The impact of pancreatic cancer screening on life expectancy: A systematic review of modeling studies

Abstract Evidence supporting the effectiveness of pancreatic cancer (PC) screening is scant. Most clinical studies concern small populations with short follow‐up durations. Mathematical models are useful to estimate long‐term effects of PC screening using short‐term indicators. This systematic review aims to evaluate the impact of PC screening on life expectancy (LE) in model‐based studies. Therefore, we searched four databases (Embase, Medline, Web‐of‐science, Cochrane) until 30 May 2022 to identify model‐based studies evaluating the impact of PC screening on LE in different risk populations. Two authors independently screened identified papers, extracted data and assessed the methodological quality of studies. A descriptive analysis was performed and the impact of screening strategies on LE of different risk groups was reported. Our search resulted in 419 studies, of which eight met the eligibility criteria (mathematical model, PC screening, LE). Reported relative risks (RR) for PC varied from 1 to 70. In higher risk individuals (RR > 5), annual screening (by imaging with 56% sensitivity for HGD/early stage PC) predicted to increase LE of screened individuals by 20 to 260 days. In the general population, one‐time PC screening was estimated to decrease LE (2‐110 days), depending on the test characteristics and treatment mortality risk. In conclusion, although the models use different and sometimes outdated or unrealistic assumptions, it seems that PC screening in high‐risk populations improves LE, and that this gain increases with a higher PC risk. Updated model studies, with data from large clinical trials are necessary to predict the long‐term effect of PC screening more accurately.


What's new?
Robust evidence on the effectiveness of pancreatic cancer screening is lacking, as published studies are mostly cohort studies, limited in size and duration. Here, the authors performed a systematic review to evaluate the impact of pancreatic cancer screening on life expectancy as determined in model-based studies. The results show that, while it could be detrimental in the general population, pancreatic cancer screening in high-risk populations may increase life expectancy. Updated model studies, with data from large clinical trials, are necessary to predict the long-term effect of pancreatic cancer screening more accurately.

| INTRODUCTION
Pancreatic cancer (PC) is mostly diagnosed at an incurable stage. 1 Early detection by screening might improve survival, but robust evidence on the effectiveness of PC screening is lacking, as published studies are cohort studies, limited in size and duration. Ideally, large clinical trials with decades of follow-up would provide solid evidence on the efficacy of screening. Since such studies are unlikely to be concluded promptly, mathematical models can be used to predict long-term outcomes, using short-term indicators from observational studies.
Mathematical models, also known as decision-analytic models, can simulate health outcomes of individual patients or a population under a variety of scenarios. They can be used to estimate the consequences of distinct screening scenarios under varying circumstances.
Several types of mathematical models exist, such as decision trees, Markov models and microsimulation models.
These models all have a similar principle of decision processes.
These processes relate to potential events that can occur. For example, an individual can be at a certain risk to develop a precursor lesion.
Based on this risk, a precursor lesion will or will not develop during his or her lifetime. If a precursor lesion develops, the associated cancer risk will determine the chance that it will progress to a malignancy.
Decision processes can be relatively simple, such as a decision tree, or more complex, such as a microsimulation model. In a decision tree the risk of events and states of nature is diagramed over a fixed time horizon. Markov models simulate a hypothetical cohort of individuals through health states over time. A microsimulation model simulates individual life histories and tracks the past health states and future events for each individual. 2 These decision-analytic models can inform policy makers on optimal screening strategies for several cancer types. Also, they can estimate the cost-effectiveness and harms and benefits of a screening program. 3,4 In the past, these models have been used to quantify the effects of screening for cancer such as breast, colon, cervix, and lung cancer in a population. [3][4][5][6] Mathematical modeling is, for example, used to translate colorectal cancer screening trial results (that showed screening effectiveness in a controlled setting) to real world health outcomes for the total population and a longer time period. 7,8 In the last decade, several mathematical models on PC screening have been developed to evaluate the efficacy of screening with a variety of screening modalities, screening strategies, and risk populations. [9][10][11][12][13] The objective of the current systematic review is to gain insight into the estimated effects of these mathematical models pertaining life expectancy (LE).

| METHODS
This systematic review was executed according to the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guideline.
The review protocol is available in PROSPERO (CRD42020168804).
We searched four electronic databases (Embase, Ovid Medline, Web of Science, Cochrane central) for studies published before 30 May 2022.
The search strategies used for each database are presented in the Supplementary Materials (Data S1).

| Inclusion and exclusion criteria
We searched for studies that used a mathematical model to evaluate the effect of PC screening on LE in different PC risk groups. Mathematical models were defined as models using given input to simulate a decision process with the help of algorithms (eg, simulation, Markov, microsimulation or decision tree models). Our search was limited to papers written in English and involving human subjects. We excluded studies that evaluated the effect of PC treatment. We also excluded reviews, in vitro studies, case reports and letters.

| Quality assessment
We adapted the ISPOR-AMCP-NPC Questionnaire 14

| Data synthesis and analysis
We performed a descriptive analysis and summarized the effect of PC screening on LE. We compared the screening effect on LE for the different risk groups in the included studies (general population, high-risk group, different risk levels/mutation carriers).

