Estimating sojourn time and sensitivity of screening for ovarian cancer using a Bayesian framework

Abstract

Background

Ovarian cancer is among the leading causes of gynecologic cancer-related death. Past ovarian cancer screening trials using combination of cancer antigen 125 testing and transvaginal ultrasound failed to yield statistically significant mortality reduction. Estimates of ovarian cancer sojourn time—that is, the period from when the cancer is first screen detectable until clinical detection—may inform future screening programs.

Methods

We modeled ovarian cancer progression as a continuous time Markov chain and estimated screening modality–specific sojourn time and sensitivity using a Bayesian approach. Model inputs were derived from the screening arms (multimodal and ultrasound) of the UK Collaborative Trial of Ovarian Cancer Screening and the Prostate, Lung, Colorectal and Ovarian cancer screening trials. We assessed the quality of our estimates by using the posterior predictive P value. We derived histology-specific sojourn times by adjusting the overall sojourn time based on the corresponding histology-specific survival from the Surveillance, Epidemiology, and End Results Program.

Results

The overall ovarian cancer sojourn time was 2.1 years (posterior predictive P value = .469) in the Prostate, Lung, Colorectal and Ovarian studies, with 65.7% screening sensitivity. The sojourn time was 2.0 years (posterior predictive P value = .532) in the United Kingdom Collaborative Trial of Ovarian Cancer Screening’s multimodal screening arm and 2.4 years (posterior predictive P value = .640) in the ultrasound screening arm, with sensitivities of 93.2% and 64.5%, respectively. Stage-specific screening sensitivities in the Prostate, Lung, Colorectal and Ovarian studies were 39.1% and 82.9% for early-stage and advanced-stage disease, respectively. The histology-specific sojourn times ranged from 0.8 to 1.8 years for type II ovarian cancer and 2.9 to 6.6 years for type I ovarian cancer.

Conclusions

Annual screening is not effective for all ovarian cancer subtypes. Screening sensitivity for early-stage ovarian cancers is not sufficient for substantial mortality reduction.

Ovarian cancer is the second-leading cause of death among all gynecologic cancers (1). It is estimated that approximately 19 680 new ovarian cancers will be diagnosed and 13 430 ovarian cancer deaths will occur in the United States in 2024 (1). Epithelial ovarian cancer accounts for 90% of all ovarian cancer cases and has several distinct histological subtypes (2,3). The most common subtype is high-grade serous carcinoma (HGSC), which accounts for approximately 52% of the cases (4). HGSC ovarian cancers, along with carcinosarcoma and not-otherwise-specified (NOS) subtypes, are classified as type II ovarian cancers. Type II ovarian cancers are typically aggressive and responsible for the majority of the ovarian cancer–attributable deaths. Type I ovarian cancers behave in a more indolent manner, tend to grow locally, and metastasize late. Type I ovarian cancers are made up of low-grade serous carcinoma (LGSC), mucinous, clear cell, and endometrioid subtypes. Although ovarian cancer is deadly at advanced stages, more than 70% of patients can survive for at least 5 years if the cancer is diagnosed at an early stage (5). Currently, however, only 30% of patients are diagnosed at an early stage (6,7), Most type I ovarian cancers are diagnosed at an early stage, whereas type II ovarian cancers are typically diagnosed at an advanced stage (8,9).

Ovarian cancer screening for the general population failed to yield statistically significant mortality reduction (10,11). The UK Collaborative Trial of Ovarian Cancer Screening (UKCTOCS) trial (10,12,13) and the Prostate, Lung, Colorectal and Ovarian (PLCO) Cancer Screening trial (11,14) evaluated the performance of multimodal ovarian cancer screening using a combination of the cancer antigen (CA) 125 biomarker and transvaginal ultrasound or transvaginal ultrasound alone. Serial CA-125 with the Risk of Ovarian Cancer Algorithm (ROCA) and subsequent transvaginal ultrasound screening, as performed in the UKCTOCS, showed statistically significant stage shift but not a statistically significant mortality reduction, suggesting that screening needs to detect ovarian cancer at an even earlier stage of the natural history of the disease.

Estimates of the sensitivity of screening and the time between when the cancer first becomes detectable via screening to symptom onset, a period commonly referred to as sojourn time, would be extremely useful to policy makers to design an effective and efficient ovarian cancer screening program. Because the objective of screening is to reduce mortality by detecting ovarian cancers at an earlier stage, the longer the sojourn time, the more likely we are to detect cancer at an early stage by screening (15). Estimation of these key parameters is challenging, because the preclinical phase of cancer is not directly observable, whereas the sensitivity of screening cannot be directly calculated because not all participants undergo definitive diagnostic testing. Thus, a statistical model is warranted for estimating these parameters using data from screening trials.

