Screening for pancreatic cancer has the potential to save lives, but is it practical?

Abstract

Introduction:

Most patients with pancreatic cancer present with advanced stage, incurable disease. However, patients with high-grade precancerous lesions and many patients with low-stage disease can be cured with surgery, suggesting that early detection has the potential to improve survival. While serum CA19.9 has been a long-standing biomarker used for pancreatic cancer disease monitoring, its low sensitivity and poor specificity have driven investigators to hunt for better diagnostic markers.

Areas covered:

This review will cover recent advances in genetics, proteomics, imaging, and artificial intelligence, which offer opportunities for the early detection of curable pancreatic neoplasms.

Expert opinion:

From exosomes, to circulating tumor DNA, to subtle changes on imaging, we know much more now about the biology and clinical manifestations of early pancreatic neoplasia than we did just five years ago. The overriding challenge, however, remains the development of a practical approach to screen for a relatively rare, but deadly, disease that is often treated with complex surgery. It is our hope that future advances will bring us closer to an effective and financially sound approach for the early detection of pancreatic cancer and its precursors.

1. Introduction

NORC, a research institute at the University of Chicago, recently estimated that only 14% of all cancers diagnosed in the United States are detected by screening [1]. Even in this background of low overall screening rates, pancreatic cancer stands out as a deadly malignancy without a widely-accepted screening program. Most patients with pancreatic cancer are not diagnosed until after the cancer has locally progressed or metastasized, and the prognosis for patients with advanced pancreatic cancer is dismal [2]. By contrast, patients diagnosed with non-invasive precancerous lesions can be cured, and patients diagnosed with low-stage invasive pancreatic cancers have promising survival rates [3–8]. These and other findings suggest that the earlier detection of pancreatic cancer and its precursors offers hope to improve survival for this disease. Indeed, Goggins and Canto have recently reported, in an uncontrolled trial, that the median survival for patients with a screen-detected pancreatic cancer (9.8 years) is more than six times longer than the median survival for patients with pancreatic cancer diagnosed outside of surveillance (1.5 years) [9].

An enormous challenge for the early detection of pancreatic cancer and its precursors is the relative rarity of the disease. While it has been estimated that 62,210 Americans were diagnosed with pancreatic cancer in 2022, the population of the country by the end of 2022 was about 334 million [2]. This means that only 0.0186 percent of the population developed pancreatic cancer that year. Screening tests with sensitivities and specificities near 100% would be needed to effectively detect most of these cancers in the asymptomatic general population without unnecessarily alarming too many people with false positives or overtreating lesions that would never have progressed even without treatment [10]. Another essential characteristic of an effective screening test is its ability to detect curable disease. For pancreatic cancer, that’s primarily Stage I disease, but few patients are diagnosed with Stage I pancreatic cancer outside of pancreatic surveillance programs. Most biomarker studies involve patients with advanced-stage disease whose biomarker alterations often reflect derangements associated with advanced disease.

The pressing question, then, is how will we get to a sensitive, specific, and cost-effective approach to the early diagnosis of pancreatic cancer and its precursors? To help answer this question, we first describe available screening approaches, as well as some potential advances on the horizon. We then review the performance characteristics needed for a pancreatic cancer screening program, and we describe ways to improve screening performance by prioritizing high-risk populations. Finally, we look forward to emerging technologies and opportunities.

2. CA19–9

Cancer antigen 19–9 (CA19–9, or sialylated Lewis (a) antigen) is one of the best-known markers of pancreatic cancer [11]. CA19–9, a monosialoganglioside, is produced by most pancreatic cancers, and CA19–9 levels are easily measured in the blood. Serum CA19–9 levels have proven to be useful in predicting resectability and in monitoring treatment response in patients known to have pancreatic cancer [12–16]. Warshaw and colleagues, for example, found that a postoperative decline in CA19–9 levels is a strong predictor of survival after surgical resection [17]. However, several other malignancies, and even a number of benign conditions, including obstructive jaundice, lung disease, liver failure, and acute and chronic pancreatitis, are also associated with elevated serum CA19–9 levels [18]. Furthermore, germline variants in the CA19–9 synthetic pathway influence serum levels, including those in FUT3, the enzyme responsible for CA19–9 synthesis [19]. Depending on race, as much as 10% of the population lack any functional FUT3 [20,21]. Pancreatic cancers that arise in these “Lewis antigen negative” individuals will not make significant amounts CA19–9, and their cancers are, therefore, generally non-detectable using this marker [19,22].

3. Other serum protein and glycoprotein markers

Other potential serum markers of pancreatic cancer have been identified using a variety of technologies, including serial analysis of gene expression (SAGE), gene expression arrays, mRNA sequencing, and mass spectrometry (Table 1) [23–29]. For example, one of genes first found to be overexpressed in pancreatic cancer by SAGE was tissue inhibitor of metalloproteinase type 1 (TIMP-1). Hanash and colleagues have found that the combination of serum TIMP1 and CA19–9 levels is more sensitive in detecting presymptomatic pancreatic cancers than CA19–9 levels alone [23,29].

Table 1:

Selected circulating protein markers of pancreatic cancer

Marker	Full name(s)	Function	References
ApoA2 isoforms	Apolipoprotein A2	Component of high-density lipoprotein particles	[30,31]
CA125	Carcinoma antigen 125, Mucin 16	O-glycosylated protein that functions in forming a protective mucous barrier	[32,33]
CA19–9	Cancer antigen 19–9, carbohydrate antigen sialyl Lewis a	A glycosphingolipid that functions in embryogenesis that is secondarily absorbed to red blood cells	[31,34–36]
CEA	Carcinoembryonic antigen, carcinoembryonic Antigen-Related Cell Adhesion Molecule 5	Cell surface glycoprotein that functions in cell adhesion	[19,33,35,37,38]
CPA1	Carboxypeptidase A1	A zinc metalloprotease that cleaves select dietary proteins	[34]
DUPAN-2	Duke pancreatic monoclonal antigen type 2	A precursor of CA 19–9	[39]
GDF15/MIC-1	Growth differentiation factor 15	A secreted ligand in the transforming growth factor-beta superfamily of proteins	[40,41]
LRG1	Leucine rich alpha-2-glycoprotein 1	Functions in protein-protein interactions, cell adhesion, and signal transduction	[42,43]
MUC5AC	Mucin 5AC	Protects mucosal surfaces, and functions in phosphatidylinositol-mediated signaling	[44]
SPP1	Osteopontin	Attachment of osteoclasts to mineralized bone matrix (hydroxyapatite)	[33,38,45,46]
THBS2	Thrombospondin-2	A glycoprotein that mediates cell-cell and cell-matrix interactions	[47–49]
TIMP1	Tissue inhibitor of metalloproteinases 1	As the name suggests, inhibits matrix metalloproteinases	[23,50]

