Ovarian cancer risk prediction: a clinical epidemiology perspective

Abstract

Ovarian cancer is a rare and highly heterogeneous disease usually detected at late stages when outcomes are poor. Population-based screening approaches have not been successful at reducing ovarian cancer mortality, but preventive bilateral salpingo-oophorectomy is highly effective at preventing ovarian cancer in high-risk populations. Ovarian cancer risk prediction models may allow identification of populations at increased risk of ovarian cancer for preventive interventions or targeted early detection. We propose a life-course approach to ovarian cancer risk prediction based on the time at which a risk model should be applied and the risk factors that are available. The discriminative ability of ovarian cancer risk prediction models published so far is limited, with areas under the curve ranging from 0.58 to 0.65 for different combinations of risk factors and genetic susceptibility markers. Currently proposed absolute risk thresholds for preventive surgery are around 4% lifetime risk. The absolute risk predicted by ovarian cancer risk models ranges from 0.6% to 2.5% lifetime risk in the general population, highlighting the need to improve ovarian cancer risk prediction models and evaluating new preventive approaches that can be offered to individuals at lower risk.

Editor’s note: The opinions expressed in this article are those of the authors and do not necessarily reflect the views of the American Journal of Epidemiology.

Introduction

While ovarian cancer incidence and mortality have been decreasing over the last decade, ovarian cancer remains the most lethal gynecologic cancer, with an overall 5-year survival of 50%.¹ Most patients lack specific early clinical symptoms and are diagnosed at late stages when the disease has spread. Conversely, treatment of early-stage ovarian cancer is highly successful and has spurred interest in ovarian cancer screening.² However, two large, randomized screening trials using CA-125 and transvaginal ultrasound, the prostate, lung, colorectal, and ovarian screening trial (PLCO) and the UK collaborative trial of ovarian cancer screening (UKCTOCS), showed no survival benefit of ovarian cancer screening in the general population.^3,4

Several factors may explain these results. Ovarian cancer is rare, and very high specificity is needed to minimize false-positive results that may lead to unnecessary invasive procedures.⁵ Many ovarian cancers likely do not originate in the ovaries, but in other sites like the fallopian tubes, suggesting that ovarian visualization may not detect disease early.⁶ Limited understanding of ovarian carcinogenesis, particularly with respect to site of origin and time from early exposures to cancer development, makes it challenging to identify better approaches to prevention and early detection.

In contrast, preventative approaches are very successful in high-risk populations, specifically among individuals with BRCA pathogenic variants, where risk-reducing bilateral salpingo-oophorectomy (RRSO or BSO) can substantially reduce ovarian cancer risk and mortality.⁷ Ovarian cancer risk prediction models can play an important role in identifying individuals at increased risk of ovarian cancer who could be targeted for preventive interventions outside of the small group of individuals with rare pathogenic germline variants. Several risk prediction models have been developed to predict future ovarian cancer risk. Here, we describe specific applications, or use cases, for ovarian cancer risk prediction models, describe existing models in the context of absolute risk and clinical action thresholds, and highlight the need to develop models with improved risk stratification for different target populations.

Overview of ovarian cancer risk prediction

Cancer risk prediction models are typically developed by combining meaningful risk information into a risk score to predict future risk of cancer.⁸ While many cancer risk prediction models are primarily exploratory, some models are developed to identify subjects who would benefit from cancer prevention approaches, such as cancer screening and prophylactic surgery, or to identify individuals who should undergo diagnostic evaluation or treatment.

The intended application for a cancer risk prediction model has important implications for model development, with respect to the risk information included, the required discriminatory ability, the absolute risk prediction, and the time in the life course when cancer risk is estimated. Absolute risk prediction can be conceptualized as a data integration process combining multiple data sources for the relative risks for the risk factors, age-specific disease rates and rates of competing events and population distribution of risk factors. Prior to clinical or public health applications, a model predicting absolute risk needs to be evaluated for calibration and discrimination in independent prospective cohorts. We refer to other publications that provide more details about model building, model calibration, and model discrimination.^8-11

The baseline risk in the target population has important implications for development and application of a risk model. The lower the baseline risk, the better the performance of a risk model needs to be to identify individuals with high absolute risk levels. Annual ovarian cancer incidence in the United States is about 10 cases per 100 000 women (ie a risk of 0.01% per year for the general female population) and it is expected that 19 710 new cases will be diagnosed in 2023.¹²  Figure 1 shows a strong increase of ovarian cancer incidence by age, with incidence around 30/100 000 by age 65 and around 40/100 000 by age 80. In the general population, the lifetime risk of ovarian cancer is estimated around 1.1%–1.5%, while the lifetime risk ranges from 20% to 40% for individuals with variants in BRCA1/2, MLH1, or MSH2.⁷

Figure 1.

