A multi-state survival model to identify risk factors for lethal ovarian cancer

Our analyses included 211,420 unique participants at risk for developing ovarian cancer across three 12-year risk periods and 1,730 participants who developed ovarian cancer. There were 865 deaths from ovarian cancer during follow-up; of these, 660 died from ovarian cancer in the same risk period as their diagnosis and were included in mortality analyses. Participants who developed ovarian cancer and either survived their risk period of diagnosis or died during their risk period of diagnosis tended to have a longer duration of estrogen hormone therapy use at baseline than non-cases (mean 1.0–1.4 years vs. 0.8 years in NHS; Table 1 ). Incident cases were more likely than non-cases to be nulliparous (7.3–7.9% vs. 5.4% in NHS; 31.3–42.4% vs. 22.6% in NHSII), and have a family history of breast cancer (7.0–7.4% vs. 6.0% in NHS; 6.4–8.1% vs. 5.9% in NHSII). Use of oral contraceptives for 10 or more years (2.6–3.1% vs. 4.5% in NHS; 7.4–10.5% vs. 10.9% in NHSII) and self-identifying as non-White (4.1–5.6% vs. 6.1% in NHS; 4.0–4.5% vs. 7.3% in NHSII) were less common among incident cases than non-cases. The most common tumor type among participants who died from ovarian cancer was type 2 (70.1% in NHS and 69.0% in NHSII). We observed several patterns of risk factor associations with lethal ovarian cancer reflecting underlying associations with incidence and mortality ( Table 2 , Figure 2 ). The most common pattern was an amplified association between the risk factor and lethal ovarian cancer due to associations with incidence and mortality in the same direction. For example, longer reproductive span (per 6 years) was associated with higher incidence (HR 1.50, 95% CI 1.41–1.59) and mortality (HR 1.42, 95% CI 1.26–1.61), resulting in a RR of 1.97 (95% CI 1.76–2.20) for lethal ovarian cancer. Other risk factors in this pattern were age at menarche (lethal ovarian cancer RR 1.13 per 1 year older, 95% CI 1.06–1.20), menopause duration (RR 1.46 per 6 years, 95% CI 1.34–1.59), nulliparity (RR 1.62, 95% CI 1.23–2.13), medium-term oral contraceptive use (5–9 vs. 0 years, RR 0.70, 95% CI 0.52–0.95), and depression (RR 1.34, 95% CI 1.01–1.79). Another pattern was relationships driven by a risk factor association with incidence. For example, despite having no association with ovarian cancer mortality (HR 1.05, 95% CI 0.70–1.58), ≥10 versus 0 years of oral contraceptive use was associated with a 35% decreased risk of lethal ovarian cancer (RR 0.65, 95% CI 0.43–0.97) due to its strong association with incidence (HR 0.60, 95% CI 0.48–0.76). The other risk factor in this pattern was duration of estrogen hormone therapy use (per 2 years, RR 1.12, 95% CI 1.08–1.17). Conversely, the association of pack-years of smoking with lethal ovarian cancer was driven by the mortality relationship (≥20 pack-years versus never smoking, incidence HR 1.06, 95% CI 0.93–1.20; mortality HR 1.25, 95% CI 1.01–1.54; lethal RR 1.25, 95% CI 1.02–1.53). The final pattern observed was an overall null association of the risk factor with lethal ovarian cancer due to opposing associations with incidence and mortality. For example, family history of breast cancer was associated with increased risk of ovarian cancer (HR 1.19, 95% CI 1.04–1.37) and decreased risk of death (HR 0.82, 95% CI 0.65–1.05), resulting in no association with lethal ovarian cancer (RR 1.03, 95% CI 0.81–1.30). Family history of ovarian cancer and tubal ligation followed a similar pattern. In general, associations were consistent across risk periods, with some exceptions ( Supplementary Table S2 ). Several associations increased in magnitude over the risk periods, such as the association of menopause duration with mortality ( P -heterogeneity=0.049) and the association of adult change in BMI with mortality (≥4 vs. −2 to 2 kg/m 2 , P -heterogeneity=0.003). Conversely, factors with waning associations with incidence over time included parity, 5–9 years vs. never use of oral contraceptives, endometriosis, and family history of ovarian cancer ( P -heterogeneity=0.005–0.047). For