In this study, we used a previously published Bayesian framework to assess and compare the sensitivity and mean sojourn time associated with several ovarian cancer screening strategies evaluated in the UKCTOCS and PLCO trials. We estimated histology-specific ovarian cancer sojourn times and discuss potential implications for future screening programs.

Methods

Bayesian estimation approach

Shen and colleagues (16) developed a Bayesian framework that jointly estimates the mean sojourn time and sensitivity of screening using a continuous time Markov model. We adopted the Shen et al. (16) approach and used data from the UKCTOCS and the ovarian component of the PLCO screening trial to estimate the mean sojourn time and sensitivity of available screening modalities for ovarian cancer. Accordingly, we modeled the ovarian cancer progression as a 3-state continuous-time Markov chain (17) composed of the following states: no clinical disease state, preclinical screen-detectable disease state, and clinical disease state. State transitions were assumed to occur sequentially from less advanced to more advanced disease states. All women started at the “no clinical disease” state and transitioned to the preclinical screen-detectable disease state according to a transition rate ${\lambda }_1$. Note that although the preclinical disease state is not directly observable and patients are asymptomatic, ovarian cancer at a preclinical state can be detected through screening based on the sensitivity (S) of the modality used. From the preclinical screen-detectable state, patients may develop ovarian cancer–related symptoms that lead to a clinical diagnosis of the disease at a rate equal to λ₂. Our goal was to estimate these 3 parameters (λ₁, λ_2, and S) such that the estimated screen-detected, interval-diagnosed, and follow-up–diagnosed ovarian cancer cases approximate the corresponding reported number of ovarian cancer cases from the screening trials. Assuming that sojourn time follows an exponential distribution, the mean sojourn time is equal to $\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{${\lambda }_2$}\right.$(18). Furthermore, we expanded the framework proposed by Shen et al. (16) to estimate the sensitivity of screening and sojourn times for early (stages I and II) and advanced (stages III and IV) ovarian cancers separately. More details on the Bayesian framework and its extension are provided in Supplementary Methods Section 1 and Supplementary Figure 1 (available online).

Data

The PLCO cancer screening trial

The PLCO cancer screening trial was a randomized controlled trial that evaluated the effectiveness of ovarian cancer screening in asymptomatic women from the general US population aged 55 to 74 years (19). Eligible women received up to 6 annual screening exams or usual care. During the first 4 years, women in the screening arm underwent CA-125 biomarker testing and transvaginal ultrasound, whereas during the last 2 years, only CA-125 testing was completed (11). A CA-125 level of 35 U/mL or higher was considered abnormal. We estimated the sensitivity of screening as conducted in the PLCO trial in 2 ways. First, we pulled screenings from all 6 years and estimated the overall screening sensitivity. Moreover, we estimated 2 separate sensitivity levels, one for the first 4 years, during which multimodal screening was used, and another level for the last 2 years, during which only CA-125 was tested. We used deidentified, individual-level data from ever-screened participants randomly assigned in the screening arm of the PLCO trial to estimate the number of screen-detected cancers, number of interval cases, and the time from the last screening to diagnosis (Supplementary Methods Section 1, Supplementary Tables 1 and 2, available online).

UKCTOCS trial

The UKCTOCS trial was a randomized controlled trial conducted in the United Kingdom that evaluated and compared the effectiveness of 11 rounds of annual multimodal ovarian cancer screening using serial CA-125 measurements followed by second-line transvaginal ultrasound vs ultrasound alone vs no screening (10). The study protocol and main findings have been published elsewhere (13,20-22). In the UKCTOCS, CA-125 testing used the ROCA, which uses serial CA-125 measurements and statistical modeling to derive a woman’s probability of developing ovarian cancer (23). Unfortunately, we were unable to access individual-level data from the UKCTOCS, so we estimated the number of screen-detected cases and interval cases for each screening arm and the time between last screening exam and diagnosis based on summary data from related literature (10,21), as described in the Supplementary Methods Section 2 and Supplementary Tables 3 through 7 (available online).