Life-course perspective of ovarian cancer risk prediction. Ovarian cancer incidence was plotted by age based on SEER22 from 1975–2020 using SEER*explorer (https://seer.cancer.gov/statistics-network/explorer/). Relevant exposure windows and applications for risk models are plotted over the life course.

Life-course approach to ovarian cancer risk prediction and specific use cases

We approach ovarian cancer risk prediction with a life-course perspective, considering two directions around the time point when a risk model is applied (Figure 1): (1) Going backwards, the risk model uses risk information that has accumulated during the life-course until that time point. (2) Going forward, the model will predict ovarian cancer risk for a specified time window in the future.

The timepoint during the life course when each risk factor can be assessed varies widely for ovarian cancer. Genetic susceptibility (high-penetrance variants like BRCA1, BRCA2, or common susceptibility loci combined into polygenic risk scores [PRS]) can be ascertained from birth but require genetic testing from the entire population evaluated for risk. Family history of cancer can be assessed at birth to some extent but may change during the life course when first-degree relatives are diagnosed with cancer. In contrast, most other ovarian cancer risk factors can only be assessed later in adulthood, such as hormonal and reproductive exposures, which limits their use in risk models required at young age.

Given the increased risk of ovarian cancer associated with germline variants in high penetrance genes BRCA1, BRCA2, MLH1, and MSH2, as well as moderate penetrance genes BRIP1, RAD51C, RAD51D, PALB2, and ATM, ovarian cancer risk assessment based on genetics plays an important role for preventive applications. However, most ovarian cancers in the population are not related to pathogenic germline variants. Ovarian cancer incidence peaks after age 60, when non-genetic risk information has accumulated, allowing to develop models with better performance for this age group. The time window of future risk that needs to be predicted by a cancer risk prediction model can range from a few years for screening and early detection approaches up to lifetime risk for preventive approaches like surgery.

Risk models for preventive interventions

An important use case for ovarian cancer risk prediction models is to identify target populations for preventive interventions. Possible preventive interventions range from chemoprevention, such as use of aspirin or hormonal prevention using oral contraceptives that have shown ovarian cancer risk reduction,^13-17 to surgical interventions, including tubal ligation, RRSO/BSO, and bilateral salpingectomy. Depending on the intervention, risk assessment may be needed early in the life course when preventive agents act early in cancer development. Some risk factors like hormonal and reproductive factors are not available or are incomplete in younger women. The risk of ovarian cancer must be high to justify risk-reducing surgery, as it is the case for carriers of pathogenic BRCA variants.^18-20 For other preventive interventions, a lower risk may be sufficient if the associated harms are low. For potential chemoprevention approaches, such as aspirin or hormones, use over a long time period is associated with potential health risks, including bleeding for long-term aspirin use,²¹ or risk of venous thromboembolism or other cancers for some hormone preparations,^22,23 respectively. Other potential harms related to clinical application of risk models need to be considered, including the possibility of communicating false negative and false positive results which can lead to false reassurance of low cancer risk, or unnecessary interventions and follow up, respectively.