example, additional births after the first birth were inversely associated with incidence in periods 1 and 2 (HRs 0.87 and 0.94) but not period 3 (HR 0.99). We conducted analyses by type 1 and type 2 tumors ( Figure 3 , Supplementary Tables S3 – S5 ). Due to limited numbers of type 1 tumors (N=131–230 across risk periods) and deaths (N=30–35 across risk periods), those estimates generally had wide CIs. Thus, apparent differences by tumor type should be interpreted cautiously. Duration of combined estrogen and progestin hormone therapy use and a history of long-term smoking had suggestively inverse associations with lethal type 1 cancer, but positive associations with lethal type 2 cancer ( P -heterogeneity=0.02–0.12). In general, though, the associations were largely similar across histotypes with respect to hazards of lethal ovarian cancer. In analyses by menopausal status ( Figure 4 , Supplementary Tables S5 – S7 ), results for premenopausal women should be interpreted cautiously as most cases occurred after menopause (N=139–224 [24–68] premenopausal and N=180–377 [87–167] postmenopausal cases [deaths] across risk periods). We observed an inverse association of breastfeeding duration with lethal ovarian cancer among premenopausal women (per 12 months, RR 0.62, 95% CI 0.45–0.85), but not postmenopausal women (RR 0.93, 95% CI 0.79–1.09, P -heterogeneity=0.02). Family history of breast or ovarian cancer was strongly associated with increased risk of lethal ovarian cancer among premenopausal women (RR 2.41, 95% CI 1.30–4.47), but not postmenopausal women (RR 0.93, 95% CI 0.70–1.23, P -heterogeneity=0.01). Nulliparity tended to have a stronger positive association with lethal ovarian cancer risk in premenopausal than postmenopausal women ( P -heterogeneity=0.07).

NHS was established in 1976 among 121,700 female U.S. registered nurses, ages 30 to 55, and NHSII was established in 1989 among 116,429 female U.S. registered nurses ages 25 to 42. Both cohorts are observational cohort studies; thus, no randomization or blinding procedures were used. Participants have been asked to complete biennial questionnaires about their medical and reproductive history and health behaviors. We examined the association of baseline (pre-diagnosis) risk factors with ovarian cancer incidence and mortality. Given the 28–40-year follow-up period in the cohorts, we divided the follow-up period into three 12-year risk periods, allowing exposures to be updated once (NHSII) or twice (NHS) ( Figure 1 ). Period 1 was 1980–1992 (NHS), period 2 was 1992–2004 (NHS) / 1993–2005 (NHSII), and period 3 was 2004–2016 (NHS) / 2005–2017 (NHSII). During the overall study period, 94% of NHS participants and 96% of NHSII participants remained under active follow-up, defined as the number of person-years in the cohort when participants are censored after their last questionnaire response divided by the total number of person years in the cohort (i.e., participants censored only upon death). The study was conducted in accordance with the ethical guidelines of the Declaration of Helsinki. The study protocol was approved by the institutional review boards of the Brigham and Women’s Hospital and Harvard T.H. Chan School of Public Health, and those of participating registries as required (Protocol numbers: 1999P003389, 1999P011114, 2022P002840). Completion of the self-administered questionnaire was considered implied consent. Participants self-reported ovarian cancer diagnoses on questionnaires. To confirm diagnoses, we obtained medical records, including pathology reports, or linked to the relevant cancer registry. We included cases corroborated a second time by the participant or next-of-kin as well as cases identified via cause of death in the National Death Index (NDI). Pathology reports were reviewed by a gynecologic pathologist to abstract tumor histology and grade, enabling classification of tumor histotype as type 1 (low grade serous, mucinous, endometrioid, clear cell, low grade mixed, borderline transitional/Brenner) or type 2 (high grade serous/poorly differentiated, high grade mixed, invasive transitional/Brenner, carcinosarcoma). 