Identifying key factors for the effectiveness of pancreatic cancer screening: A model‐based analysis

Abstract Pancreatic cancer (PC) survival is poor, as detection usually occurs late, when treatment options are limited. Screening of high‐risk individuals may enable early detection and a more favorable prognosis. Knowledge gaps prohibit establishing the effectiveness of screening. We developed a Microsimulation Screening Analysis model to analyze the impact of relevant uncertainties on the effect of PC screening in high‐risk individuals. The model simulates two base cases: one in which lesions always progress to PC and one in which indolent and faster progressive lesions coexist. For each base case, the effect of annual and 5‐yearly screening with endoscopic ultrasonography/magnetic resonance imaging was evaluated. The impact of variance in PC risk, screening test characteristics and surgery‐related mortality was evaluated using sensitivity analyses. Screening resulted in a reduction of PC mortality by at least 16% in all simulated scenarios. This reduction depended strongly on the natural disease course (annual screening: −57% for “Progressive‐only” vs −41% for “Indolent Included”). The number of screen and surveillance tests needed to prevent one cancer death was impacted most by PC risk. A 10% increase in test sensitivity reduced mortality by 1.9% at most. Test specificity is important for the number of surveillance tests. In conclusion, screening reduces PC mortality in all modeled scenarios. The natural disease course and PC risk strongly determines the effectiveness of screening. Test sensitivity seems of lesser influence than specificity. Future research should gain more insight in PC pathobiology to establish the true value of PC screening in high‐risk individuals.


| INTRODUCTION
Pancreatic cancer (PC) has one of the poorest survival rates of all human cancers and is ranked among the top five of cancer-related deaths. 1 Although the general incidence is low (lifetime risk of 1.3%), it is substantially increased in certain high-risk groups, with relative risks up to 132 in Peutz Jeghers Syndrome, for instance. 1 Approximately 10% of PC cases are found in individuals with an inherited increased risk for this fatal disease. This high-risk population includes carriers of PC-prone gene mutations (eg, BRCA2, CDKN2A and LKB1) and first-degree relatives of familial PC patients. Familial PC patients have at least (a) one first-degree relative with PC, (b) one seconddegree relative <50 years at time of PC diagnosis or (c) two relatives with PC without a known gene mutation. 1 Several studies have shown that screening individuals at high risk for PC can detect precursor lesions and early stage cancer. 1 However, we lack solid evidence that the benefits of screening (reduced PC-related morbidity and mortality) outweigh its harms, such as patient burden, overdiagnosis and overtreatment. Determination of the effectiveness of screening requires large-scale and prolonged follow-up.
Computer simulation models have proven their value in estimating the long-term impact of screening using short-term indicators such as PC incidence and detection rates. The Microsimulation Screening Analysis (MISCAN) model, for example, has been used for the evaluation and optimization of screening programs for cancer of the cervix, breast, colon, prostate, lung and esophagus. [2][3][4][5][6][7] This model has also been applied to test hypotheses on disease development 3 and test characteristics. 4 In order to create an accurate simulation model, detailed information on the disease is necessary. Current knowledge gaps in PC screening include the natural disease course, test characteristics (sensitivity, specificity) for different disease stages, and the PC risk level.
Decision-analytic modeling can be used to explore such uncertainties and quantify the impact of individual factors on the effect of screening. 2,4 Identification of the factors with the strongest impact on screening may guide future research.
In our study, we aim to identify key parameters that define the effect of PC screening and to analyze their impact on the effect of PC screening in a high-risk population using the microsimulation model MISCAN. Consequently, we will identify areas to which further research should be directed. Here, using a microsimulation model, the authors explored uncertainties concerning the early detection of pancreatic cancer and analyzed the impact of these uncertainties on the effect of screening. In all simulated scenarios, screening was associated with reduced pancreatic cancer mortality. The effectiveness of screening was most strongly impacted by characteristics of natural disease course and level of pancreatic cancer risk base cases with contrasting disease pathways, to look at the effect of different progression patterns on screening outcomes. In the first base case, we simulate a path where all lesions progress from preclinical disease to clinical cancer, in an average of 14.3 years, based on the scarce evidence available. 9 We assume durations are exponentially distributed (Weibull distribution, Shape 1) and that the duration of each disease stage is correlated with the next: in other words, when a lesion is fast growing in one stage, it does so in the next. Durations varied based on random selection in the distribution. There are no indolent lesions in this base case.
Since the average dwelling time of 14.3 years is debated, we model a second base case, in which both faster progressive and indolent (slow developing) lesions are simulated. In case of slow developing disease, lethal PC will never occur in the course of a lifetime. Preinvasive stages will at most progress to preclinical cancer Stage I. This is assumed to take over 30 years. In this base case, the preinvasive lesions that do progress to For both base cases, we use PC lifetime risk as a calibration target. We have created a high-risk population by multiplying the PC lifetime risk of the general population by 10. Figure 1 shows the prevalence of each disease stage by age, for both the "Progressiveonly" and "Indolent Included" pathways.