Risk models for selecting ovarian cancer screening participants

Some ovarian cancer screening trials conducted in high-risk populations and the general population have shown a stage shift in cancers detected in the screening arm; however, ovarian cancer screening in the general population has not reduced ovarian cancer mortality in two large screening trials.^3,4,24,25 Risk prediction can be used to identify a subset of the general population at increased risk of ovarian cancer that could benefit from ovarian cancer screening. The advantage of targeting screening to a higher risk population is better absolute risk prediction and a better population-level risk to benefit ratio, since fewer low-risk individuals are subjected to unnecessary interventions.⁵

The goal of a risk model to identify screening participants is to exclude a substantial proportion of the general population with low to average risk from the screening population, while ensuring that as many cancers as possible will occur in the population for which screening is recommended. For example, this approach is used to select participants for CT-based lung cancer screening, relying on risk factors like smoking history.²⁶ Since most ovarian cancers occur after the age of 50, risk prediction models to identify a target population for ovarian cancer screening can be applied when most hormonal and reproductive risk information is available and can be combined with genetic information.

Risk models for early detection

A risk model for early detection and treatment decisions would require prediction of a high absolute risk of cancer over a very short time so that diagnostic procedures can detect a pre-clinical or early-stage cancer. It is currently not clear what the time window is between possible detectability of ovarian cancer and clinical manifestation; a risk model for early detection would likely need to be applied at an age where ovarian cancer incidence is highest, eg after age 55 years. Since preclinical lesions such as serous tubal intraepithelial carcinomas are microscopic, imaging has limited lead time for early detection. Interventions such as cytologic sampling of the fallopian tubes or detection of molecular signatures in vaginal secretions have been evaluated, but the optimal approach to diagnostic evaluation is not defined.^27-29 The shorter the time window, the more frequent risk assessment would need to be conducted to identify individuals who need diagnostic evaluation. Current models based on epidemiologic risk factors and genetics do not achieve sufficient risk stratification for such an application, and risk is typically estimated for a longer period. Further research is needed to evaluate whether biomarkers can be added to risk models to allow short-term risk prediction, for example tests like CA-125, or new approaches like cell-free DNA, which have shown some promise in detecting ovarian cancer as part of multi-cancer detection efforts. While promising, initial liquid biopsy studies were conducted in symptomatic patients and future studies are needed to evaluate preclinical detection prospectively.^30,31

Overview of published ovarian cancer risk prediction models

Table 1 summarizes key characteristics and metrics of ovarian cancer risk prediction models that were developed using risk factor data and/or genetic information and that included an independent study dataset (either subsets of the population not included in model building or separate studies) to validate the risk model applied to the general population, or to individuals at increased genetic risk. Three models were built using various epidemiologic risk factors,^32-34 three using a PRS,^35-37 and two using a combination of both.^38,39 Core epidemiologic risk factors included in all models with this information were reproductive history variables and oral contraceptive use. All models published after 2000 further included menopausal hormone use. A wide range of risk factors was included across different models, ranging from 4³³ to 12³⁸ variables. While many variables were similar across models, the risk estimates were derived from different sources, leading to variation in effect sizes. Notably, salpingectomy, which confers strong protection against ovarian cancer and is now standard of care when performing permanent sterilization or hysterectomy with ovarian retention, even for benign indications,⁴⁰ is not incorporated in published risk models due to the low prevalence prior to 2015.

Table 1.