14 Our analysis included incident, confirmed ovarian cancer cases identified through May 2016 (NHS) or May 2017 (NHSII). Cause of death was confirmed via NDI linkage. We used a prospective approach in which an exposure was examined in relation to ovarian cancer incidence and mortality if data were available from a questionnaire administered at or before the beginning of the risk period ( Supplementary Table S1 ), except for self-identified race in NHS, which was asked in 1992. We evaluated the impact of age on lethal ovarian cancer risk by modeling components of age (which sum to age at baseline) that have been associated with risk of ovarian cancer incidence and/or mortality 13 , 15 , 16 : age at menarche, reproductive span (defined as age at menopause [if postmenopausal] or age at the beginning of the risk period [if premenopausal] minus age at menarche), and duration of menopause (defined among postmenopausal women as age at the beginning of the risk period minus age at menopause; zero among premenopausal women). Age at menopause was determined from self-reported age when natural periods ended or age at bilateral oophorectomy. For women who underwent a premenopausal hysterectomy, we used median imputed age at menopause, stratified by current smoking status and ever use of hormone therapy. We examined reproductive and hormonal variables that are established or putative risk factors for ovarian cancer (or specific ovarian cancer subtypes), including parity, breastfeeding duration, oral contraceptive use duration, tubal ligation, hysterectomy, and use of estrogen, combined estrogen and progestin, or other hormone therapy. 13 , 17 – 19 Other established, putative, or emerging risk factors evaluated included endometriosis confirmed by laparoscopy (available in NHSII only), cigarette smoking pack-years, depression, low-dose aspirin use, first degree family history of breast cancer, first degree family history of ovarian cancer, body mass index (BMI) at age 10, change in BMI from age 10 to 18, and change in BMI from age 18 to age at menopause (if postmenopausal) or at the beginning of the risk period (if premenopausal). 13 , 20 – 28 BMI at age 10 was derived based on the nine-figure Stunkard somatotype pictogram 29 , which was included on the 1988 (NHS) and 1989 (NHSII) questionnaires following previously used methods. 26 Depression was defined as a score ≥8 on the 10-item Center for Epidemiologic Studies Depression Scale 30 (NHS only), positive response to a screener question for DSM-III-R major depressive episode 31 (NHSII only), or report of depression diagnosis or regular anti-depressant use in the past 2 years. For incidence and mortality analyses, inclusion criteria at the beginning of each risk period were no prior cancer diagnosis, having at least one intact ovary, non-missing data on components of age (i.e., age at menarche and age at menopause [if postmenopausal]), and being under active follow-up (i.e., last questionnaire was returned after baseline for that period). Mortality analyses included incident cases that developed during the risk period (see participant flow charts in Supplementary Figures S1 – S3 ). We used multivariable Fine and Gray subdistribution hazard models to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for incident ovarian cancer diagnosis (or ovarian cancer-specific mortality) in relation to the set of evaluable risk factors available at the beginning of each risk period. 32 We analyzed variables that had generally linear associations as continuous variables, per standard deviation increment higher (e.g., per SD older age at menarche or longer duration of menopause). To evaluate parity, we included an indicator for nulliparity (yes vs. no) and a continuous variable representing each additional birth after the first birth. The remaining factors were analyzed as categorical variables. To maintain adequate sample size in each risk period, we included missing indicator variables for breastfeeding, the BMI variables, depression, and low-dose aspirin, which were not asked until later in follow-up in NHS and NHSII. For incidence