Prediction of Pancreatic Cancer Risk in Patients with New-Onset Diabetes

In the USA, pancreatic cancer was the fourth leading cause of cancer-related deaths in 2024 despite its absence among the top 10 most diagnosed cancers this year [1]. This pattern has also been observed globally [2] and reflects the poor prognosis of pancreatic cancer at diagnosis. This is due in part to the fact that pancreatic cancer is diagnosed at a late stage: the survival rate for stage 4 cancer is 3.1% relative to 44% for localized disease [1]. A shift toward detecting pancreatic cancer at earlier stages would double survival rates [3]. However, there are technical hurdles to achieving this. The subtle presentation of early-stage cancer and its low prevalence in the general population would increase morbidity from unnecessary testing of this group [4]. However, certain high-risk groups do benefit from screening, including those with genetic mutations, family history of pancreatic cancer, or certain syndromes with a greater than 5% lifetime risk of pancreatic cancer [5]. Additionally, patients with intraductal papillary mucinous neoplasms have a lifetime risk of greater than 6% [6].

Approximately 90% of patients with pancreatic cancer do not harbor a genetic risk [5]. Among these are patients with diabetes, who are at an increased risk of pancreatic cancer relative to the general population [7]. In this group, a distinction is made between new-onset diabetes, diabetes that is diagnosed within 3 years of pancreatic cancer onset, and long-standing diabetes, diagnosed greater than 3 years from pancreatic cancer onset. While risk is elevated in both groups, it is markedly elevated in the former group compared to the latter [8, 9]. Long-standing diabetes doubles the risk of pancreatic cancer, new-onset diabetes triples the risk relative to the general population, and diabetes diagnosed within 1 year increases the risk sevenfold [10, 11]. These distinct risk profiles reflect the different processes underlying type 2 diabetes and pancreatic cancer-associated diabetes. Type 2 diabetes is a progressive condition in which obesity and physical inactivity lead to increasing insulin resistance and eventual beta cell dysfunction – a gradual process that typically occurs over 10 years or longer [12]. In contrast, pancreatic cancer-associated diabetes is a paraneoplastic phenomenon caused by the cancer itself: the tumor induces changes that abruptly disrupt blood glucose control. These changes often precede the visible onset of pancreatic cancer by 24–36 months [8], and this type of diabetes frequently resolves after tumor removal [13]. Therefore, when a patient with normal blood glucose levels suddenly develops diabetes without the typical years of prediabetes, pancreatic cancer should be considered as a possible cause. Additionally, this metabolic destabilization can manifest in patients with existing diabetes or prediabetes as a sudden deterioration in blood glucose control [14‒16].

New-onset diabetes, even without a history of prediabetes, is almost always type 2 diabetes. However, there are several differences between new-onset diabetes caused by type 2 diabetes and new-onset diabetes associated with pancreatic cancer:

• 1.

  Weight changes: Type 2 diabetes is associated with weight gain, while pancreatic cancer-related diabetes is associated with weight loss [10, 11, 17].

• 2.

  Age of onset: Pancreatic cancer-related diabetes tends to occur at an older age compared to type 2 diabetes. The time between the onset of diabetes and the diagnosis of pancreatic cancer is also shorter in these cases [18, 19].

• 3.

  Risk factors: Patients with pancreatic cancer-related diabetes have higher rates of smoking, gallstone disease, and chronic pancreatitis and are less likely to be obese [18].

• 4.

  Gender differences: While there may be subtle differences in the risk between genders, they are insignificant, and the risk is elevated in both men and women [20, 21].

• 5.

  Lipid and adipose tissue changes: Pancreatic cancer-related diabetes is associated with decreases in serum lipids and adipose tissue [22‒24], whereas type 2 diabetes is associated with hyperlipidemia.

Long-standing diabetes promotes the development of pancreatic cancer in several ways. Long-standing diabetes is associated with hyperinsulinemia: hyperinsulinemia induces increased secretion of digestive enzymes that promote local inflammation [25]. Elevated insulin levels can lead to overstimulation of the PI3K/AKT and RAS/MAPK pathways that promote cancer cell proliferation and increased protein synthesis [26‒28]. This overstimulation can also be mediated through interactions between insulin and insulin-like growth factors on the same cellular pathways [27]. These effects from hyperinsulinemia on pancreatic cancer risk have been shown in nondiabetic populations, which suggest that hyperinsulinemia, independent of hyperglycemia, confers pancreatic cancer risk [28, 29]. Hyperglycemia may contribute to the development of pancreatic cancer through overstimulation of the STAT3-MYC pathway [30], and inducing mutations of KRAS, the leading oncogenic mutation seen in 90% of pancreatic cancers [31]. Hyperglycemia can increase intracellular levels of TGF-B1, a cytokine that promotes the epithelial-mesenchymal transition and selective destruction of beta cells [32, 33].