Several technologies have been employed to identify other potential new protein and glycoprotein blood markers since the discovery of TIMP1 [24,51–54]. For example, Cao and colleagues, as part of the Clinical Proteomic Tumor Analysis Consortium (CPTAC), conducted a comprehensive proteomic, phosphoproteomic, and glycoproteomic analysis of 140 pancreatic cancers. They identified over 200 proteins expressed at higher levels in tumors compared to normal pancreas and macrodissected normal pancreatic ducts [24]. Importantly, many of these proteins were similarly abundant in low-stage cancers, and a number of them were predicted to be secreted proteins, making them attractive targets for early detection [24]. Glycoproteomic analyses in this study identified 75 N-linked glycoproteins upregulated more than two-fold in the cancers [24]. As expected, mucin-type O-linked glycoproteins associated with CA19–9 were found to be upregulated in pancreatic malignancy [24]. As with the proteomic analysis, many of the glycoproteins identified were also upregulated in low-stage disease, suggesting that they may be good markers for early detection [24]. Thus, a host of potential new markers of early pancreatic cancer have been identified, and several, such as thrombospondin 2, have been shown to improve upon the diagnostic accuracy of CA19–9 alone [48].

The added value of these newly discovered blood-based proteins, when used individually as markers for pancreatic cancer, is unclear. Boyd and colleagues conducted a meta-analysis of 28 primary studies and found that while several novel protein biomarkers have moderate diagnostic accuracy, they do not outperform CA19–9 in distinguishing between patients with and without pancreatic cancer [11]. When they are used in combination, however, these new markers may, however, be useful in Lewis antigen non-secretors. Kane and colleagues conducted a similar analysis using PRISMA standards and found that pooling biomarkers, particularly TIMP-1, CEA, CA125, and CA19–9, can improve the diagnostic accuracy over CA19–9 alone [55].

4. Using genotyping to improve blood-based protein and glycoprotein markers

One challenge to using the protein and glycoprotein markers identified to date is that the expression of these markers can be influenced by inherited gene variants, creating significant person-to-person variability in the normal ranges of the blood levels of these markers [19,34,56]. Abe and colleagues proposed to overcome this challenge by using germline genotyping to create germline variant-specific tumor marker reference ranges for CA19–9, CA125 and carcinoembryonic antigen (CEA)[19,34]. Tanaka and colleagues performed a similar study evaluating serum carboxypeptidase A (CPA) and CA19–9 levels, coupled with determinations of germline CPA1 and germline Lewis and secretor gene genotypes, in 345 controls and 190 patients with pancreatic cancer and found that the combination of germline variant-stratified CPA and CA19–9 levels achieved specificity levels >98% for pancreatic cancer [34]. Thus, “personalized” biomarker marker reference ranges offer one potential route for improving the early detection of pancreatic cancer.

5. Branch chain amino acids

Mayers and colleagues reported that elevated plasma levels of branched chain amino acids are a predictive marker for developing pancreatic cancer [57]. Using blood samples from four large prospective cohort studies, the team found that the branched chain amino acids isoleucine, leucine and valine, were significantly elevated in some individuals years before they developed pancreatic cancer [57]. The authors hypothesized that early pancreatic cancers are associated with an increased whole-body breakdown of proteins and that the resultant increase in branched chain amino acids could be used as a marker for the early detection of pancreatic cancer [57].

In addition, it appears that dietary intake high in branched chain amino acids is associated with pancreatic cancer risk and that branched chain amino acid metabolism can contribute to the growth of pancreatic cancer [58–62]. Subsequent studies have validated these findings and have confirmed that branched chain amino acids can be elevated years, and sometimes even a decade, before patients become symptomatic and are diagnosed with pancreatic cancer [62].

6. Metabolites

Alterations in the blood levels of a number of metabolites have been reported in patients with pancreatic cancer [43,63]. Mahajan and colleagues reported that a test incorporating the serum levels of multiple metabolites with CA19–9 could achieve an area under the curve (AUC) of >90% in distinguishing patients with pancreatic cancer from controls [63]. Similarly, Fahrmann and colleagues reported that a plasma metabolite panel, which included acetylspermidine, diacetylspermine, an indole-derivative, and two lysophosphatidylcholines, improved the performance of CA19–9, LRG1 and TIMP1 in detecting low-stage pancreatic cancer [43].

One of the challenges of discovering metabolites specific for pancreatic cancer, and indeed of any blood-based protein or glycoprotein marker of pancreatic cancer, is that patients with pancreatic cancer often have a number of comorbidities in addition to their cancer that can dramatically impact the analytes measured. For example, virtually all cancers that arise in the head of the pancreas obstruct the distal common bile duct as it runs through the pancreas, and, as noted with CA19–9, this biliary obstruction will alter the metabolism of some analytes [64]. Thus, markers identified by comparing healthy controls to patients with pancreatic cancer may be detecting non-specific physiologic changes, such as biliary obstruction, rather than the development of pancreatic cancer itself. As a result, alterations in analytes may not be specific for pancreatic cancer if applied to real-world populations [64].

7. Exosomes

Extracellular vesicles function in cell-cell communication [65]. Exosomes, a type of extracellular vesicle, have an average diameter of 100 nanometers and can contain a variety of materials including nucleic acids and proteins [65]. Cancer cells may secrete exosomes, which have been shown to increase the susceptibility of distant organs to metastatic seeding and growth [66]. Since exosomes are shed into extracellular spaces, including blood, exosomal “cargo” is an attractive target for early detection. Indeed, Nakamura and others reported that exosomal miRNAs in the blood could serve as markers for early pancreatic cancer [67–70]. As is true for the blood protein markers, the combination of exosomal markers with serum CA19–9 levels may prove more sensitive and specific than any one marker in isolation [71].