analyses, person-time was calculated from the date of return of the risk period baseline questionnaire until date of ovarian cancer diagnosis, development of a competing risk (i.e., death, diagnosis with another cancer other than non-melanoma skin cancer, or bilateral oophorectomy), or the risk period cutoff date, whichever was earliest. For mortality analyses, person-time was calculated from the date of ovarian cancer diagnosis until date of death from ovarian cancer, development of a competing risk (i.e., death from a cause other than ovarian cancer), or the risk period cutoff date, whichever was earliest. Data from NHS and NHSII were pooled for periods 2 and 3. To obtain estimates for endometriosis, models were run using NHSII only. The multi-state model for lethal cancer has been described previously. 11 Briefly, if C I F t , x , z _ = cumulative incidence of lethal ovarian cancer over t years among disease-free women at time 0 (e.g., 1980 in risk period 1) with main exposure x , and other risk factors z _ , one must first develop ovarian cancer at time t 1 years ( 0 < t 1 < t ), and then die of ovarian cancer at some time over the next t – t 1 years, or C I F t , x , z _ = ∫ t 1 = 0 t I 1 t 1 , x , z _ C I F 2 t − t 1 , x , z _ d t 1 where I 1 t 1 , x , z _ = incidence of ovarian cancer at time t 1 years, adjusted for competing risks, C I F 2 t − t 1 , x , z _ = cumulative incidence of ovarian cancer mortality among patients with ovarian cancer during t – t 1 years after diagnosis, adjusted for competing risks. The relative risk (RR) for lethal ovarian cancer among cancer-free women at baseline with x+1 versus x is equivalent to R R t = C I F t , x + 1 , z _ / C I F t , x , z _ . The log RR(t) is approximated by w 1 β 1 + w 2 β 2 where β 1 and β 2 are the beta coefficients for incidence and mortality, respectively, associated with a one-unit difference in x and w 1 and w 2 are weights that depend on the baseline hazard and survival function for incidence and mortality post-diagnosis, respectively, over the risk period. To generate a single estimate for risk factors examined in more than one risk period, we calculated the inverse variance weighted average of the beta coefficients. To calculate 95% CIs, we assessed the standard error of the weighted average beta coefficient by taking the square root of the inverse of the sum of the inverse variances across periods. To quantify variability in incidence, mortality, and lethal beta coefficients across periods as well as heterogeneity between weighted average lethal relative risk estimates in analyses by tumor type and menopausal status, we used random effects meta-analysis to calculate the Q-statistic for heterogeneity. We used Wald tests to quantify heterogeneity in the weighted average incidence and mortality beta coefficients for each factor. In analyses by tumor subtype, incidence of the other type of ovarian tumor or tumors with unknown type was considered a competing risk. Due to the limited number of type 1 tumors, we modified the coding of several covariates, including oral contraceptive use (combined the top two categories), family history of breast or ovarian cancer (combined into one variable), and BMI change from age 18 to menopause (combined the lowest and missing categories), to allow the models to converge. For consistency, we made the same modifications for the type 2 tumor models. Analyses among premenopausal and postmenopausal women were based on menopausal status at the beginning of the risk period. Due to the lower number of cases and deaths among premenopausal women, we modified the coding of family history of breast or ovarian cancer and BMI change from age 18 to menopause and modeled the cumulative incidence of lethal ovarian cancer over 11 years to ensure model convergence. These modifications were also applied to postmenopausal models. All analyses were conducted using SAS Software version 9.4 (RRID:SCR_008567). Further information about the procedures to obtain and access data from the Nurses’ Health Studies is described at https://www.nurseshealthstudy.org/researchers (contact email: [email protected] ).