| Assumptions for screening, surveillance and treatment
For both base cases, we simulated a hypothetical cohort of 1 000 000 high-risk individuals from birth until death with and without screening.
When screening is implemented, these individuals are offered yearly or 5-yearly screening from the ages of 50 to 75. Individuals are assumed to adhere to all screening rounds. The screen test characteristics are equivalent to a combination of both an endoscopic ultrasonography (EUS) and a magnetic resonance imaging (MRI), as they provide the most accurate pancreatic imaging. 1 The screen test is considered positive when a preinvasive or invasive lesion is detected, rightfully or wrongfully. Surveillance tests are defined as additional tests, performed every 6 months after a positive screen test until the lesion was resected ( Figure 2). Test characteristics of the surveillance test are assumed identical to those of the screen test.
As explained in the abovementioned paragraph on natural disease course, disease development occurs in different stages. We assume test characteristics differ between these stages, based on preliminary results of a PC screening study cohort ( Figure 2). We assume that 90% of normal pancreata (no lesions) are correctly identified on imaging and 10% will be diagnosed as false positive. These individuals will be referred to more intensive follow-up (surveillance test) or surgical resection.

| Sensitivity analyses
The effect of different sensitivity analyses on NNS for both pathways in case of annual screening are visualized in Figure 3. An overview of all sensitivity analyses on different outcome measures is provided in Table 3 (annual screening) and in Appendix Tables 2 and 3 (5-yearly screening).

| Treatment mortality
Varying treatment mortality from 3% to 5% resulted in an increase in The interval cancer rate is presented as the number of cancer cases (per 100 000 LYs) in the first 5 years after a negative screening test and in the total period after a negative screening test (including after age 75) (screen detected [SD] cancer cases after a negative screening test are not included).
The sensitivity of the screening test has a much smaller effect on mortality reduction. The specificity of the test, however, is of particular importance for the number of surveillance tests (associated with burden) needed to prevent one PC death.
The influence of the test sensitivity on the NNS was negligible. This is partly caused by the fact that the probability that someone is referred for resection after a positive test was not varied (Appendix In this exploratory analysis, the only included negative aspect of treatment was a 3% to 5% surgical mortality risk. Therefore, a potential limitation is that we omitted other negative aspects, such as morbidity and loss in quality of life, as a result of screening, surveillance or resection. These aspects need consideration, since pancreatic surgery is associated with significant morbidity (40%-60%), 10,11 such as delayed gastric emptying, wound infections and pancreatic fistulae, in addition to diabetes mellitus and/or exocrine insufficiency as late complications. Integrating these harms in the model will likely result in a less favorable effect. Furthermore, the assumed 100% attendance rate for PC screening might have led to an overestimation on the population effect. However, literature shows that high risk individuals are likely to participate in PC screening. 14 Also, we did not consider systematic false-negative test results in our model. The sensitivity of a test is important, but repeated testing in a screening program generally enables missed lesions to be detected in a next screening-round.
However, some lesions may be systematically missed due to their size or (pre-existing) parenchymal changes.
The NNS estimated for breast, cervical, colorectal and prostate cancer screening ranges from 1000 to 2000. 15,16 Our estimation that approximately 500 persons need to be screened to prevent one PC death is significantly lower. Although many of the factors that influence the NNS in our model are based on assumptions, none of the sensitivity analyses resulted in a NNS of >1500. Moreover, the screening instruments (ie, EUS, MRI) used to detect (preinvasive) PC are more invasive than those used in currently implemented screening programs. Also, they are more expensive (eg, approximately €700 for EUS, 17 as compared to the <€60 spend by other screening programs 5,7,18  intervention. This may be due to the assumed high morbidity rates and loss in quality of life after total pancreatectomy.
To conclude, we showed that the natural disease course of PC and its precursor lesions is one of the determining factors of the success of pancreatic screening in high-risk individuals. The risk for developing PC in the target population is another factor that plays a major role. Test sensitivity has a minor influence. Both base cases show that under plausible assumptions PC screening might be promising in a high-risk population. The current study underlines the importance of continued research pertaining the development of PC and differential risks within specific target populations. This is of interest not only to improve test strategies based on imaging, but also on biomarkers in serum and secretin stimulated pancreatic juice.