8. Circulating tumor DNA

One of the most active targets for the early detection of pancreatic and other cancers is circulating tumor DNA (ctDNA). Pancreatic cancer, like almost all other cancers, is driven by the accumulation of somatic mutations in a set of well-defined genes [52,72–74]. DNA harboring these cancer-specific somatic mutations is released in small quantities into the blood, and this ctDNA can be detected using modern sequencing methods [33,75–79]. Here we briefly describe some of the many different cancer-specific changes that can be detected in ctDNA.

The most common alterations used in ctDNA tests are somatic intragenic mutations [33,77]. The KRAS gene is somatically mutated in ~95% of pancreatic cancers, and extremely sensitive tests have been developed to detected rare mutant KRAS alleles admixed with thousands of wild-type KRAS alleles [75,80–82]. For example, MacGregor-Das and colleagues determined the plasma levels of mutant KRAS and GNAS in 67 patients with pancreatic cancer and in 73 healthy controls using digital next generation sequencing. They found that a third of the patients with low-stage cancer had detectable mutations in their blood, while mutant alleles were rarely detected in healthy individuals [80].

Aberrantly methylated DNA is another somatic cancer-specific change that can be detected in blood [78,83–86]. Ying and colleagues reported that a four gene methylation panel, which included the ADAMTS1, BNC1, LRFN5, and PXDN genes, was accurate in detecting pancreatic cancer, and Kandimalla and colleagues reported a separate methylation panel that produced areas under the curve (AUCs) of 0.85 in distinguishing patients with low-stage pancreatic cancer from controls [87,88].

Velculescu and others have described novel approaches to cancer detection based on the patterns of DNA fragmentation detected in the blood [76,79]. They reported that cell free DNA in blood samples from healthy individuals reflected the nucleosomal patterns of normal white blood cells, while patients with cancer had abnormal circulating DNA fragmentation patterns. These circulating DNA fragmentation patterns could be used to accurately identify the presence of a cancer and, remarkably, often predict the organ of origin of that cancer [79]. Adding to these observations, Mouliere and colleagues found that integrating the analyses of DNA fragment size improved the detection of cancer-specific mutations in ctDNA [76]. This fragmentation-based approach is being commercialized through the biotech company Delfi.

Jamshidi and colleagues recently compared some of the many approaches to detecting cancer-specific DNA in blood [89]. Using data from the Circulating Cell-free Genome Atlas study, they were able to compare whole-genome methylation, single nucleotide variants, somatic copy number alterations, and fragmentation pattern approaches. They found that the whole genome methylation approach was among the most sensitive of the methods and best predicted cancer origin. Some of the authors of the Jamshidi study are affiliated with the cancer screening company Grail, which is pursuing a methylation-based approach to multi-cancer detection.

Combining ctDNA markers with other markers will likely improve the accuracy of early detection tests. For example, Cohen and colleagues reported high sensitivity and specificity with a blood test that combines ctDNA testing for somatic mutations with a panel of serum protein biomarkers [38]. In this study, the combinatorial approach achieved a specificity of 99.5% with sensitivity of 64% [38].

Unfortunately, ctDNA levels are significantly lower in patients with low-stage cancers than they are in patients with high-stage disease, meaning that approaches based on ctDNA are more likely to detect advanced cancers than they are to detect early curable cancers [89]. For example, the company GRAIL recently reported the findings for a trial screening 6,662 individuals over the age of 50 using their multi-cancer early detection test [90]. Only 36 cancers were detected, and about half of them were advanced cancers (stage III or IV). The United Kingdom is currently performing a large, randomized trial to evaluate whether the GRAIL test (Galleri) can reduce the incidence of advanced cancers (stage III and IV) 3–4 years after randomization [91].

Another commercialized cancer screening test from the company Thrive Earlier Detection has clinical data available to evaluate. In a large non-randomized study of Thrive’s multi-analyte blood-based screening test performed by some of the authors of this review, 26 cancers were detectable among 9,911 apparently healthy subjects using a combined ctDNA and protein approach [92]. Of the 26 cancers identified, 17 were advanced (stage III or stage IV), one was stage unknown, and only eight were low-stage cancers (stage I or II) [92]. In addition, some of the 26 patients with screen detectable cancers already showed symptoms. The challenges of ctDNA-based screening are significantly greater if the goal is to detect curable precancerous lesions, because non-invasive lesions, even large precancers with high-grade dysplasia, appear not to release significant amounts of mutant DNA into the blood [93].

Although mutant DNA from precancerous lesions in the pancreas may not be shed into the blood in significant quantities, it is frequently shed into the pancreatic duct system (since precursor lesions arise in the ducts), and this DNA will eventually pass into the stool [94–97]. Yu and colleagues tested secretin-stimulated pancreatic juice samples collected at the time of endoscopy and found that mutant DNA could be detected in patients with invasive pancreatic cancer as well as patients with an intraductal papillary mucinous neoplasm [94]. Of note, the study included at-risk patients who were not known to have pancreatic cancer, and four of these at-risk study subjects later developed pancreatic cancer. Remarkably, two of these four had SMAD4/TP53 gene mutations detected in their pancreatic juice over a year before they were clinically diagnosed with cancer. These results highlight the potential value of biosamples obtained from sources closer to the organ being screened.

Nucleic acid-based screening assays have already appeared in clinics. There are at least six companies with tests for the early detection of cancer testing in various phases of development (Table 2). Three are developing tests based on methylation of ctDNA (EarlyDiagnostics, GRAIL and Singlera), one a test based on hydoxy-methylation of ctDNA (Bluestar/ClearNote Health), one uses ctDNA fragment length (DELFI), and finally one is developing a test based on the combination of ctDNA mutations and tumor protein markers (Thrive/Exact Biosciences).