The Surveillance, Epidemiology, and End Results Program

We extracted data on ovarian cancer incidence and survival by stage and histology from the Surveillance, Epidemiology, and End Results (SEER) 18 database for the years 2000-2018 (4). We derived hazard rates on survival by histology relative to HGSC for PLCO and applied them to calculate histology-specific sojourn times. We used HGSC as our reference group because most ovarian cancer cases recorded in SEER, the PLCO, and UKCTOCS were HGSC. Because the UKCTOCS’s histology-specific reported outcomes combined HGSC and LGSC as serous carcinoma ovarian cancers, we followed a similar approach to derive survival hazard rates relative to serous carcinoma from the UKCTOCS.

Statistical analysis

We estimated the overall sojourn time for all ovarian cancers combined, along with the sensitivity of the different screening modalities tested in the PLCO and UKCTOCS trials using the Bayesian framework proposed by Shen et al. (16). We used relevant literature and the SEER registry to inform the initial prior distributions for each estimated parameter. We validated our mean sojourn time estimates using 2 statistical tests (16)—namely, we used the Pearson’s χ² test to assess whether statistically significant differences between the observed and predicted number of ovarian cancer cases exist, and evaluated the quality of our model predictions by calculating the posterior predictive P value (24). A posterior predictive P value close to .5 indicates good fit for the model.

Histology-specific sojourn time

Because the histology subtypes used in the PLCO and UKCTOCS trials differed, we derived the histology-specific sojourn times for HGSC, LGSC, mucinous, clear cell, endometrioid, carcinosarcoma, and NOS subtypes in the PLCO trial and for serous, clear cell, endometrioid, and carcinosarcoma in the UKCTOCS trial by adjusting the overall sojourn time based on the survival hazard rate calculated from the SEER registry. More information about the methodology used to derive histology-specific sojourn times is presented in the Supplementary Methods Section 3, Supplementary Table 8 (available online).

Results

We informed the Bayesian framework using past related literature from the PLCO and UKCTOCS trials. The sensitivity of ovarian cancer screening was previously reported to range from 65% to 90% (13,14). Hence, we considered a β distribution with parameters α  =  5 and β  =  2 as the prior distribution for the sensitivity of available screening modalities, such that the mode of screening’s sensitivity would be 80%. Despite some differences in the screening modalities tested in the PLCO and UKCTOCS trials, we considered the same prior for all modalities because the reported sensitivities do not differ significantly between modalities and none of them showed statistically significant mortality benefit. The prior distribution for the ovarian cancer incidence rate (λ₁) was assumed to be Uniform, with a range of 0.0005 to 0.01 based on age-adjusted incidence rates of ovarian cancer estimated from the SEER-18 registry (4). The prior distribution for λ₂ was set to Uniform, with a range of 0.2 to 3.0, based on the study by Broder et al. (25), which reported sojourn time for ovarian cancer ranging from less than 1 year to 3 years. Finally, we assumed that the number of screen-detected cases at every screening exam followed a binomial distribution, whereas the number of interval cancer cases and those detected during the follow-up period followed a Poisson distribution.

Sojourn time and screening sensitivity based on the PLCO data

We applied the Bayesian framework to estimate the overall ovarian cancer sojourn time (all histologies and stages combined) and overall sensitivity of screening over the 6 years of screening in the PLCO trial using the aforementioned prior distributions to estimate the λ₁, λ₂, and S parameters. The posterior distributions for each of the iterations are shown in Figure 1, and the summary of the final posterior distributions for the 3 parameters is shown in Table 1. The mean overall sojourn time for ovarian cancer in the PLCO data was estimated as 2.1 years (95% confidence interval [CI] = 1.9 to 2.4 years). The overall sensitivity of screening over the 6 years in the PLCO trial was 65.7% (95% CI = 60.2% to 71.2%). Table 2 depicts the goodness of fit of the expected number of cases per screening round, as estimated by the model, compared with the observed number from the PLCO trial. The χ² test showed no statistically significant differences between observed and expected values (P = .502). Moreover, the posterior predictive P value was estimated as .469 (Figure 1, A), confirming that the models fit the PLCO data well.

Figure 1.

Posterior distributions of 3 parameters and scatter plot of predicted and observed log likelihood ratio discrepancies for the Bayesian model in the Prostate, Lung, Colorectal and Ovarian screening trial (A) for all 6 y of screening combined and (B) modeling the first 4 years (multimodal screening) and last 2 years (cancer antigen 125 testing alone) separately.

Table 1.