Summary of published ovarian cancer risk prediction models

Model	Epidemiological risk factors	SNPs	Population for model building	Population for model validation	Calibration (E/O, 95% CI)	Discrimination (AUC, 95% CI)
Rosner Epidemiology 2005	Model 1: duration of premenopause, duration of menopause, age at first birth minus age at menarche, birth index, duration of OC use, tubal ligation Model 2: duration of premenopause, duration of menopause, parity, duration of oc use, tubal ligation	None	Theoretical model to estimate ovarian cell divisions	78 504 women from NHS; 382 ovarian cancers; 106 618 from NHS2; 90 ovarian cancers	1.00	0.60 0.57,0.62
Pfeiffer PLoS Med 2013	Family history of breast or ovarian cancer, duration of MHT use, parity, OC use	None	PLCO, NIH-AARP 151 165 and 58 282 women, respectively. 597 284 cases, respectively	Independent data from NHS 56 638 women, 377 cases	1.08 0.97,1.19	0.59 0.56,0.63
Li Br J Cancer 2015	Full model: Age at menarche, parity, # full term pregnancies, age at first live birth, duration of breast-feeding, miscarriages, OC use, duration of OC use, IUD, menopausal status, age at menopause, HRT use, duration of HRT use, hysterectomy, unilateral ovariectomy, BMI, prevalent diabetes Selected model: Parity, # full term pregnancies, OC use, duration OC use, menopausal status, age at menopause, HRT use, duration HRT use, unilateral ovariectomy, BMI	None	EPIC 202 206 women, 791 ovarian cancers	EPIC separate data left out from model building	0.90 0.81,1.01	0.64 0.58,0.70 for full model 0.64 0.57,0.70 for selected model
Clyde AJE 2016	Age at diagnosis/interview, age at menarche, OC use, pregnancy history, breastfeeding, tubal ligation, endometriosis, family history of breast/ovarian cancer, BMI, aspirin use, menopausal status, hysterectomy, MHT use	17	OCAC Training set 7586 controls, 4662 ovarian cancers	OCAC validation set 1926 controls, 1131 cases	Calibration plot	Model without SNPs: 0.65 model with SNPs: 0.66
Kuchenbaecker JNCI 2017	None	17	Synthetic model for general PRS, internal data for BRCA-specific SNPs	CIMBA 15 252 BRCA1 (breast cancer = 7797, ovarian cancer = 2462) and 8211 BRCA2 (breast cancer = 4330, ovarian cancer = 631)		0.58 (0.56,0.60) for BRCA1 carriers; 0.63 (0.60,0.67) for BRCA2 carriers
Yang J Med Genet 2018	None	15	Synthetic model approach	UKCTOCS 750 EOC cases, 1428 controls		0.58 0.55 , 0.60 for the overall PRS; 0.60 0.57–0.63 for the serous PRS
Lee J Med Genet 2022	Parity, OC use, MHT use, endometriosis, tubal ligation, BMI, height	15	Synthetic model approach	374 cases, 1587 controls from UKCTOCS	1.05	0.61
Dareng Eur J Hum Genet 2022	None	22 up to 17 000	OCAC 23 564 invasive non-mucinous EOC ovarian cancers, 40 138 controls	UK Biobank, European ancestry: 657 cases/198 101 controls African ancestry: 368 cases/704 controls Asian ancestry: 2841 cases/4828 controls BRCA1/2 carriers from CIMBA BRCA1: 2053 cases/16 862 controls BRCA2: 717 cases/11 620 controls		European 0.59 CIMBA BRCA1 carriers 0.56 CIMBA BRCA2 carriers 0.61 Models based on OCAC and CIMBA data (UK biobank): European 0.60 African 0.58 Asian 0.53

The number of SNPs in risk models has increased from 15 to over 40 with the continued discovery of genome-wide significant loci through large Genome-Wide Association Study (GWAS) efforts.⁴¹ One study expanded PRS to include SNPs beyond genome-wide significant loci, with the largest set including 17 000 SNPs.³⁵ Most models were developed for and validated in the general population,^32-35,37-39 while one was developed for populations with BRCA mutations.³⁶ Due to the limited availability of risk factor and genetic data for non-European ancestry populations, all models were primarily developed for European ancestry populations. One study conducted validation of a PRS-based risk model in small subsets of the population with African and/or Asian ancestry.³⁵

Most risk factor-based models reported the expected/observed number of cases to assess calibration, which ranged from 0.90 to 1.08, indicating adequate to excellent calibration in independent validation studies. As the standard reported measure of discrimination, the area under the curve (AUC) estimates ranged from 0.59 to 0.64 for risk-factor based models, from 0.58 to 0.60 for PRS-based models alone,^35,37 and 0.61 to 0.65 for models combining risk factor data with a PRS.^38,39 The PRS model developed for BRCA populations showed AUCs of 0.58 and 0.63 for BRCA1 carriers and BRCA2 carriers, respectively.³⁶ The expansion of PRS to include larger numbers of SNPs beyond the genome-wide significant loci had minimal impact on discrimination.³⁵ Importantly, PRS-based models can achieve discrimination and absolute risk prediction similar to risk-factor based models, with the advantage that the genetic information is available at birth. Combining SNPs and risk factors led to a slight increase of AUCs compared to models with either SNPs or risk factors.³⁸

Ovarian cancer risk prediction for clinical and public health applications

The usefulness of a risk model is determined by the ability of the model to identify a subset of the population with an absolute risk that would justify specific clinical management or public health interventions. For example, a risk-based approach has been developed for cervical cancer prevention and is the foundation of current screening and management guidelines.^42,43 Absolute risk of cervical precancer is estimated for combinations of risk factors and test results and evaluated in the context of clinical action thresholds that determine screening intervals, diagnostic evaluation, and treatment.