Table 2:

Nucleic Acid-based screening assays for early detection¹

Company/PI	Test Name	Sample	Analyte/Assay	Status	Clinical Trials (NCT number)	References2
Bluestar Genomics/ ClearNote Health	Avantect pancreatic cancer test	PB	5-OH-methyl-cytosine	Development/ Live	NCT03869814 NCT05188586 NCT05188573	[98,99]
Delfi	NA	PB	ctDNA, molecule length	Development, anticipated 2023	NCT04825834 NCT05306288	[79,100–102]
EarlyDiagnostics	cfMethyl-Seq	PB	ctDNA methylation	Development	ND, Trials planned for liver and lung cancer	[103–105]
Exact Biosciences/ Thrive Early Detection	CancerSEEK/ Detect-A	PB	ctDNA and protein markers	Development	NCT04213326 NCT04213326	[92]
Grail	Galleri	PB	ctDNA methylation	Live	NCT05481697 NCT05611632 NCT05205967	[106–108]
Singlera Genomics	PanSeer, PDACatch	PB	ctDNA methylation	Development	NCT05336058 NCT05159544 NCT05431621 NCT05485077 NCT05444491 NCT05432128 NCT05536089 NCT05336539 NCT03828396 NCT05626985 NCT03685669	[109–113]

^1.Assays for earlier detection of cancer are also referred to as multi-cancer early detection (MCED), in addition to liquid biopsy.

^2.As determined by an author or funding from the company, or use of the test in the study. NA: not available. ND: none detected in clinicaltrails.gov. PB= peripheral blood.

9. Autoantibodies

The somatic mutations that accumulate in neoplastic cells as they progress to invasive pancreatic cancer may create neoantigens when the genes altered are translated into proteins, and some people will develop antibodies to these neoantigens [114–117]. As reviewed by Dumstrei and colleagues, these novel autoantibodies have the potential to serve as markers for the detection of pancreatic cancer [114]. Their review concluded that single autoantibodies make poor markers, while multiple serum autoantibodies, when combined with other blood markers, such as CA19–9, can have better diagnostic performance. Nonetheless, most reports on the value of autoantibodies have not been validated across multiple studies, and the overall value of this approach has not been established [114].

10. Imaging

A variety of imaging modalities have been used to detect pancreatic neoplasia, as well as the changes that can occur in the body secondary to pancreatic cancer (Table 3) [118–124]. Endoscopic ultrasound (EUS) is probably the most sensitive of the pancreatic imaging modalities, and it has been used to detect asymptomatic precursor lesions, including intraductal papillary mucinous neoplasms, small invasive pancreatic cancers, and subtle secondary changes in the pancreas suggestive of multiple microscopic precursor lesions (pancreatic intraepithelial neoplasia (PanIN)) [120–126].

Table 3:

Comparison of imaging modalities

Modality	Region of Interest	Resolution	Ionizing Radiation	Advantages	Disadvantages
Pancreatic protocol CT	Whole abdomen	Millimeter	Yes	• Easily accessible  • Relatively cheap  • High spatial resolution for detection of small detections and local staging  • Detection of distant metastatic disease	• Small masses can be occult on CT  • Concerns with radiation safety limit its potential use as screening modality  • Contraindicated in patients with renal insufficiency or contrast allergy
MRI/MRCP	Whole abdomen	Millimeter	No	• Accurate detection of pancreatic cysts and suspicious features  • Accurate detection of pancreatic duct abnormality  • Can detect pancreatic masses that are occult on CT  • Can be performed without contrast for patients with renal insufficiency or contrast allergy	• Less accessible and more expensive than CT  • More operator dependent than CT
Endoscopic ultrasound	Pancreas and adjacent organs only	Millimeter	No	• Accurate characterization of pancreatic mass and pancreatic duct abnormality  • Obtain fine needle aspiration for tissue diagnosis	• Invasive procedure  • Rare procedural complications  • Limited assessment of disease extent outside pancreas

CT= Computed tomography, MRCP=magnetic resonance cholangiopancreatography, MRI= magnetic resonance imaging, US= ultrasound

Computed tomography and magnetic resonance imaging can also provide detailed images of the pancreas, and a number of subtle changes, including dilatation of the main pancreatic duct, focal pancreatic atrophy and an abrupt cut-off of the duct, may be present years before patients are diagnosed with pancreatic cancer [127–129]. Image analysis with 3D rendering techniques like Cinematic Rendering have the potential for earlier lesion detection by looking at numerous features, such as pancreatic texture, rather than simply looking for a mass [130]. Textural changes may prove to be among the earliest finding and have often been the critical finding in review of pre-diagnostic CT scans (Figure 1).

Figure 1:

Computed tomography of a patient with pancreatic cancer (arrows). Standard axial (A) and coronal (B) views demonstrating a hypointense mass in the body of the pancreas with upstream dilatation of the main pancreatic duct. Cinematic rendering (C and D), in similar planes, highlights the mass.

Harinck and colleagues compared endoscopic ultrasound to magnetic resonance imaging in a screening study of 139 asymptomatic high-risk individuals [131]. They found that the two imaging techniques are complementary, each with their own strengths and weaknesses.

Improvements in both image resolution and image analysis are on the horizon [132]. For example, photon counting CT has the potential to deliver 5 times higher resolution than standard CT [132,133]. In addition, several recent technological advances in image analysis have allowed investigators to detect features in images that the human eye might not perceive, and these have the potential to improve the sensitivity and specificity of imaging as a screening tool. Chu and colleagues, for example, applied radiomics, the mathematical integration of features such as shape, size, volume and texture, to CT of the pancreas and showed that radiomics features helped differentiate scans from patients with pancreatic cancer from scans from healthy controls, with sensitivities of 100% and specificities of 98.5% [134]. As discussed later in the section on artificial intelligence, these and other new technologies have the potential to improve the performance of imaging techniques.

The general challenge with using imaging for screening is that these tests tend to be expensive, some involve radiation (CT scans), and some (EUS) are invasive. Therefore, imaging-based screening programs cannot easily be applied to the general population.