| Models
Our search resulted in 514 articles, of which 11 underwent full text review after title and abstract screening ( Figure 1). Four papers were excluded based on study type (review) and/or model outcome (LE was not reported). One paper was found by evaluation of reference lists.
Altogether, eight articles met the inclusion criteria, published between 2003 and 2021. Table 1  Because of a paucity of available data on cystic lesions, similar test characteristics were assumed for high risk cysts.
Rulyak et al used an EUS sensitivity of 90% for the detection of dysplasia based on expert opinion. 12 An abnormal EUS was defined as two or more of the following abnormalities: heterogenous parenchyma with echogenic foci, hypoechoic nodules, hyperechoic main duct wall or discrete masses. A positive test led to a diagnostic Endoscopic Retrograde Cholangiopancreatography (ERCP) for confirmation. Koopmann et al used preliminary data from a large surveillance cohort to assess a stage dependent sensitivity (60%-99%) for a combined EUS/MRI test. 13 An overview of the extracted data from the included models is presented in Table 2.

| Screening in a high-risk population
Four studies reported on the effect of screening in different high-risk populations. 9,10,12,13 The effect of PC screening on LE depended on the relative risk (RR) for PC of the screened population and the test characteristics (Table 3). Pandharipande

| Screening individuals with pancreatic cystic lesions
Two studies evaluated the effect of screening in individuals with pancreatic cysts. 15

| Screening in the general population
Two studies evaluated the effect of PC screening in the general population. 9,11 The first reported that one-time screening either at age

| Risk of bias
Especially in the field of PC, where data are scarce and evidence is based on small studies, the risk of bias regarding input or target data is relatively high. Figure 2 provides an overview of the risk of bias of the included articles per item. Bias assessment per study is available in the Supplementary Materials (Data S1).

| DISCUSSION
This systematic review evaluates model-based analyses on the effect of PC screening on LE. PC screening in the general population led to decreased LE in most analyses, as screening benefits were outweighed by false positive tests and surgical mortality risks. For higherrisk individuals, screening was more beneficial, but this conclusion strongly depends on model assumptions such as test characteristics.
The effect of surveillance on LE seemed small in the slightly increased risk group (RR < 4.5). However, in this analysis, an unrealistic low test sensitivity (56%) for HGD/early stage PC were used in this analysis.
As a result, these predictions likely underestimate the potential effect of screening. In individuals at higher risk (RR 6.4-60, eg, CDKN2A), both annual and one-time screening improved LE, even with a poor screening test. Although most papers were based on outdated and unrealistic assumptions, they do show that PC screening can be worthwhile. In the future, predictions from these models will improve when they are updated with reliable input data from ongoing longterm studies.
In higher-risk individuals (RR: 30), the expected median LE gain of 260 days for annual PC screening seems worthwhile. In comparison, colorectal cancer screening with an annual FIT test in the general population is estimated to increase LE by 89 days per screened individual. 18 For lung cancer, 30 days are gained by annual screening with a low dose CT of 50 to 75-year old males who smoke. 19 However, LE gain is not the only parameter to determine effectiveness of a screening program. Other factors, such as impact on quality of life, but also harms and costs should be considered. Moreover, morbidity associated with the screening methodology or the surgical treatment (pancreatic fistula, exocrine pancreatic insufficiency and diabetes) was not taken into account in most studies, although this is highly relevant.
The effect of screening on LE varies within the models because of different model assumptions, for instance on the natural history. In two models, PC could only evolve from "pancreatic dysplasia" or adenomas. 12,15 Distinct precursor types such as IPMN and PanIN, that can progress from LGD to HGD into cancer, were not defined. The models that did simulate both precursors in separate pathways assumed that 90% of PC developed from PanIN and 10% from cystic lesions. [9][10][11] If the proportion of cyst derived cancers is actually higher, as some studies indicate, [20][21][22] these models would also underestimate the efficacy of screening.
Input on cyst prevalence is needed to simulate a population at (increased) risk for pancreatic precursor lesions and PC. [9][10][11]13 The reported cyst prevalence in the general population varies from 2.6% to 55%. This range is caused by differences in imaging techniques, populations and cyst definition (size and type). Pandharipande et al used an imaging based cyst prevalence of 4.6% at age 50 as a calibration target. An increased cyst prevalence was assumed in the models that simulated a higher risk group, as is consistent with literature. 23,24 Next to differences in natural history, the models incorporated different screen scenarios and assumed different screen tests charac- The effect of PC screening in ongoing clinical trials is mainly based on early detection of cancer rather than resection of HGD cyst, as published data shows that the number of resected HGD lesions is low (n = 0, 31 n = 0, 23 n = 10 25 ). Precursor lesions that have been resected were mainly LGD. As the majority of LGD cysts will never progress to PC (given the high cyst prevalence and low PC incidence), the impact of a resected LGD cyst on life expectancy will be low and possibly negative because of treatment complications. Trials also show that patients with early detected and resected PC still have a limited life expectancy. 23 The 5 year survival rate of T1 patients was reported to be 30.6%. 32 For future research, focus should therefore lay on the detection and treatment of HGD precursor lesions. The detection of early stage cancer, even with annual screening, seems challenging and screen detected early stage PC remains associated with a decreased LE.
Unfortunately, current screen tests are underperforming in the detection of high-grade dysplastic precursor lesions. Despite the existence of multiple risk prediction tools, it remains difficult to classify cysts and determine their grade of dysplasia correctly, based on imaging. 27 Also, test specificity for PanIN lesions is nowhere near 97%, while this lesion is believed to account for 50% to 90% of all PC. Thus, it is important to search for new and better screen tests. In this regard biomarkers hold promise, not only to detect early cancer, but also to search for timely identification of (high grade dysplastic) precursor lesions. 33 In conclusion, even though most included papers used out-