While long-standing diabetes can induce pancreatic cancer, pancreatic cancer in turn can induce glycemic dysfunction through several mechanisms. One mechanism is through peripheral insulin resistance: skeletal muscle has been shown resistant to insulin in pancreatic cancer but not other cancers [34]. Another mechanism is through destruction of beta cells which release insulin. Pancreatic cancer exosomes, extracellular vesicles secreted by malignant pancreas cells, have been shown to express the proteins adrenomedullin, vanin-1, and the microRNA mir-19a, all of which mediate beta-cell dysfunction [35‒37]. Adrenomedullin may also mediate lipolysis seen in pancreatic cancer [38]. Recently, serpin E1 and CCL5 were identified as two inflammatory cytokines that mediate beta-cell dysfunction in pancreatic cancer-related diabetes [39]. Finally, pancreatic cancer can induce diabetes through disruption of GLP-1 and GIP, two hormones that work synergistically with insulin [40].

Antidiabetic medication initiation (including insulin and metformin) may show elevated short-term pancreatic cancer risk due to reverse causation, as worsening glycemic control can be an early cancer manifestation [14]. Similarly, starting anticoagulation or stopping hypertensive medications shows elevated short-term cancer risk, reflecting cancer-induced hypercoagulability and weight loss [41, 42].

Metformin reduces long-term pancreatic cancer risk and improves survival through AMPK-mediated mTOR pathway suppression [43‒47]. Insulin’s long-term effects remain unclear due to confounding by advanced age (itself a risk factor) and concurrent medication use [48‒50]. Long-term use of sulfonylureas may increase pancreatic cancer risk, particularly in Asian populations, though interpretation is limited by the concurrent use of multiple anti-glycemic agents [49, 50]. Initial concerns about incretin-based therapies increasing cancer risk have been subsequently attributed to reverse causation in more recent studies [51, 52].

Previous studies have established that screening for new-onset diabetes may be cost-effective at less than USD 100,000 per quality adjusted life year when two criteria are met: the 3-year incidence of pancreatic cancer is at least 2%, and a minimum of 25–26% of screened cases are localized [53, 54]. Although these studies were criticized for using stage distribution data derived from patients with genetic risk factors [55] and employed different imaging modalities for screening, they nonetheless provide an important benchmark for determining the cost-effectiveness of cancer screening. Notably, this threshold is lower than the suggested 10% lifetime risk cutoff used for screening patients with genetic predisposition to pancreatic cancer. This difference is justified because the risk of cancer in new-onset diabetes is concentrated within 3 years of diabetes onset, as opposed to the lifetime risk associated with genetic predisposition [56]. It is estimated that ∼0.8% of those with new-onset diabetes will develop pancreatic cancer, and of the cases detected, only 16% can be resected [57, 58]. These findings fall below the benchmarks of cost-effectiveness.

Several clinical models have been developed to identify new-onset diabetes patients at higher risk of pancreatic cancer, with the 2018 END-PAC model being particularly significant [59]. This model uses age, weight change, and blood sugar change to assess cancer risk, but while initial results were promising, three validation studies showed lower performance [60‒62]. The model’s strict criteria for defining new-onset diabetes limited patient inclusion [60‒62], and its development in a predominantly Caucasian population led to variable performance across ethnicities [60]. Despite concerns about false positives and cost-effectiveness [60‒62], a recent meta-analysis confirmed the model's utility, showing 56% sensitivity and 82% specificity with a score threshold of 3 or higher [63]. In contrast, Boursi and colleagues [64] had developed a more complex model in 2017 using UK healthcare data, incorporating 11 variables including laboratory values (HbA1c, hemoglobin, cholesterol, creatinine, alkaline phosphatase), medications, and demographic information. This comprehensive approach achieved stronger performance metrics (AUC 0.82, 44.7% sensitivity, 94% specificity at 1% risk threshold), later validated in a US population [65]. While Boursi's model showed superior discrimination compared to the END-PAC model, both approaches face limitations in real-world implementation due to their inability to account for missing data [66, 67]. Boursi and colleagues also developed a variant model for prediabetic patients, substituting body mass index change with alanine aminotransferase as an input [68].