11. New-onset diabetes

Pancreatic cancers cause a number of profound changes in the body. Some of these changes occur before patients develop symptoms, and these changes may therefore serve as early detection markers. Perhaps the most notable of these are alterations in blood glucose levels [135–140]. Chari and others have shown that many or even most patients with pancreatic cancer develop either hyperglycemia or overt diabetes, sometimes years before their cancer is diagnosed [141]. However, there are challenges with detecting early pancreatic cancer based on a diagnosis of new-onset diabetes (NOD) alone; first, only one in several hundred individuals with NOD will have pancreatic cancer, and second, diabetes associated with pancreatic cancer becomes more likely with increasing tumor burden [136,139,142,143]. Nonetheless, several groups have suggested a potential simple approach to screening for pancreatic cancer: all elderly patients with new onset diabetes could be prioritized for pancreatic imaging [135–140].

12. Other clinical features of pancreatic cancer

Muscle wasting is also commonly found in patients with pancreatic cancer and, like new onset diabetes in the elderly, muscle wasting could serve as an early indicator of pancreatic cancer [144–146]. For example, in the IMPACT study out of Italy, 73% of individuals with pancreatic cancer had sarcopenia at presentation [145]. Similarly, many patients with pancreatic cancer develop depression, and some even develop depression before they are diagnosed with cancer [147,148].

The problem with indirect clinical markers of pancreatic cancer is that they are common, non-specific, and all too often the health care professionals who care for these medical problems are unaware of their association with pancreatic cancer. Individuals who develop diabetes are treated by primary care physicians or endocrinologists, not oncologists. Even if these specialists are increasingly becoming aware of the association between physiological changes and pancreatic cancer, they have not been provided with well-established diagnostic algorithms to pursue this possibility.

13. Screening without intent

Artificial intelligence (AI) is one possible approach to overcoming many of the challenges noted above. For example, AI-based algorithms have been used to identify patterns in health care records that might indicate an increased likelihood of having pancreatic cancer [149–151]. Such algorithms are only as good as the information available in the medical record, but could eventually be used to automatically notify the patient’s health care provider that their patient has an increased risk of developing pancreatic cancer, and these algorithms could even suggest the best next steps to evaluate the patient. Such an approach, however, would require careful clinical validation to ensure it is providing actionable results without causing undue alarm.

AI can also be applied to images, and AI-based algorithms have the potential to be trained to detect early curable pancreatic cancers in CT scans [152–155]. These AI-based algorithms can achieve similar diagnostic performance to subspecialized academic radiologists and can potentially elevate the performance of an average community radiologist to the level of an expert. AI-based algorithms may have a role in early detection as reviews of pre-diagnostic CT scans from patients who subsequently developed pancreatic cancer have identified small, curable pancreatic cancers that went undetected by radiologists [118,119]. These small cancers produced subtle changes, such as focal atrophy of the gland, faint enhancement changes in the gland, dilatation of the main pancreatic duct, or an abrupt cut-off of the duct [118,119]. AI-based algorithms can be trained to recognize these subtle changes on pre-diagnostic (median time to diagnosis 386 days) CT scans with AUC up to 0.98 in a retrospective study, which can lead to significantly earlier diagnosis [153]. Since so many people in the United States have CT scans for reasons unrelated to their pancreas, AI-based algorithms could run in the background with the potential to detect early, curable pancreatic cancers [151,156]. Importantly, the application of AI-based algorithms to routine abdominal imaging would not incur the expense, added radiation exposure, and inconvenience of a population-wide screening program. Instead, AI-based algorithms can take advantage of what is already being done for other clinical indications and serve as a peer review mechanism to identify potentially clinically significant findings that are overlooked by radiologists [156,157]. On the other hand, relying solely on ad-hoc imaging will limit the potential reach of this approach since most people do not undergo abdominal imaging.

14. Low incidence as a barrier to screening

As noted earlier, although pancreatic cancer is deadly when it strikes, it is a relatively rare disease. Data from the Surveillance, Epidemiology and End Results (SEER) program of the National Cancer Institute have been used to estimate that, in the United States, the annual rate of new cases of pancreatic cancer is 13.3 per 100,000 for men and 11.1 per 100,000 for women [158]. Screening for a disease that strikes such a small fraction of the population is problematic [10,89]. One could improve the odds by focusing on the older adults, as 90% of pancreatic cancers occur in individuals 55 and older. However, tests with anything short of near perfect specificity will lead to numerous false positive results, needlessly alarming healthy people and potentially leading to unnecessary, costly, and potentially harmful interventions (Figure 2) [10,158–160]. Conversely, tests with low sensitivity will lower the number of true positive tests, thereby increasing the cost of the test per cancer detected and potentially offer false reassurance to participants. For example, we previously estimated that if 100,000 Americans over the age of 55 were screened using a test with a specificity of 98% and a sensitivity of 100%, the test would produce 1,999 false positive results for every 68 true positive results [10]. Screening tests for diseases such as pancreatic cancer that have a low annual incidence of disease must therefore have extraordinarily high specificity to avoid the negative consequences of large numbers of false positive results relative to true positives [89,159]. As Gray and colleagues stated in their classic 2008 paper on screening, “All screening programs do harm; some do good as well, and, of these, some do more good than harm at reasonable cost“ [159]. The question then becomes: how can one improve both pre-test probability and test performance to reduce the number of false positives?

Figure 2:

The rarity of pancreatic cancer in the general population (top left) remains a significant challenge to screening for the disease. Screening the general population, even with a highly sensitive and specific test, will result in numerous false positives (frowning faces) relative to true positives (smiling faces). Efforts to define populations with the greatest risk (lower left), will improve the ratio of true positives to false positives. (Grey indicates individuals affected with a high-grade precancer or early invasive pancreatic cancer. In the interest of space, the proportions of individuals affected in each group are not accurate).

15. Defining individuals with the greatest risk

Any successful screening program for pancreatic cancer must include efforts to focus the program on those with a risk of developing the disease that is sufficiently increased to overcome the adverse risks of screening. Age, family history of pancreatic cancer, germline gene status, and new onset diabetes mellitus are all associated with an increased risk, and the magnitude of this risk has been quantified for each, making them plausible approaches to identifying high-risk groups for screening (Table 4).