Summary of estimated parameters, mean sojourn time, and sensitivity of screening from the Prostate, Lung, Colorectal and Ovarian screening trial

Analysis	Parameter	Mean (95% Confidence interval)	SD
6 y of screening combined	λ1	0.000669 (0.000639 to 0.000699)	0.000016
	λ2	0.467 (0.416 to 0.515)	0.025
	Mean sojourn time, y	2.1 (1.9 to 2.4)	0.1
	Sensitivity	0.657 (0.602 to 0.712)	0.028
Multimodal screening (first 4 y) and cancer antigen 125 alone (last 2 y), separately	λ1	0.000664 (0.000632 to 0.000695)	0.000016
	λ2	0.496 (0.449 to 0.547)	0.026
	Mean sojourn time, y	2.0 (1.8 to 2.2)	0.1
	Sensitivity (multimodal)	0.808 (0.743 to 0.874)	0.035
	Sensitivity (cancer antigen 125)	0.362 (0.296 to 0.432)	0.0354

Table 2.

Comparison of observed vs predicted number of cases in the Prostate, Lung, Colorectal and Ovarian screening trial^a,b

Screening		6 y of screening combined			Multimodal screening (first 4 y) and cancer antigen 125 alone (last 2 y), separately
		Observed	Predicted	Residual	Observed	Predicted	Residual
1	No. of screened women	28716	28716.9	0.9	28716	28713.1	‒2.9
	No. of positive case	28	27.1	‒0.9	28	30.9	2.9
	No. of interval case	4	5.5	1.5	4	3.9	‒0.1
2	No. of screened women	27523	27525.1	2.1	27523	27525.3	2.3
	No. of positive case	17	14.9	‒2.1	17	14.7	‒2.3
	No. of interval case	7	6.3	‒0.7	7	4.9	‒2.1
3	No. of screened women	26566	26569.2	3.2	26566	26568.6	2.6
	No. of positive case	16	12.8	‒3.2	16	13.4	‒2.6
	No. of interval case	6	8.7	2.7	6	7.4	1.4
4	No. of screened women	25406	25409	3	25406	25408.2	2.2
	No. of positive case	15	12	‒3	15	12.8	‒2.2
	No. of interval case	8	12.1	4.1	8	10.7	2.7
5	No. of screened women	20111	20105.4	‒5.6	20111	20110.1	‒0.9
	No. of positive case	4	9.6	5.6	4	4.9	0.9
	No. of interval case	7	4.6	‒2.4	7	6.3	‒0.7
6	No. of screened women	22186	22181.4	‒4.6	22186	22185.8	‒0.2
	No. of positive case	7	11.6	4.6	7	7.2	0.2
	No. of interval case	8	5.4	‒2.6	8	8.3	0.3
Follow-up	No. of cases after screening	107	103.3	‒3.7	107	108.9	1.9

^aOne-sensitivity model: ${\chi }^2=12.321$, $P=.502$.

^bTwo-sensitivity model: ${\chi }^2=3.657$, $P=.994$.

We further estimated the ovarian cancer sojourn time and screening’s sensitivity separately for the first 4 years of the PLCO trial (multimodal screening with CA-125 and transvaginal ultrasound) and the last 2 years of the trial (CA-125 test alone). In this analysis, the overall ovarian cancer sojourn time was 2.0 years (95% CI = 1.8  to 2.2 years), which is comparable to our estimate, which combines all 6 years of screening in the PLCO trial (Table 1). The sensitivity of multimodal screening with annual CA-125 screening followed by transvaginal ultrasound, as was done in the first 4 years of the PLCO trial, was estimated to be 80.8% (95% CI = 74.3% to 87.4%), whereas screening’s sensitivity using the CA-125 test alone was estimated to be 36.2% (95% CI = 29.6% to 43.2%). The predicted ovarian cancer cases did not differ significantly from the observed cases (χ²  P = .994, posterior predictive P value = .606) (Table 2, Figure 1, B).

Moreover, we estimated the sensitivity of ovarian cancer screening and sojourn time for early-stage and advanced-stage ovarian cancers. Screening’s sensitivity for early-stage ovarian cancers (39.1%, 95% CI = 34.9% to 43.3%) was significantly lower than the sensitivity for advanced-stage cancers (82.9%, 95% CI = 78.0% to 87.8%) (Table 3;  Supplementary Figures 2 and 3, Supplementary Table 9, available online). Our findings were consistent when we analyzed the sensitivity of multimodal screening and CA-125 separately; sensitivity for early-stage ovarian cancers was estimated to be 46.6% (95% CI = 41.3% to 51.7%) with multimodal screening and 22.6% (95% CI = 16.2% to 29.1%) with CA-125 test alone, whereas the sensitivity for advanced ovarian cancers was estimated to be 93.5% (95% CI = 89.0% to 98.0%) with multimodal screening and 42.7% (95% CI = 33.1% to 52.7%) with CA-125 alone (Table 3;  Supplementary Figures 2 and 3, Supplementary Table 9, available online). The time staying in the preclinical early-stage and advanced-stage states was estimated at approximately 1 year (Table 3).