In this study, we evaluated pre-diagnosis risk factors for lethal ovarian cancer over 12 years in two large, prospective studies. Our analyses identified risk factors that had strong associations with lethal ovarian cancer overall due to the amplifying impact of having similar direction of association for incidence and mortality. This applied to many reproductive factors and depression as well as family history of breast or ovarian cancer for premenopausal women. Further, some lethal ovarian cancer risk factors we identified reflected exposures associated strongly with incidence that had little or no association with mortality. These included higher risk with longer estrogen hormone therapy duration and lower risk with long-term oral contraceptive use. Importantly, we identified a risk factor for lethal ovarian cancer that is not associated with incidence, but only mortality (pack-years of smoking), and clarified risk factors that had no impact on risk of lethal ovarian cancer overall due to opposing associations with incidence and mortality (e.g., family history of breast or ovarian cancer and tubal ligation). In general, the risk factor associations we observed with incidence and mortality were consistent with the prior literature. 10 , 13 , 15 , 16 , 20 – 22 , 24 , 26 , 27 , 33 – 38 One notable difference was for pre-diagnosis hormone therapy use, which has been associated with improved ovarian cancer survival in several, 34 , 39 , 40 but not all, 41 prospective studies. Here, pre-diagnosis use of estrogen only and combined estrogen and progestin were not associated with ovarian cancer survival, including in analyses limited to postmenopausal women and for type 1 or type 2 tumors. However, since the goal of our analyses was to assess baseline exposures with 12-year risk of lethal ovarian cancer, we did not update exposure status after baseline, which could explain differences in results compared to other studies with updated exposures. This study had several limitations. We identified incident ovarian cancer cases via self-reports on questionnaires and linkage of deceased participants to cancer registries, potentially leading to outcome misclassification for participants who did not report their ovarian cancer and survived throughout the study period. However, given the large sample size and the rarity of ovarian cancer, the percentage of participants classified as non-cases who were true ovarian cancer cases is expected to be very low, which would result in minimal impact to the effect estimates. Further, the long follow-up allowed for comprehensive assessment of death outcomes for cases. We examined baseline risk factors with risk of lethal ovarian cancer, thus we did not capture changes in exposures in the 12 year risk period after baseline or post-diagnosis, which may have different associations with survival. 39 , 42 To partially address this, we split the follow-up into three 12-year periods, allowing us to update exposures and evaluate premenopausal and postmenopausal exposures. However, the relatively short periods limited the number of evaluable cases and deaths and led to low statistical power for less common exposures and in some secondary analyses. Further, in secondary analyses, we examined risk of lethal type 1 or type 2 ovarian cancer, rather than individual histotypes, due to limited case numbers and, thus, insufficient statistical power for rarer histotypes. While there are some differing risk factor associations across histotypes within the type 1 category, risk factor associations with incidence of low-grade serous, endometrioid, and clear cell tumors (the majority of type 1 tumors) are generally similar (uncentered correlation similarity metric of 0.71–0.75). 13 Combining histotypes with different risk factor profiles would be expected to attenuate associations. Future studies with larger sample sizes should examine associations with individual histotypes. To preserve the sample size available for analyses, we used the missing indicator method. A simulation study found that the missing indicator method generally yields unbiased results when the missingness is independent of the outcome (as in a prospective study setting) and is not extreme (i.e. <25%). arXiv:2111.00138 In our study, missingness was ≤16% for all variables except the early life BMI variables which were missing in 20%-32% of NHS participants ( Table 1 ); thus, results for early life BMI, especially among postmenopausal women (primarily NHS participants), should be interpreted cautiously. Our analyses assumed associations with pre-diagnosis exposures were the same across risk periods. However, we observed several factors with waning (e.g., parity, oral contraceptive use duration) or strengthening associations (e.g., menopause duration) with incidence or mortality over the three time periods examined, which could be due to chance or suggest an impact of recency of exposure. For example, we observed a protective association of oral contraceptive use for 5–9 years versus never use with ovarian cancer risk in the first and second, but not the third, risk period. More research is needed to determine if this could be due to use of newer oral contraceptive formulations that