Table 4:

Risk factors for pancreatic cancer

Risk Factor	Prevalence and Associated Risk	References
Non-genetic
Cigarette Smoking	1.7 fold increased risk vs to never smokers.	[161]
Family history of pancreatic cancer	2.5 fold increased risk for individuals with a first degree relative with pancreatic cancer, risk increases with additional affected relatives and/or family members who developed cancer before age 50.	[162]
Long Standing Diabetes	1.5 – 2-fold increased risk of PDAC for individuals with diabetes >3 years in duration.	[163–166]
New onset Diabetes	<0.3% to 0.8% of new diabetes develop pancreatic ductal adenocarcinoma primarily within one year of diabetes diagnosis.	[143,167,168]
Obesity	1.6 fold increased risk in obese vs normal weight.	[169]
Genes
ATM pathogenic variant	Occurs in 1–3% of PDAC cases, 6.5 fold increased risk.	[170–172]
BRCA1	Occurs in 1% of PDAC, 2–4 fold increased risk.	[173–176]
BRCA2 pathogenic variant	Occurs in 2–7% of PDAC cases, 3.5–10 fold increased risk.	[174–180]
CDKN2A pathogenic variant	Occurs <0.5% of PDAC cases, 34 -fold increased risk.	[175,181]
CPA1/CPB1 variants	<1%; lifetime risk not defined	[182,183]
GWAS Variants & Polygenic Risk Scores	European ancestry populations: 1q32.1 (NR5A2), 1p36.33 (NOC2L), 2p13.3 (ETAA1), 3q29 (TP63), 5p15.33 (CLPTM1L, TERT), 7p14.1 (INHBA), 8q21.11 (HNF4G) 8q24.21 (MYC), 9q34.2 (ABO), 13q12.2 (PDX1), 13q22.1 (KLF5), 16q23.1 (BCAR1), 17q12 (HNF1B) 17q25.1 (LINC00673), 18q21.32 (GRP), and 22q12.1 (ZNRF). China: 21q21.3, 5p13.1, 21q22.3, 22q13.32 and 10q26.11: Japan 6p25.3, 12p11.21 and 7q36.2. In Europeans ~4 fold higher risk for those with the top 10% of polygenic risk (combined weighted average of associated loci).	[184–189]
Mismatch Repair Gene	Occurs in up to 1% of PDAC, 8 fold increased risk.	[190]
PALB2 pathogenic variant	Occurs <0.5% of PDAC cases, up to 10-fold risk.	[191]
PRSS1	Very Rare; >40 fold.	[192–194]
STK11	Very Rare: 10–40 fold.	[195–197]

Having a first-degree family member with pancreatic cancer increases one’s risk of developing precancerous and invasive pancreatic cancer, and having multiple family members increases the risk further [162,198–202]. The impact of family cancer history on pancreatic cancer risk is illustrated in a prospective study of 21,141 individuals in 4,433 families enrolled in the National Familial Pancreas Tumor Registry, conducted by Porter and colleagues [162]. They found that individuals in familial pancreatic cancer kindreds, defined as families in which at-least two first-degree family members had been diagnosed with pancreatic cancer, had a 4.86-fold increased risk of developing pancreatic cancer themselves [162]. Furthermore, the risk in the familial pancreatic cancer kindreds increased as the number of first-degree relatives (FDRs) with pancreatic cancer increased. Individuals with one FDR with pancreatic cancer had a 3.46-fold increased risk, those with two FDRs with pancreatic cancer had a 5.44-fold increased risk, and those with three or more FDRs with pancreatic cancer had a 10.78-fold increased risk [162]. The risk was even higher in families in which one of the family members developed pancreatic cancer before the age of 50 [162,203].

The risk associated with a family history of cancer can be refined if the pathogenic germline variant driving that risk can be identified (Table 4) [122,170–172,182,194,204–214]. For example, individuals with a pathogenic germline ATM variant have a 6.5-fold increased risk of developing pancreatic cancer [172]. Looking at other genes, carriers of pathogenic germline BRCA2 variants have a 3.5–10-fold increased risk of developing pancreatic cancer, carriers of a pathogenic germline CDKN2A variants have a 34-fold increased risk, and carriers of a pathogenic germline STK11 variant have a 10–40-fold increases risk [122,170–172,182,194,204–214]. Establishing age-specific risk estimates for all pancreatic cancer susceptibility genes will be important to optimize clinical surveillance protocols and early detection initiatives.

New onset diabetes, as discussed earlier, can be the first symptom of pancreatic cancer. While early studies suggested that elderly individuals with new-onset diabetes have up to an eight-fold increased risk of developing pancreatic cancer compared to healthy individuals, population-based studies in the Veterans Administration health system have shown that this risk is lower [141,167].

Other risk factors can be added to the mix (Table 4). For example, cigarette smokers have double the risk, obesity increases risk, and individuals of non-O blood groups are almost twice as likely to develop pancreatic cancer than are individuals of blood type O [215–217]. Unfortunately, even when one combines most of the leading risk factors together (including cigarette smoking, obesity, diabetes, family history and non-O ABO blood group), relatively few people have a combination of risk factors suggesting a risk sufficient to warrant screening with currently available tests [218]. As a result, it is challenging to identify a significant number of individuals who have a very high risk of developing pancreatic cancer and who would most benefit from screening for pancreatic cancer [218]. Furthermore, those with the greatest individual risk for pancreatic cancer may not contribute most to the total incidence of pancreatic cancer. As a result, a highly-targeted pancreatic cancer screening program with good performance is not guaranteed to markedly reduce the total population-level mortality. Most pancreatic cancers, for instance, are not associated with a known high-risk germline genetic predisposition [215]. To achieve the greatest population-level impact, moderate or even low-risk individuals might need to be included in a screening program, exposing many of them to unnecessary testing. This has been referred to as the “prevention paradox [219].”