The AUC, an estimate of the discrimination between cases and controls over the whole range of risk estimates, is the most widely used metric to assess performance of ovarian cancer risk prediction models. The AUC summarizes the complexity of risk model performance in a single value that can be compared across models. However, the AUC alone is inadequate to evaluate the usefulness of a risk model.⁵ Most importantly, the AUC does not consider disease prevalence. For a rare disease such as ovarian cancer, a higher AUC is required to achieve a certain absolute risk compared to a common disease. Figure 2 highlights the importance of baseline risk for risk prediction.⁵ The four graphs include ROC curves with AUCs ranging from 0.6 to 0.9. Plots are shown for different baseline risk, ranging from one-year ovarian cancer risk prediction in the general population (baseline risk 0.01%) to the lifetime risk of ovarian cancer in women with pathogenic BRCA variants (baseline risk 20%). A risk model with an AUC of 0.6 has limited discrimination ability and can predict an absolute risk of 2.5% only at the highest percentiles of the risk score for lifetime risk in the general population. AUCs of 0.8 or 0.9 are needed to predict a 10-year risk of 2.5%. The situation is very different in a population with high baseline risk, (eg pathogenic BRCA variant carriers) where modest AUCs can lead to high absolute risk prediction. These projections serve as a “reality check” for the potential clinical utility of a risk model with a certain discriminative ability, or AUC.

Figure 2.

Relationship of AUC and absolute risk of ovarian cancer. Receiver operating characteristic (ROC) curves, plotting the sensitivity and specificity of a risk model over the whole range of thresholds, are superimposed on contours of absolute risk for different prior risk estimates. The four graphs include ROC curves with AUCs ranging from 0.6 to 0.9, with the diagonal line (AUC of 0.5 indicating no discrimination) added as a reference. The plots are created for different baseline risk, ranging from 1-year ovarian cancer risk prediction in the general population (baseline risk 0.01%) to the lifetime risk of ovarian cancer in women with pathogenic BRCA variants (baseline risk 20%).

Clinical risk thresholds

Currently, there is a limited set of effective clinical or public health recommendations available for the prevention of ovarian cancer. Most interventions have been evaluated in the context of high genetic risk. RRSO is a surgical procedure that is offered for BRCA carriers.⁷ By removing the likely site of origin of ovarian cancer, RRSO is highly effective at preventing ovarian cancer, with hazard ratios around 0.21.⁴⁴ This translates into a substantial overall mortality reduction (HR 0.32) and a stronger reduction in ovarian cancer mortality (HR 0.06).⁴⁵ However, oophorectomy and the associated premature surgical menopause, has well-documented harms, which include risk of osteoporosis, cardiovascular diseases, and adverse effects on mood, sexual health and cognition, in addition to detrimental effects on fertility and surgical risks. Hormone Replacement Therapy (HRT) use can compensate for some of the effects of premature menopause but itself may increase risk of breast cancer.⁴⁶ Therefore, benefits and harms for RRSO need to be carefully evaluated when considering expanding the indications for this efficacious preventive intervention. Several recent benefits-harms analyses and cost-effectiveness analyses suggest that a lifetime risk as low as 4% could warrant RRSO, particularly after childbearing has been completed.²⁰

Bilateral salpingectomy has been evaluated based on the fact that most high-grade serous ovarian carcinomas originate in the fallopian tubes. Removal of fallopian tubes before malignant transformation and migration of malignant cell clones to the ovaries has high potential to prevent ovarian cancer. This procedure is ovary-preserving, avoiding premature menopause. However, not all ovarian cancers originate in the fallopian tubes, and even for those types that do, salpingectomy performed after migration of transformed cells cannot prevent cancer. There is currently no clear prospective evidence about the optimal time window and risk reduction after bilateral salpingectomy; several studies are underway.⁴⁷ Another approach under evaluation is bilateral salpingectomy followed by delayed bilateral oophorectomy. This approach combines benefits of both approaches, but the efficacy is still under evaluation and the approach is not recommended outside of clinical trials.⁴⁸