Table 3.

Estimated mean sojourn time and screening sensitivity, stratified by ovarian cancer stage, using data from the Prostate, Lung, Colorectal and Ovarian screening trial

Model	Parameter	Mean (95% Confidence interval)	SD
6 y combined	λ1	0.000649 (0.000623 to 0.000675)	0.000013
	λ2	0.931 (0.884 to 0.978)	0.024
	Mean time staying at early-stage preclinical state	1.1 (1.0 to 1.1)	0.03
	λ3	0.127 (0.112 to 0.142)	0.008
	Mean sojourn time for early-stage disease, y	7.9 (7.1 to 8.9)	0.5
	λ4	0.928 (0.880 to 0.979)	0.025
	Mean sojourn time for advanced-stage disease, y	1.1 (1.0 to 1.1)	0.03
	Sensitivity, early-stage disease	0.391 (0.349 to 0.434)	0.021
	Sensitivity, advanced-stage disease	0.829 (0.780 to 0.878)	0.025
Multimodal screening (first 4 y) and cancer antigen 125 alone (last 2 y), separately	λ1	0.000643 (0.000617 to 0.000670)	0.000014
	λ2	0.952 (0.912 to 0.991)	0.020
	Mean time staying at early-stage preclinical state	1.1 (1.0 to 1.1)	0.02
	λ3	0.129 (0.115 to 0.144)	0.008
	Mean sojourn time for early-stage disease, y	7.8 (6.9 to 8.7)	0.5
	λ4	0.948 (0.909 to 0.989)	0.021
	Mean sojourn time for advanced-stage disease, y	1.1 (1.0 to 1.1)	0.02
	Sensitivity of multimodal screening, early-stage disease	0.466 (0.413 to 0.517)	0.027
	Sensitivity of multimodal screening, advanced-stage disease	0.935 (0.890 to 0.980)	0.023
	Sensitivity of cancer antigen 125 test, early-stage disease	0.226 (0.162 to 0.291)	0.033
	Sensitivity of cancer antigen 125 test, advanced-stage disease	0.427 (0.331 to 0.527)	0.050

Sojourn time based on the UKCTOCS data

We applied the same framework to the 2 screening arms of the UKCTOCS trial. The iterative updates of the posterior distributions from the UKCTOCS data are shown in Figure 2, and the summary of the final distributions is shown in Table 4. The mean sojourn time in the transvaginal ultrasound arm was estimated to be 2.4 years (95% CI = 2.2  to 2.5 years). The mean sojourn time in the multimodal arm of the UKCTOCS trial was estimated to be 2.0 years (95% CI = 1.8 to 2.1 years). The sensitivity of the multimodal screening arm was better than the ultrasound screening arm (93.2.%, 95% CI = 88.9% to 97.3%, vs 64.5%, 95% CI = 60.4% to 68.8%). The P values calculated from χ² tests were .634 and .999 in the multimodal screening and ultrasound arms, respectively, which show that there was no statistically significant difference between observed and predicted outcomes (Supplementary Section 4, Supplementary Table 10, available online). The high quality of our estimates was also supported by the posterior predictive P value, which was estimated as .532 in multimodal screening arm and .640 in the ultrasound screening arm (Figure 2).

Figure 2.

Posterior distributions of 3 parameters and scatter plot of predicted and observed log likelihood ratio discrepancies for the Bayesian model in the (A) multimodal screening arm and (B) ultrasound screening arm of the UK Collaborative Trial of Ovarian Cancer Screening trial.

Table 4.