16. Costs

The last words of J. Gray’s statement on screening, “All screening programmes do harm; some do good as well, and, of these, some do more good than harm at reasonable cost,” remind us that costs need to be considered when instituting a screening program [159]. One estimate of the financial costs, using Medicare and national average pricing, suggested that a screening program based on magnetic resonance imaging/magnetic resonance cholangiopancreatography could be affordable if applied to a high-risk population [220]. In particular, using Medicare data, the cost per “year of life added” was $638.62 for screening individuals with the Peutz-Jeghers syndrome and $945.33 for individuals with hereditary pancreatitis [220]. The costs were even lower, only $356.42, for obese smokers over the age of 50 with new-onset diabetes [220]. Corral and colleagues came to a similar conclusion, that screening can be financially reasonable for individuals with a very high (>20-fold) increased risk of pancreatic cancer [221].

The direct financial costs are, however, not the only costs of screening. One also has to consider the anxiety generated and the real risk of harm caused by false positive results. False positives can lead to additional testing and sometimes even to unnecessary invasive procedures. Unneeded biopsies may cause bleeding or pancreatitis in a minority of cases, while surgical overtreatment of low-risk precancerous lesions can lead to substantial recovery times, post-operative complications, and lifelong insulin dependence.

17. Actual results from screening

We have described the basis for screening and theories on how to improve the odds of screening, but what about real-life efforts to screen for pancreatic cancer? As shown in Table 5, there have been many studies targeting at-risk populations using different technologies [120–126]. Most studies employed combinations of imaging modalities, including endoscopic ultrasound, magnetic resonance imaging, and computed tomography. Imaging modalities are often alternated during sequential screening rounds, as each modality has its own advantages and disadvantages. For example, endoscopic ultrasound is a more sensitive and specific than CT, particularly for early disease, but requires an invasive approach [222]. CT, on the other hand, exposes the individual being screened to radiation.

Table 5:

Screening studies that included at least 100 study participants

First author	Year	Total number of participants	Primary mode of screening	Number of PDACs	Surgically resected PDACS	Surgeries for high-grade dysplasia	Surgeries for benign/low-risk lesions	Reference
Al-Sukhni	2012	262	MRI	3	1	0	2	[123]
Canto	2018	354	EUS, CT, MRI	14	10	10	23	[120]
Dbouk	2022	1461	EUS, CT, MRI	10	8	3	5	[9] Overlap with Canto (2018)
Harinck	2016	139	EUS, MRI	1	1	0	1	[131]
Klatte	2022	347	MRI, EUS	36	27	0	8	[122] Overlap with Vasen (2016)
Konings	2017	186	MRI, EUS	2	1	0	2	[223]
Lachter	2018	123	EUS	1	1	0	0	[224]
Ludwig	2011	109	MRI, EUS	1	1	1	4	[225]
Overbeek	2022	366	MRI, EUS	10	6	0	8	[121]
Paiella	2019	187	MRI, EUS	5	2	0	2	[226]
Vasen	2016	411	MRI, EUS	15	11	4	14	[125]
Zubarik	2011	546	CA19–9, EUS	1	1	0	1	[124]

CT=computed tomography, EUS=endoscopic ultrasound, MRI= magnetic resonance imaging, PDAC= pancreatic ductal adenocarcinoma

Table 5 describes selected studies of pancreatic cancer screening that included at least 100 subjects. In these studies, surgically resectable pancreatic cancers were detected by screening in a minority of participants. Many, though not all, of these screen-detected cancers were low stage [120–126,223,225,226]. For example, Dbouk and colleagues enrolled 1,461 high-risk individuals in the Cancer of the Pancreas Screening (CAPS) program. Seven of the 10 patients who were diagnosed with pancreatic cancer in this study had stage I disease, and three individuals with imaging abnormalities resulting from high-grade dysplasia were also successfully detected [9]. Similarly, Klatte and colleagues, detected 36 pancreatic cancers when they screened 347 individuals with a germline pathogenic variant in the CDKN2A gene [122]. In their study, 83% of the cancers were considered resectable at the time of imaging, and one-third were Stage I [122]. These and similar studies have shown that potentially curable cancers and high-risk precancerous lesions can be identified in highly-selected cohorts.

These screening programs, however, came at the cost of resecting a number of benign or low-risk lesions, thereby exposing those patients to an unnecessary major surgical procedure. Unnecessary surgery represents the most substantial documented harm of pancreatic cancer screening, and to date this harm has not been entirely preventable in practice. The Dbouk et al study, for example, reported that five patients underwent surgery for concerning imaging findings that turned out to be related to low-grade PanIN and intraductal papillary mucinous neoplasms. Although none of these patients had a significant surgical complication, the risk of low-grade dysplasia progressing to pancreatic cancer is quite low, so it’s debatable whether these five patients benefitted from having their pancreas resection. In the Klatte et al study, seven patients underwent pancreaticoduodenectomy or distal pancreatectomy for benign low-risk lesions. All of these patients were alive at follow up. Studies of low-risk pancreatic cyst surveillance have also reported significant rates of overtreatment. Tamburrino and colleagues found that in a cohort of 961 patients receiving surveillance for a cyst, 40 were surgically overtreated while only 16 were classified as correctly treated [227]. Although immediate surgical complications have been rare, these procedures can be life-altering for patients. Many develop diabetes, some develop difficult to control diabetes, and many require substantial recovery time [228]. Any widely-adopted screening program would need to minimize such adverse events.

As noted earlier, individuals with screen-detected pancreatic cancers live much longer than individuals whose cancers are detected because they developed symptoms [9]. However, because pancreatic cancer screening studies have not used randomized controlled methodology, positive results are subject to the possibility of lead-time and length-time biases potentially inflating the apparent survival benefit [229]. But the accumulating results of surveillance using imaging (EUS/MRI) techniques demonstrate that, in the right populations, the screening of asymptomatic high-risk individuals is possible and will detect some surgically resectable early stage cancers, likely leading to improved survival for those individuals.

18. Looking forward

Screening for pancreatic cancer presents enormous opportunities, and yet it also poses enormous challenges. Advanced pancreatic cancers are deadly, with little hope for cure. Yet high-grade precancerous lesions and low-stage invasive cancers can be cured. The practical challenges to finding and treating these curable lesions are substantial (Figure 3). Pancreatic cancer is a rare disease in the overall population, and the organ lies deep in the abdomen. False positive results from screening the general population will be too common and can produce real patient harm. The challenge, then, is to develop a screening program that minimizes patient harm, maximizes patient benefit, and does so at a reasonable cost to society [159].