Screening with transvaginal ultrasound and CA-125 testing offered to the entire population has not provided a mortality benefit, but it should be evaluated whether these modalities are successful in combination with a risk model that can identify those most likely to benefit. Lastly, there is currently not sufficient data on benefits and harms of chemopreventive approaches for ovarian cancer in the general population and at what ovarian cancer risk level they could be considered. In BRCA populations, chemoprevention using combined oral contraceptives can increase breast cancer risk short term while reducing long-term endometrial and ovarian cancer risk.⁴⁹

Performance of risk models in a clinical context

When determining the clinical utility of risk models, one must consider the level of absolute risk predicted in context with possible clinically actionable thresholds. Figure 3 displays absolute risk estimates reported by ovarian cancer risk models developed for the general population and for individuals with BRCA mutations. In the general population, risk models can predict a 1.8%–3.8% absolute risk over several decades for the highest percentiles of risk scores and a risk around 0.7% in the lowest percentiles of risk.^33,37,39

Figure 3.

Absolute risk prediction in the context of clinical action thresholds. Absolute risk estimates generated by different risk models (identified by first author on the plot) are plotted on a log scale of absolute risk. The time periods of risk prediction and the components of risk models are indicated below the x-axis. When comparing risk estimates across models, we must consider that models were developed for different populations and predict risk for different time intervals, so each model should be evaluated in the context of absolute risk thresholds on its own merits. The general population risk is estimated at 1.3%. For risk models, top and bottom 5% of the risk score is plotted. The model by Pfeiffer et al. does not indicate percentiles but describes high-risk and low-risk risk factor combinations. Absolute risk thresholds for surveillance and prophylactic surgery are indicated at 2% and 4%, respectively.

Among carriers of a pathogenic BRCA variant, the baseline risk of 20% is far above the threshold for recommending preventive surgery and both the risk in the highest and lowest percentiles suggest recommending prophylactic surgery.³⁶ However, there is a well-documented higher risk of ovarian cancer in BRCA1 carriers of 39%–58% compared to a 13%–29% risk in BRCA2 carriers.⁷ In addition, BRCA2 carriers present later with an average age of diagnosis similar to the general population. Thus, risk reducing BSO may be delayed in this population. Refining risk prediction models even in the high-risk population may aid in counseling patients regarding timing for risk reducing surgery. In addition, with broader indications for and improved access to genetic testing, previously underrepresented minorities and patients without classic family histories of cancer will present for risk evaluation. Updated models in diverse patient populations will be key for counseling these patients. Lastly, the ovarian cancer risks associated with moderate penetrance genes such as BRIP1, RAD51C, RAD51D, ATM, and PALB2 are lower ranging from 2% to15% with current expert guidelines ranging from “manage patients based on family history” to recommending BSO starting at 45–50. In this group, risk modeling could alter management of patients.⁷

A critical aspect of ovarian cancer risk prediction is how risk is distributed in the population and how cancers are distributed across risk groups. Figure 4 shows the distribution of ovarian cancer risk predicted by two risk models^37,39 (SNPs only and SNPs+risk factors) and by BRCA testing for the general population. Both general population risk models can predict an absolute risk of greater than 2% for a sizable subset of the population (20% for SNPs only and 34% for SNPs and risk factors). However, only a very small subset of the population in the risk model with SNPs and risk factors have a risk above the proposed prophylactic surgery threshold of 4%. Communicating risk estimates for health outcomes is challenging; shared decision-making between patients and providers is particularly critical when predicted risk levels do not have clear clinical guidance.⁵⁰ In contrast, population-based BRCA testing results in two very distinct risk groups: a small subset of the population (around 0.4%) with a BRCA pathogenic variant that has a lifetime risk of 20% or greater and most of the population without a pathogenic variant that has a risk around 1.3%.

Figure 4.