Summary of estimated parameters, mean sojourn time, and sensitivity of screening from the UK Collaborative Trial of Ovarian Cancer Screening trial

Arm	Parameter	Mean (95% Confidence interval)	SD
Multimodal screening	λ1	0.000585 (0.000568 to 0.000603)	0.000009
	λ2	0.507 (0.474 to 0.541)	0.017
	Mean sojourn time, y	2.0 (1.8 to 2.1)	0.1
	Sensitivity	0.932 (0.889 to 0.973)	0.022
Ultrasound screening	λ1	0.000555 (0.000538 to 0.000572)	0.000009
	λ2	0.425 (0.393 to 0.459)	0.017
	Mean sojourn time, y	2.4 (2.2 to 2.5)	0.1
	Sensitivity	0.645 (0.604 to 0.688)	0.022

Estimation of histology-specific sojourn time

We estimated the histology-specific survival hazard rates relative to HGSC and applied them to the overall mean sojourn time estimates to derive histology-specific sojourn times from the PLCO and UKCTOCS trials (Table 5). The estimated sojourn times for type II ovarian cancers (HGSC, carcinosarcoma, and NOS) ranged from 0.8 to 1.8 years, whereas sojourn times for type I ovarian cancers (LGSC, mucinous, clear cell, and endometrioid) ranged from 2.9 to 6.6 years.

Table 5.

Estimated hazard rate and sojourn time in each histology in the PLCO and the UKCTOCS trials^a

PLCO trial	HGSC	LGSC	Mucinous	Clear cell	Endometrioid	Carcinosarcoma	NOS
Hazard rate	1	0.37	0.49	0.45	0.27	1.54	2.02
Mean sojourn time = 2.0 y (PLCO multimodal screening arm [first 4 y] and CA-125 alone [last 2 y], separately)	1.7	4.5	3.4	3.7	6.2	1.1	0.8
Mean sojourn time = 2.1 y (PLCO = 6 y combined)	1.8	4.7	3.5	3.8	6.6	1.1	0.9
UKCTOCS trial	Serous		Mucinous	Clear cell	Endometrioid	Carcinosarcoma	NOS
Hazard rate	1		NA	0.46	0.26	1.64	NA
Mean sojourn time = 2.0 y (UKCTOCS multimodal screening arm)	1.7		NA	3.6	6.3	1.0	NA
Mean sojourn time = 2.4 y (UKCTOCS ultrasound screening arm)	1.3		NA	2.9	5.1	0.9	NA

^aCA=125, cancer antigen 125; HGSC = high-grade serous carcinoma; LGSC = low-grade serous carcinoma; NA = not applicable; NOS = not otherwise specified; PLCO = Prostate, Lung, colorectal Ovarian trial; UKCTOCS = UK Collaborative Trial of Ovarian Cancer Screening trial.

Discussion

Ovarian cancer screening trials for the general population failed to demonstrate statistically significant mortality reduction, tempering any enthusiasm for implementation of a screening program for the general population. Although a statistically significant stage shift was observed in the multimodal screening arm of the UKCTOCS trial, screening was not effective in reducing ovarian cancer mortality. To better understand why screening did not yield a mortality benefit, a more robust understanding of the preclinical phase of the disease and screening’s opportunity window to detect ovarian cancer is needed. Toward that goal, estimation of the ovarian cancer sojourn time—that is, the time from when the disease is screen detectable until clinical diagnosis—will provide important insights and inform future policy. Considering that the preclinical phase is not directly observable, mathematical modeling is warranted to estimate ovarian cancer sojourn time and the sensitivity associated with alternative screening modalities.

In this study, we used a previously published Bayesian framework to estimate the sojourn time for ovarian cancer, using data from the largest randomized controlled screening trials available to date: the PLCO and the UKCTOCS trials. We estimated the overall ovarian cancer sojourn time as well as histology-specific sojourn times, along with the sensitivity of 3 different screening modalities evaluated in past ovarian cancer screening trials: CA-125 with transvaginal ultrasound, serial CA-125 with transvaginal ultrasound, and transvaginal ultrasound alone. To the best of our knowledge, this study is the first to estimate ovarian cancer sojourn time associated with the above screening modalities.

Our findings provide insights into the effectiveness of readily available ovarian cancer screening modalities. We found that serial CA-125 testing using the ROCA with second-line transvaginal ultrasound has approximately 12% higher sensitivity than the multimodal screening test without the ROCA. None of the available screening modalities, however, has sufficient sensitivity to detect early-stage ovarian cancers. Our conclusions are consistent with findings from the UKCTOCS trial, which demonstrated a statistically significant stage shift in the multimodal screening arm compared with no screening but not in the ultrasound screening arm (10). Screening with CA-125 and transvaginal ultrasound in the PLCO trial also failed to demonstrate statistically significant stage shift (11). Furthermore, our screening sensitivity estimates are comparable to those reported in the PLCO and UKCTOCS trials. Screening’s sensitivity in the PLCO trial was reported to be 64.8% for type II and 85.7% for type I ovarian cancers (14), whereas in the UKCTOCS trial, screening’s sensitivity was 89.5% and 75.0% for the multimodal screening and ultrasound screening arms, respectively (13).