Figure 3:

Barriers to screening for pancreatic cancer.

One can easily envision that in the not too distant future, AI-based algorithms will be integrated into electronic patient medical records, looking for family history, germline gene status, smoking, obesity, and new onset diabetes [230]. These systems could alert health care providers when a patient appears to be at risk for pancreatic cancer. In parallel, AI-based algorithms will be quietly running in the background of radiology practices, analyzing the many abdominal CT scans that are performed every day. These algorithms will look for subtle changes in the pancreas, unseen by the radiologist, which could alert practitioners of the patient’s cancer risk. In addition, advances in non-invasive DNA-based screening tests will improve on the sensitivity and specificity of existing ctDNA approaches, increasing the yield of low-stage cancers and decreasing false positives. New protein and glycoprotein-based markers of pancreatic cancer will also continue to be discovered, enhancing the arsenal of biomarkers available for screening programs. No single marker will be sufficient; instead, a combination of markers, used in the right setting, will provide the sensitivity and specificity needed to screen for pancreatic cancer in a cost-effective manner.

As we move forward, our zeal to help those at-risk for developing pancreatic cancer must always be balanced with an understanding of the costs of screening. These costs not only include financial burdens to the medical system and patients, but also the psychological and medical harms of false positive results. Surgical overtreatment of low-risk precursor lesions will especially need to be minimized. Exciting advances should also, whenever possible, be subjected to the rigor of randomized controlled trials to prove their benefit in real-world populations.

AI-based applications will present their own challenges. These include potential legal and privacy issues when using data obtained for other clinical indications. Patients should be given the opportunity to provide informed consent before being subjected to any screening program, especially one using a novel modality such as artificial intelligence. Patients and clinicians alike may feel distrust or discomfort when it comes to applying an opaque third-party algorithm to intimate personal health data. The use of ancillary AI testing could also increase patient costs in a fee-for-service payment system.

19. Conclusion

At first glance, pancreatic cancer seems to be an ideal target for screening and early detection. It is a highly lethal cancer, and most patients do not develop signs and symptoms until late in the course of their illness. A variety of methods have been developed to screen for early disease. From blood CA19–9 levels to circulating tumor DNA, each method has advantages and disadvantages, and it is likely that the best approach will incorporate multiple complementary modalities. It is also likely that artificial intelligence, whether applied to electronic patient medical records or to image analysis, will play a role in identifying the highest risk individuals. Screening for pancreatic cancer, however, must take into account the rarity of the disease in the general population, the costs of false positive test results, and the real potential for overtreatment of low-risk lesions.

20. Expert Opinion

If we wait until patients develop symptoms from pancreatic cancer, most will die within months of their diagnosis. There are, however, a number of potential approaches to the earlier diagnosis of pancreatic cancer and its precursors in asymptomatic individuals. The unresolved challenge is that none of the existing approaches have the sensitivity and specificity needed to effectively screen the general population. While combining tests can improve sensitivities, even these combinations fall short. Existing approaches to screening, therefore, have to be focused on populations at greatest risk of the disease. Towards this end, screening efforts to date have targeted individuals with a strong family history of pancreatic cancer or a known pathogenic germline variant that predisposes to pancreatic cancer. Other groups with a heightened cancer risk that have not yet been clinically targeted include cigarette smokers, obese individuals, and adults with new onset diabetes mellitus. Screening programs that target high-risk groups and utilize imaging modalities like magnetic resonance imaging and endoscopic ultrasound have generated promising results. These programs, however, have thus far been relatively expensive and limited in scope.

Artificial intelligence (AI) has the potential to be a significant advance. AI-based algorithms can be applied to medical records and to imaging. AI-based algorithms can be trained to scour electronic patient medical records and imaging scans for subtle patterns, which may not be recognizable to the human eye, yet that indicate an increased risk of pancreatic cancer. Clinicians could then be alerted and even guided on the best ways to evaluate their patients further. AI-based algorithms have already been built that can evaluate computed tomography (CT) and other scans for subtle changes caused by low-stage, curable, pancreatic neoplasms. These algorithms could quietly run in the background, evaluate the millions of CT and other scans that are already being performed for other indications, and have the potential to identify changes in the pancreas that radiologists could easily miss in their day to day practice. Again, the program could alert the radiologist to look more closely at the images of the pancreas, and even suggest the best next steps to evaluate the subtle AI-detected changes. This approach has the advantage of not requiring additional testing until an abnormality is discovered. But enthusiasm about AI technologies must be tempered by the yet unproven real-world benefit and privacy concerns of these tools.

The growing emphasis on early detection of pancreatic cancer and its precursors should be balanced with a consideration of the costs and harms of screening. Costs are incurred during acquisition of the screening cohort, initial testing, confirmatory imaging and biopsies, surgery and other treatments, and post-treatment follow up and care. Screening can also cause serious patient harm. This harm includes the psychological stresses that individuals will feel when they are told that they are at increased risk, the worries from an initial result, and the medical complications that can occur from unnecessary biopsies and surgeries. The pancreas lies deep in the abdomen, and surgical resection of the head of the gland (the pancreatoduodenectomy) is associated with a 1–2% operative mortality rate and significant morbidity. If this surgery is done for a false positive screening test or low-risk precursor lesion, we have done significant harm to the individual.

As we look forward to the next five to ten years, we hope that new technical advances will improve the sensitivity and specificity of screening for pancreatic cancer. We also anticipate a better understanding of who is at risk, and of which groups will benefit most from screening. The development of new AI-based tools might improve screening approaches and help refine the identification of at-risk populations. The challenge, then, will be to fulfill these goals in a cost-effective manner that minimizes patient harm and to prove, prospectively and rigorously, the real-world value of these interventions before widespread implementation.

Article highlights:

• Most patients with pancreatic cancer present with advanced, incurable disease.

• Early detection and treatment offer the best hope for cure.

• A variety of approaches, from novel blood tests to artificial intelligence driven algorithms to interpret CAT scans, offer hope for improved early detection.

• All early detection approaches, however, come with added costs and potential harms from false positives.

• New technical advances are needed to improve the sensitivity and specificity of screening for pancreatic cancer.