AI and Epigenetics: A New Opportunity for Early Pancreatic Cancer Detection in Australia

Abstract

Pancreatic ductal adenocarcinoma (PDAC) is now the third leading cause of cancer-related death in Australia and is poised to overtake colorectal cancer mortality globally [1, 2]. Despite decades of clinical trials, five-year survival has stagnated at approximately 14% [3]—a failure largely attributed to PDAC's typically late-stage presentation, early systemic spread, cold immunogenicity, and historically undruggable KRAS mutations [4]. The primary risk factors for pancreatic cancer include tobacco use, chronic pancreatitis, and obesity, alongside non-modifiable factors such as older age, a family history of the disease, and inherited genetic syndromes. Despite recent advances in KRAS targeted therapies [5], early diagnosis and surgical resection remain the only viable option of cure for PDAC. Of the five most common causes of cancer-related death in Australia, four have some form of screening and early detection available in Australia. Breast and colorectal cancer have mature screening programs; prostate cancer screening in the form of serum PSA (although controversial) is easily available, and low-dose chest CT in high-risk patients has become part of the National Lung Cancer Screening Program (NLCSP) as of July 2025. Currently, there is no established screening tool for PDAC in the asymptomatic population. The recently commenced Australian Pancreatic Cancer High-Risk Screening Program (APRISE) is a non-randomized trial of screening high-risk (with familial or genetic history) patients to undergo endoscopic ultrasound (EUS) and MRI imaging. While APRISE is a critical step forward, these investigation modalities are resource-intensive and require specialized gastroenterological and radiological expertise found primarily in metropolitan centers. Our proposed strategy seeks to augment this framework by providing a cost-effective, less invasive entry point that could eventually transition high-risk surveillance to the primary care setting.

The introduction of various artificial intelligence (AI) driven imaging assessments has potentially unearthed a cost-effective modality of pancreatic cancer screening. Numerous published AI models, such as the PANDA model, have detected pancreatic lesions on CT imaging with 92.9% sensitivity and 99.9% specificity [6, 7]. The multi-center PANORAMA study has recently shown AI systems as superior to standard radiological assessment of pancreatic cancer (AUROC of 0.92 compared to 0.88 for radiologists) [8]. Given the overlap in risk factors between pancreatic cancer and lung cancer, could the non-contrast imaging in the NLCSP be extended to the abdomen to perform AI-based screening for pancreatic cancer? However, given the taxonomy of harms of screening (physical effects, psychological effects, financial strain, and opportunity costs [9]) for PDAC, an additional simple ancillary test with high specificity may be required.

While CA19-9 is a validated tool for monitoring treatment response and disease progression, it has little utility as a primary screening tool; up to 22% of certain ethnic groups are Lewis-antigen negative non-secretors, and the marker frequently yields false positives in benign biliary disease. Conversely, 5-hydroxymethylcytosine (5hmC) signatures in cell-free DNA offer a stable, tissue-specific ‘fingerprint’ of early tumorigenesis with reported specificities exceeding 95% [10, 11], and are already available in Australia [12].

During early tumor development, 5hmC abundance typically decreases globally across the tumor genome while the remaining 5hmC regions undergo extensive redistribution. These distinct patterns of hydroxymethylation remain stable throughout disease progression. As 5hmC profiles also reflect the identity of the tissue they originated from, identifying these patterns in cell-free DNA (cfDNA) circulating in the blood allows for the early, non-invasive detection of cancer and the prediction of the tumor's tissue of origin.

Analytical validation of a 5hmC-based test demonstrated 68.3% sensitivity for early-stage (I/II) PDAC and 96.9% specificity [13], with similar tests such as Avantect and CancerGuard^R detecting early pancreatic cancer with greater than 95% specificity [14]. Additionally, a novel four-biomarker blood panel combining aminopeptidase N, polymeric immunoglobulin receptor, CA19-9, and thrombospondin-2 has demonstrated greater than 87% sensitivity and 95% specificity for early-stage PDAC in two independent validation cohorts [15].

A pragmatic opportunity for early PDAC detection lies in tandem with the National Lung Cancer Screening Program (NLCSP). The NLCSP provides low-dose CT (LDCT) scans every 2 years to high-risk individuals, primarily long-term smokers aged 50–70. Given the significant overlap in risk factors—most notably tobacco use and age—this cohort is already at an elevated risk for pancreatic malignancy. By extending the field of view of these existing scans to include the upper abdomen, AI-driven diagnostic tools could screen a population at high risk for pancreatic cancer, without significantly increased radiation or clinical burden. Integrating PDAC screening into this established infrastructure would leverage existing radiology workflows and maximize the clinical yield of the $264 million investment already committed to the NLCSP. Combining this with serum 5hmC based measurements could potentially increase the proportion of patients eligible for curative surgical resection, which remains the only cost-effective pathway to long-term survival.

Geographical disparities remain significant barriers to healthcare access in Australia. Australians in rural and remote areas often face prohibitive travel requirements and out-of-pocket expenses for specialized imaging like MRI or EUS. A blood-based screening tool using 5hmC technology would be inherently more accessible, in line with the National Cancer Plan 2023–2033. The 2019 USPSTF recommendation against pancreatic cancer screening in asymptomatic adults was a pragmatic response to the tools of its time, citing a lack of high-accuracy tests and the significant harms of invasive follow-up [16]. However, emerging AI-enabled image assessment and blood-based molecular tests has led to a renewed focus on the risk–benefit balance for targeted PDAC screening. With AI-driven CT assessments and 5hmC signatures offering non-invasive, high-specificity validation, the taxonomy of harms of screening has changed. Australia may be well positioned to evaluate this approach through existing screening infrastructure such as the NLCSP. Before any national rollout, a targeted pilot study should evaluate diagnostic performance, investigation burden and harms, cost-utility, implementation feasibility, equity of access, and the ethical and privacy implications of AI-enabled image assessment and proprietary blood-based technologies. We must move beyond the elusive goal of early diagnosis and embrace a multimodal strategy that offers PDAC patients their only realistic chance at a cure.

K.K. is the recipient of the Royal Australasian College of Surgeons John Loewenthal Grant, the Medtronic Pancreatic Surgery Fellowship, and the Australian Government Research Training Program offset.

All authors are in agreement with the content of the manuscript. The manuscript has not been published previously and is not under consideration elsewhere.

The authors declare no conflicts of interest.

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.