Our findings provide some indication of why screening trials failed to demonstrate statistically significant mortality reduction. We demonstrated that annual screening frequency may not be the optimal screening frequency for ovarian cancer. Considering that ovarian cancer sojourn time was estimated between 2.0 and 2.4 years combined with the very low sensitivity of screening at early stages of the disease, annual screening is unlikely to detect ovarian cancer cases early enough in their natural history to significantly improve mortality outcomes. This is also evident from the UKCTOCS trial’s failure to translate the observed statistically significant stage shift into a mortality benefit, which suggests that we need to detect ovarian cancer even earlier in the disease progression for available treatments to be effective. Our findings are consistent with previously published modeling work suggesting that annual screening for ovarian cancer is unlikely to be effective in reducing ovarian cancer mortality (26,27).

The differences in histology-specific sojourn times highlight the heterogeneity of the disease. Most noteworthy, we showed that type II ovarian cancers (HGSC, carcinosarcoma, and NOS) have a much shorter sojourn time, ranging from 0.8 to 1.8 years, than type I ovarian cancers (LGSC, mucinous, clear cell, endometrioid), which have sojourn times between 2.9 and 6.6 years. It is important to note that type II ovarian cancers account for most ovarian cancer deaths. Hence, even though annual screening may shift the time of detection for type I ovarian cancers to earlier disease stages, it is unlikely to be effective in detecting type II ovarian cancers sufficiently early in the natural history of the disease. This finding provides another possible explanation for the ineffectiveness of screening in yielding mortality benefit, despite the observed stage shift. Most of the screen-detected cases in the UKCTOCS trial’s multimodal screening arm were type II ovarian cancers, which may partly explain the lack of mortality benefit. At the same time, it is possible that the stage shift stemmed predominately from screen-detected type I ovarian cancers, which are easier to detect at earlier stages.

An effective ovarian cancer screening program may require more frequent screening (eg, every 6 months) to provide more opportunities to detect the disease at an early stage. Policy makers, however, must consider the balance between the potential benefits and harms associated with such a screening program. Existing screening modalities upon an abnormal finding require subsequent invasive diagnostic procedures, such as biopsies, to reach a definitive diagnosis. These invasive procedures could cause anxiety, physical harm, and financial burden to the individual. More frequent screening will unavoidably increase the potential harms associated with downstream diagnostic management of abnormal screening findings. Hence, the feasibility of more aggressive ovarian cancer screening programs is conditioned on the development and availability of new and improved screening modalities, which can provide definitive cancer diagnosis and reduce the false-positive rate compared with the existing screening modalities. Alternatively, annual screening may become effective if new screening modalities are developed that can detect ovarian cancer even earlier than existing modalities (ie, improving the sensitivity of screening, thus prolonging the sojourn time of ovarian cancer). Improvement in screening’s sensitivity, however, is likely to compromise the specificity of the screening modality and thereby increase the false-positive findings. Therefore, more research is warranted to determine whether an acceptable threshold of the harm/benefit trade-off for ovarian cancer screening exists and what that threshold may be. Decision-analytic modeling approaches were used to formally assess the trade-offs associated with cancer interventions for other cancer sites (28-31), and these approaches could be used to guide the development of future ovarian cancer screening programs. The recent decline in ovarian cancer incidence along with the relatively low prevalence of the disease pose an additional challenge in future screening trials—specifically, to be sufficiently powered to assess mortality benefit, future trials will require enrollment of hundreds of thousands of participants, with cost implications that may question their feasibility. Future ovarian cancer screening trials may require targeting women at high risk of developing ovarian cancer rather than the general population to increase the prevalence of the disease among the screen-eligible individuals. Identifying women at high ovarian cancer risk would necessitate a formal risk assessment with sufficiently high accuracy. Past attempts to develop accurate prediction models for ovarian cancer were not successful because they produced models with modest predictive performance (32-36). More research is warranted to develop better ovarian cancer risk prediction models. Considering the challenges in developing an effective screening for ovarian cancer, at the present time, primary prevention for ovarian cancer is perhaps the most rational approach to achieve meaningful reductions in ovarian cancer morbidity and mortality. A plethora of primary preventive interventions have been associated with notable reduction in ovarian cancer risk, including bilateral salpingo-oophorectomy (37,38), oral contraception (39-41), tubal ligation (42), and opportunistic salpingectomy (43-45).