require longer duration of use to have a protective effect 43 or a waning effect with time since last use. 44 Nonetheless, future analyses may benefit from incorporating time since last exposure, particularly for premenopausal exposures. Prior studies of ovarian cancer mortality risk factors have reported varying strength of associations by time since diagnosis 45 , 46 ; modifying the multi-state model to incorporate variability in associations of factors with short-term versus long-term survival might improve the accuracy of lethal RR estimates. Our mortality analyses were conducted among participants with similar education and relatively similar access to healthcare. Thus, the associations we observed between pre-diagnosis factors and mortality may not be generalizable to a population with different health care access. Due to limited data on debulking status and treatment available in these cohorts, we were not able to evaluate the extent to which treatment mediated the associations between the pre-diagnosis factors and risk of mortality. Further, the prediction of risk of lethal ovarian cancer may differ as novel treatments that substantially improve prognosis become available. In conclusion, we used a multi-state model to identify exposures that predict 12-year risk of developing lethal ovarian cancer. Model adaptations could expand its applicability, including with case-control data (with appropriate weighting of controls) and incorporating more than three states (e.g., adding a precursor lesion state). Overall, this approach could be used as part of a new framework for ovarian cancer risk prediction, with the long-term goal of developing a prediction model for lethal ovarian cancer. However, it is important to acknowledge that an ovarian cancer diagnosis is a life-altering event that can have significant acute and long-term impacts on health and well-being whether or not it causes death. Thus, the lethal ovarian cancer model should be considered complementary to, rather than a replacement of, ovarian cancer incidence models. Risk factors that are associated with incidence, but not associated with lethal cancer risk, must still be considered when analyzing the benefits versus risks of initiating an intervention. The lethal ovarian cancer model has particular value in highlighting factors with amplified importance due to associations with both incidence and mortality as well as factors whose importance could be overlooked due to associations with mortality, but not incidence. A high risk of lethal ovarian cancer may alter the risk-benefit ratio of certain prophylactic or interception interventions—especially those with inherent risks, such as surgery or medications.

Given the lack of effective screening, about half of ovarian cancer diagnoses occur after distant metastasis when prognosis is poor with a 5-year survival of about 30%. 1 The majority of ovarian cancer diagnoses occur among women at average genetic risk, 2 – 4 a population in whom application of ovarian cancer preventive strategies that have inherent risks is difficult to justify. For example, bilateral salpingo-oophorectomy is highly effective for nearly eliminating risk of ovarian cancer mortality, 5 but, due to the serious risks of abrupt surgical menopause in premenopausal women (e.g., reduced bone density and increased risk of cardiovascular disease, metabolic disorders, neurologic deficits, mortality 6 , 7 ), it is only recommended for women with specific risk factors (e.g., BRCA1/2 mutations). Similarly, several studies have observed lower ovarian cancer risk among women using daily low-dose aspirin. 8 , 9 Aspirin use currently is not recommended for ovarian cancer risk reduction; however, if the level of evidence in support of aspirin use reaches the threshold for guideline consideration in the future, it would not be appropriate to recommend for all women given its risks (e.g., bleeding). One potential approach to reduce ovarian cancer mortality is to better identify high-risk populations among women at average genetic risk who could be offered preventive strategies that are not appropriate to apply to the general population due to their inherent risks. Previous research observed distinct risk factor profiles by tumor aggressiveness, 10 defined as time between diagnosis and death, suggesting pre-diagnosis factors may be useful to identify high-risk populations. Expanding this line of research to identify pre-diagnosis risk factors for developing lethal ovarian cancer could be an impactful, complementary approach to reduce ovarian cancer mortality. We previously developed a multi-state survival model and applied it to determine pre-diagnosis factors related to risk of lethal breast cancer among cancer-free postmenopausal women. 11 To evaluate established risk factors for ovarian cancer incidence or mortality in relation to 12-year risk of developing lethal ovarian cancer, we applied this model to cancer-free women enrolled in two prospective cohort studies, Nurses’ Health Study (NHS) and NHSII. Given known heterogeneity in risk factor associations by histologic subtype and menopausal status, 12 , 13 we additionally conducted analyses for type 1 and type 2 tumors and by menopausal status.