Recent prediction models for pancreatic cancer in new-onset diabetes have demonstrated various innovative approaches [69‒73]. Chen et al. [69] evaluated three methods of incorporating HbA1c data: change in HbA1c, rate of change, and direct value incorporation, all proving equally effective. Cichosz et al. [70] developed a model focused exclusively on laboratory values while deliberately excluding subjective measurements like weight, achieving an AUC of 0.78. Chen et al. [71] compared machine learning algorithms using diverse clinical inputs including laboratory values, diagnoses, and medications, achieving an AUC of 0.91 with linear discriminant analysis. Ali et al. [72] developed a women-specific model using only medication data, finding that rapid progression to insulin therapy indicated increased cancer risk, confirming previous findings [14]. Clift et al. [73] created the most comprehensively validated model, incorporating symptoms, laboratory values, and clinical data, using decision curve analysis to demonstrate clinical utility. All models share three key features: age as a primary predictor, incorporation of diabetes progression (through HbA1c changes, biochemical trajectories, or medication changes), and 2- to 4-year prediction windows. While studies have proposed incorporating diverse data types like genetics and microbiome data for pancreatic cancer prediction [74], recent research demonstrates that genetic markers alone outperform clinical variables. However, the integration of both genetic and clinical information yielded the most effective predictive model [58]. A notable distinction among the aforementioned models is their definition of diabetes onset: using glycemic criteria [59, 68], diagnostic codes [64, 70, 73], first antidiabetic medication [71, 72], or prediabetic markers [58, 68]. This variation may impact cancer detection given the narrow window between diabetes onset and cancer manifestation [75].

Three notable models have been developed for pancreatic cancer screening in the general population. The Danish study [76] analyzed 6.2 million patient records and uniquely incorporated temporal patterns in disease codes, achieving an AUC of 0.88 in Danish cohorts but dropping to 0.71 in US validation. The Prism model [77] innovatively examined data from 6 to 18 months before cancer diagnosis, matching cases with controls to predict near-term cancer risk. Chen et al. [78] developed an approach using just five predictors (age, abdominal pain, weight change, HbA1C, and ALT change), achieving comparable results on validation (AUC = 0.77 internal, 0.71 external validation). All models incorporated physical symptoms like abdominal pain and jaundice, which proved highly predictive of cancer within 6 months. However, for longer-term prediction, these symptoms became less important, while factors like diabetes and gallbladder disease gained significance. Chen’s early detection model, which targeted cancer cases diagnosed 3 months after diabetes onset, showed reduced performance excluding abdominal pain. This pattern across models suggests that they may be more sensitive to metastatic disease, as physical symptoms typically indicate advanced rather than early-stage pancreatic cancer.

Biomarkers currently serve no screening role in pancreatic cancer detection. The primary biomarker, CA 19-9, while valuable for surveillance, faces significant limitations. These include genetic expression constraints affecting 5–10% of the population and nonspecific elevation in various conditions like jaundice and pancreatitis, resulting in an inadequate positive predictive value of <1% in general population screening [79, 80]. Recent research demonstrates enhanced utility of CA 19-9 through strategic combinations. In new-onset diabetes patients, the combination of elevated CA 19-9 (>37 IU/mL) with high bilirubin levels (>1.7 mg/dL) achieved 74% predictive accuracy for pancreatic cancer [81]. Further improvements emerged through metabolomic analysis, particularly for patients with chronic pancreatitis and new-onset diabetes [82, 83].

Additional biomarker candidates show promise through distinct mechanisms. Adiponectin levels, typically suppressed in type 2 diabetes due to increased visceral fat, are characteristically elevated in pancreatic cancer due to fat loss [23, 84, 85]. Leptin demonstrates sex-specific utility, offering predictive value specifically in male patients [86]. The hormonal disruption pattern extends to pancreatic polypeptide, showing blunted response [87], and altered glucagon/insulin ratios resulting from selective islet cell destruction [88]. Emerging candidates include thrombospondin-1, osteoprotegerin, and S100A8 N-terminal peptide [89].

Despite well-established links between new-onset diabetes and pancreatic cancer, current prediction models and biomarkers lack prospective validation for screening in real-world settings. Nevertheless, the poor prognosis of pancreatic cancer warrants vigilance, particularly in patients presenting with concerning features. These include new-onset diabetes in older patients (>50 years) without obesity, especially when accompanied by unexplained weight loss, rapid progression to insulin requirement, or deteriorating glycemic control despite appropriate therapy. Other concerning patterns include decreasing serum lipids and the absence of traditional metabolic syndrome features. For these high-risk patients, baseline laboratory evaluation and imaging may be warranted, with subsequent monitoring guided by clinical findings rather than fixed intervals. Until prediction models and biomarkers are validated through prospective studies, clinicians must rely on careful clinical judgment while being mindful of healthcare resources.

Salman Khan has no conflicts of interest to disclose.

This study was not supported by any sponsor or funder.

Salman Khan wrote the manuscript in its entirety.