Our study is, to the best of our knowledge, the first that quantified the sojourn time for ovarian cancer. Broader et al. (25) reported ovarian cancer sojourn time estimates based on the expert opinion of experienced experts. According to their report, the sojourn time of stage I, II, and III ovarian cancers were 3 years, 2 years, and less than 1 year, respectively. Broader and colleagues’ estimates were collected through questionnaires, so they are not based on data nor any mathematical analysis. Numerous studies exist that provide sojourn time estimates for other cancer sites with established screening programs (46-49). Our sojourn time estimates for ovarian cancer (2.0-2.4 years) are significantly shorter than the sojourn times of other cancer sites. For example, sojourn time for prostate cancer was estimated between 11.3 and 12.6 years (49), for lung cancer between 3 and 6 years (50,51), between 1.98 and 4.3 years for breast cancer (46,52,53), and between 3.6 and 4.3 years for colorectal cancer (47). The short sojourn time for ovarian cancer is biologically plausible considering the anatomy of the ovaries and the female reproductive system in general. Particularly for HGSC, assuming that most cancers originate in the tubal epithelium, there are no physical barriers to restrict the spread of malignant cells to the surface of the ovary (stage I ovarian cancer). Similarly, no physical barriers exist to limit the spread of cancer from the surface of the ovaries to the peritoneum (stage III ovarian cancer). In part, this may explain the low prevalence of stage II ovarian cancer because it is likely that there is no intermediary stage II disease. The shorter ovarian cancer sojourn time underlines the challenges of developing an effective ovarian cancer screening program for the general population.

Our study has limitations. Despite multiple efforts, we were unable to obtain individual-level data from the UKCTOCS trial. Although we informed our models using published summary data from the UKCTOCS trial, there is no way to guarantee the accuracy of those estimates. Moreover, it was impossible to directly estimate histology-specific sojourn times from the screening trials because in the PLCO trial, the numbers for histologies other than HGCS were not adequate to allow reliable estimation of sojourn times, and we did not have histology-specific outcomes from the UKCTOCS trial. Hence, our histology-specific sojourn time estimates assumed a fixed growth rate for ovarian cancer across the different stages of the disease, which may not be the case.

In conclusion, we showed that annual screening for ovarian cancer is unlikely to be effective in reducing ovarian cancer mortality given the relatively short opportunity window to detect type II ovarian cancers. Multimodal screening using serial CA-125 testing and transvaginal ultrasound has the highest sensitivity among the existing modalities, but it is not sufficient to detect early-stage ovarian cancer. More research is warranted to evaluate the clinical utility of more frequent ovarian cancer screening with multimodal screening and to develop better screening modalities for ovarian cancer.

Data availability

The data underlying this article are available upon request to the PLCO Cancer Data Access System (https://cdas.cancer.gov/plco/). All analytical results generated during this project have been made available in the manuscript and Supplementary Materials.

Author contributions

Sayaka Ishizawa, PhD (Data curation; Formal analysis; Investigation; Methodology; Validation; Visualization; Writing—original draft);Jiangong Niu, PhD (Data curation; Formal analysis; Methodology; Writing—review & editing);Martin Tammemagi, DVM, MSc, PhD (Methodology; Writing—review & editing);Ehsan Irajizad, PhD (Investigation; Methodology; Writing—review & editing);Yu Shen, PhD (Investigation; Methodology; Writing—review & editing);Karen Lu, MD (Writing—review & editing);Larissa Meyer, MD, MPH, FACOG, FACS (Conceptualization; Formal analysis; Investigation; Methodology; Project administration; Resources; Supervision; Validation; Writing—review & editing);Iakovos Toumazis, PhD (Conceptualization; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Supervision; Validation; Visualization; Writing—original draft; Writing—review & editing).

Funding

This work was supported by a grant awarded to I.T. from The University of Texas MD Anderson Cancer Center Duncan Family Institute for Cancer Prevention and Risk Assessment. The funders did not play a role in the design of the study; the collection, analysis, and interpretation of the data; the writing of the manuscript; or the decision to submit the manuscript for publication.

Conflicts of interest

I.T. received research support from Break Through Cancer unrelated to the written work research project. L.A.M. received research support from Break Through Cancer, AstraZeneca, and the National Cancer Institute for work unrelated to the written work research project. K.L. received research support from Break Through Cancer for work unrelated to the written work research project. All other authors have nothing to disclose.