P = 0.036). A sensitivity analysis specifically for lung cancer, that additionally included alcohol intake and smoking as covariates, yielded a higher HR (1.35; 95% CI, 0.98-1.86), although it was not statistically significant (P = 0.066) due to fewer events (157/25 469). The association between ET and cancer risk was not significant for colorectal (HR, 1.15; P = 0.150), pancreatic (HR, 0.99; P = 0.947) or bladder (HR, 0.86; P = 0.542) cancer. One-, 5-and 10-year absolute risks for breast, ovarian and lung cancer at selected ages and ETs are shown in Table 3. For example, the 10-year risk in women at age 60 with an ET of 10 mm is 4.5% for breast cancer, 1.2% for ovarian cancer and 0.9% for lung cancer. Table 2 also shows the results of the Cox models for all nine cancers, using only the baseline ET measurement. A significant association at the 5% level was noted only for ovarian cancer (HR, 1.22; 95% CI, 1.04-1.42; P = 0.015). HR was < 1 for each of the three cancers not fitted by a joint model (gastric: HR, 0.79; liver: HR, 0.86 and NHL: HR, 0.81), but with P-values > 0.1. Figure 3 shows the estimated hazard functions for the six cancers from the joint models, calculated for ET values of 2.5 mm and 5 mm. Note that overlap of the 95% CI should not be used as an indicator of the significance of the HR values, as the respective hazard functions are not independent. The hazard functions for most of the cancers show increasing risk with age, although the Weibull distribution forces the functions to be monotonic. The exception is breast cancer, for which, after about age 65, there is a gradual decline in the risk associated with ET.

DISCUSSION
In this prospective cohort study of 38 105 postmenopausal women with a combined total of 267 567 ET measurements and over 400 000 years of follow-up, we report on the association between ET measured using TVS, as a functional surrogate for cumulative circulating estrogen, and the risk of nine non-endometrial cancers that may be affected by estrogen exposure. A doubling of ET was associated with an increased risk of breast (21%), ovarian (39%) and lung (25%) cancer. There was no association with pancreatic cancer, a non-significant increased risk of colorectal cancer and a non-significant decreased risk of bladder, liver and gastric cancers and NHL.
Key strengths of this study include its prospective multicenter design, the large cohort size, measurement of ET using a standard protocol and similar ultrasound machines across the centers and over time, completeness of cancer notification through linkage to national cancer registries and long-term follow up of the cohort. We report for the first time on the association between ET and eight (ovarian, lung, colorectal, pancreatic, bladder, gastric, liver and NHL) cancers. The only prior study investigating the association between ET and breast cancer 24 either used only the baseline ET measurement or incorporated serial Data are given as n or n (%). *Will differ at other scans. †Approximately 77% completed, 3.5 years post-randomization. BMI, body mass index; HRT, hormone-replacement therapy; OCP, oral contraceptive pill. measurements inappropriately. We used joint models, which we believe are an improvement over standalone survival models, including time-varying Cox models. An additional strength is the development of age-related normal range reference curves for ET measurements (Figure 2), which may be helpful in a clinical setting. There is inherent intrasubject variability in ET over time, which may have obscured the association with cancer risk, particularly in a baseline-only model. Other limitations include the lack of data on histological subtype, which may have indicated stronger associations for those deemed to be estrogen-dependent. Similarly, we did not have data on HRT type, nor potential confounders such as exercise and diet.
less day-to-day biological variability than does the level of circulating estrogen. Furthermore, ET likely reflects the cumulative estrogen exposure, and hence, is a better measure of cancer risk than is estradiol (E2) level. A doubling of ET resulted in a HR of 1.21 for breast cancer, compared with the HR of 2.00 reported by Felix et al. 24 , when comparing the risk in women with ET ≥ 5 mm vs < 3 mm. However, the latter report was based on 91 breast cancers compared with 1398 in this analysis, which therefore provides more precise estimates and much stronger statistical association. Breast cancer was the only analysis for which the estimated hazard function did not rise monotonically with age, and use of a cubic spline demonstrated a notable decline in the risk associated with increasing ET in women over 65 years of age. This may reflect the age range (50-70 years) of the NHS breast-screening program and associated overdiagnosis in women of these ages. In postmenopausal women, an increase in breast cancer risk has been reported previously in women with high levels of endogenous sex hormones, including E2 and estrone, with odds ratios of 2.00 and 2.19, respectively, when the highest quintiles are compared to the lowest 3 .
We found a strong association between a doubling of ET and ovarian cancer (HR, 1.39; P = 0.018). Trabert et al. 31 reported a modest association of higher levels of estrone and other estrogen metabolites with ovarian cancer, which was limited to non-serous histotypes. The link between ovarian cancer and exogenous hormones is now well established, with data from the Women's Health Initiative (WHI) suggesting that estrogen-plus-progestin therapy may increase ovarian cancer risk, most notably for serous histotypes (relative risk, 1.53) 6 .
The other statistically significant finding was for lung cancer, for which doubling of ET was associated with a 25% increased risk, which would not have been found using the baseline ET value alone (HR, 1.03; P = 0.65). This increased risk persisted even after inclusion of the  additional covariates of alcohol and smoking, which were available for only a subset of women. This is in keeping with the latest belief that estrogen is responsible for augmenting the risk of lung cancer in smokers 9,12 . One suggestion is that circulating estrogen plays a role in modifying ER-beta levels, which are known to inhibit tumor growth 10 .
Colorectal cancer was the only other common cancer in our study (n = 495), and despite an increase in risk (HR, 1.15), the association with a doubling of ET was not significant at the 5% level (P = 0.15). Although increasing circulating estrone has been reported to be associated with an elevated risk of colorectal cancer 15 , the majority of publications suggest that estrogen and its ER-beta receptor are protective for this cancer 16,17 . Data from the WHI study 32 and population-based registries in Norway 33 have demonstrated a decrease in colorectal cancer risk with short term/current use of estrogen-plus-progestin HRT.
Investigation of the five other assessed cancers was limited by their low incidence, although the association was estimated to be essentially null only for pancreatic 274 Burnell et al.
cancer. For bladder (either model), gastric and liver cancer and NHL, the HRs were around 0.85 or lower and it is conceivable that a significant association might be found with greater numbers. These indications of protective effects are broadly consistent with findings from the limited extant studies [18][19][20] .
One of the benefits of using ET as a proxy/surrogate marker for circulating estrogen is its relative ease of measurement. Serum E2 levels are usually measured using a radioimmunoassay. Mass spectrometry is used increasingly because it provides greater specificity and sensitivity, especially for the low E2 concentrations observed in postmenopausal women 34 . However, in addition to cost and run time, there are limited data available on the discrepancies between individual spectrometry-based assays. Findings from one study indicated that interfering compounds might cause E2 levels to be 10-times higher than the true value 35 .
In conclusion, in this study exploring the association between ET measured using TVS and the risk of nine non-endometrial, potentially hormone-sensitive, cancers in postmenopausal women, we found that high and/or increasing ET was associated significantly with an increased risk of breast, ovarian and lung cancer. This suggests that ET may merit inclusion in future risk prediction models for these cancers. While our findings need further validation, clinicians might wish to assess appropriately women with high or increasing ET measurements on TVS who do not have endometrial cancer.