Distribution of ovarian cancer risk and ovarian cancer cases in the population. The first bar graph (A) shows the distribution of ovarian cancer risk predicted by two exemplary risk models (SNPs only and SNPs+risk factors, both validated in the UKCTOCS population) and by BRCA testing (positive vs. negative) in the general population. The second bar graph (B) shows the distribution of cases according to the risk groups predicted by the models and population BRCA testing.

In the general population, between 29% and 52% of ovarian cancer cases occur in individuals with a risk of 2% or higher, depending on the model used. In contrast, only 14% of cases are in the group of individuals with a pathogenic BRCA variant. These metrics are important to evaluate benefits and harms for different preventive approaches, with the theoretical benefit determined by the number of cases in a risk stratum and the potential harms related to the preventive action applied to the population in that stratum. For example, the SNP-based risk model predicts that 29% of all cases will occur in 20% of the population, while the combined risk model can predict that 52% of all cases will occur in 34% of the population. In contrast, BRCA testing predicts that 14% of all cases will occur in 0.4% of the population, sparing the majority of the population from potential harms of an intervention, but missing 86% of cases.

Summary

There is a great need for ovarian cancer risk prediction approaches that can improve prevention and reduce mortality of ovarian cancer. Risk prediction and prevention efforts work well for those at high genetic risk. Risk models for the general population show some promise but need improvement to allow predicting ovarian cancer risk in a range where preventive measures or clinical interventions could be considered for a subset of the population. Several limitations need to be addressed to achieve clinically relevant ovarian cancer risk prediction.

First, it is important to evaluate risk models in the context of their specific use case and to select models that are most likely to achieve required performance. For preventive approaches, which require long-term risk prediction early in life, epidemiologic risk models that rely on a completed hormonal and reproductive history cannot be used but need to rely on genetic information or family history. For early detection approaches, extensive risk information is available, but accurate short-term risk prediction for clinical decision-making requires additional risk markers like blood-based biomarkers or imaging.

Second, better understanding the natural history and multi-step carcinogenesis of ovarian is a critical foundation of risk prediction efforts. Recent advances in molecular biology and molecular pathology of ovarian cancer have shown that many relevant precursors may originate outside of the ovaries, which has important consequences for preventive interventions, like tubal ligation of salpingectomy, or imaging-based early detection approaches. Better understanding of ovarian carcinogenesis can also help develop disease models that can inform benefits-harms and cost-effectiveness analyses. Identifying new ovarian cancer risk factor associations, and including recently established risk factors, such as salpingectomy, inflammation-related exposures, and medication use, may improve model performance. Given the profound heterogeneity of ovarian cancer, subtype specific risk models should be developed and evaluated.¹⁶ For example, endometriosis is strongly associated with the rare endometroid and clear cell subtypes, while smoking is only associated with the rare mucinous subtype.¹⁶

Third, expanding the range of preventive interventions, like chemoprevention with hormones or aspirin, and clinical interventions that can be adapted to different ovarian cancer risk levels is a critical task. For each approach, the benefits and harms of possible interventions need to be assessed. So far, these analyses have led to proposing a lifetime 4% absolute risk threshold for prophylactic BSO. A lower, to be defined, threshold for surveillance or other preventive measures will make it easier for ovarian cancer risk models to achieve useful risk prediction for a larger population. A promising approach is to consider a sequential use of risk prediction models and other assessments to maximize benefits and minimize harms. For example, a risk model can be used to identify a population in which additional biomarkers can be measured for early detection. Similarly, non-specific symptoms that are not useful when evaluated in the general population may have predictive or diagnostic value in the background of elevated risk.

Developing better approaches to ovarian cancer prevention and early detection for the general population that will have a meaningful impact on ovarian cancer mortality is possible, but requires a concerted, multi-disciplinary effort. Given the strong disparities of ovarian cancer survival, it is critical to develop risk models and preventive approaches for diverse populations. Current understanding of risk factors, biology, and genetics is largely based on studies in populations of European ancestry. While the general principles laid out here also apply to other rare and highly fatal cancers, the natural history, risk model performance, and available options for clinical interventions are highly cancer specific and the same model performance may have very different implications at different cancer sites.

Funding

NCI Intramural Research Program: Z01 CP010124.

Conflict of interest

The authors report no conflicts of interest.

Data availability

The article uses published or publicly available data.