New-onset diabetes and pancreatic cancer: Current evidence on absolute risk and screening feasibility

Abstract

Pancreatic ductal adenocarcinoma (PDAC) remains a highly lethal malignancy largely due to delayed diagnosis, prompting growing interest in early clinical signals that may facilitate earlier detection. New-onset diabetes (NOD) has emerged as one such signal, supported by accumulating epidemiologic evidence demonstrating a temporal association between diabetes onset and subsequent PDAC diagnosis. Recent large population-based studies have confirmed an increased relative risk of PDAC among individuals with glycemically defined NOD, particularly within the first few years after diabetes onset. However, despite this association, the absolute incidence of PDAC in NOD populations remains low, raising critical questions regarding the feasibility and clinical justification of population-wide screening strategies. This discrepancy between relative risk elevation and absolute event probability represents a central challenge in translating epidemiologic findings into actionable clinical practice. Risk stratification models, such as the Enriching NOD for Pancreatic Cancer score, have been proposed to address this gap, yet their real-world performance highlights persistent limitations when applied as stand-alone tools for diagnostic decision-making. This minireview synthesizes recent epidemiologic, modeling, and screening-related evidence to clarify the clinical role of NOD in PDAC detection. Current evidence supports viewing NOD as a gatekeeper for stepwise risk enrichment rather than a screening indication and favors risk-enriched, stepwise diagnostic strategies, with future work focused on prospective validation, integrated biomarkers, and actionable absolute-risk thresholds.

Key Words: Pancreatic cancer; New-onset diabetes; Risk stratification; Absolute risk; Early detection

Core Tip: New-onset diabetes (NOD) is a reproducible clinical signal associated with pancreatic ductal adenocarcinoma, but its absolute cancer risk remains low. This minireview summarizes recent prospective cohort data, risk stratification models, and screening-related evidence showing why NOD should not be used as a screening indication. Instead, it should be considered a gatekeeper for stepwise risk enrichment, with future efforts focused on clinically actionable absolute-risk thresholds, biomarkers, and prospective validation.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) remains one of the most lethal malignancies, with an overall 5-year survival rate below 10%, largely because most patients are diagnosed at an advanced stage when curative treatment is no longer feasible[1,2]. Early-stage PDAC is associated with substantially improved survival following surgical resection, underscoring the clinical importance of earlier detection strategies. However, population-wide screening of asymptomatic individuals is not currently recommended due to the low incidence of PDAC and the absence of validated, cost-effective screening modalities[2]. Accordingly, early detection efforts have shifted toward clinically identifiable groups with enriched risk[3,4].

New-onset diabetes (NOD) has emerged as a potential clinical signal for occult PDAC. In this review, glycemically defined NOD refers to newly detected diabetes based on standard glycemic criteria, including hemoglobin A1c ≥ 6.5%, fasting plasma glucose ≥ 126 mg/dL (7.0 mmol/L), 2-hour plasma glucose ≥ 200 mg/dL during an oral glucose tolerance test, or random plasma glucose ≥ 200 mg/dL in the presence of symptoms, in an individual without previously known diabetes[5]. In the context of pancreatic cancer detection, the clinically relevant window is usually recent-onset diabetes, commonly within 1-3 years, particularly in adults aged ≥ 50 years[6].

Epidemiologic studies have consistently demonstrated an increased relative risk of PDAC among individuals diagnosed with diabetes shortly before cancer detection[7-9], supporting the hypothesis that diabetes may represent a paraneoplastic manifestation of early tumor development rather than a conventional metabolic disorder. These observations have positioned NOD as an attractive target population for risk-based early detection efforts. Biologically, NOD in the setting of PDAC may represent more than coincidental type 2 diabetes. Unlike conventional long-standing type 2 diabetes, cancer-associated NOD can arise as a paraneoplastic metabolic state characterized by rapid glycemic deterioration, weight loss, peripheral insulin resistance, and β-cell dysfunction. Tumor-derived mediators, including adrenomedullin-containing exosomes, have been implicated in impaired insulin secretion and pancreatic cancer-associated diabetes[10]. More broadly, diabetes has been linked to gastrointestinal malignancies beyond PDAC; for example, the association between type 2 diabetes mellitus and hepatocellular carcinoma has been reviewed in relation to chronic hyperinsulinemia, inflammation, and metabolic reprogramming[11]. However, the NOD-PDAC relationship is clinically distinct because diabetes may be an early manifestation of occult pancreatic cancer rather than only a long-term carcinogenic risk factor.

Nevertheless, important barriers limit the direct translation of NOD-associated risk into screening practice. Although relative risk is elevated, the absolute incidence of PDAC among individuals with NOD remains low, generally well below 1% over several years of follow-up, rendering population-wide diagnostic evaluation inefficient and potentially harmful[7,12]. Recent large population-based analyses using glycemically defined NOD further reinforce this discrepancy between relative and absolute risk, highlighting the need for cautious interpretation when considering screening strategies[12]. Collectively, these findings suggest that while NOD is a valuable epidemiologic marker, it is insufficient as a standalone indication for PDAC screening in routine clinical practice.

The aim of this minireview is to synthesize recent epidemiologic studies, risk stratification models, and screening-related evidence linking NOD and PDAC, with particular emphasis on absolute risk and clinical feasibility. Rather than treating NOD as an isolated screening indication, this review clarifies what the current evidence supports in practice: Recognition of NOD as a clinically identifiable entry signal, followed by stepwise risk enrichment using clinical features, prediction models, and biomarkers before targeted diagnostic evaluation is considered.

EPIDEMIOLOGIC ASSOCIATION BETWEEN NOD AND PDAC

Epidemiologic evidence linking NOD to PDAC has accumulated over the past two decades, consistently demonstrating a temporal relationship between diabetes onset and subsequent cancer diagnosis. In a population-based study of 2122 Rochester, Minnesota residents aged 50 years or older who first met standardized criteria for diabetes, Chari et al[7] found that 18 individuals (0.85%) were diagnosed with pancreatic cancer within 3 years, corresponding to an observed-to-expected ratio of 7.94 compared with the general population. These findings established that NOD identifies a population with markedly enriched relative risk, while also showing that the absolute probability of cancer remains approximately 1%.

The temporal pattern of diabetes in PDAC further supports the concept that hyperglycemia may, in some patients, represent a manifestation of occult cancer rather than conventional type 2 diabetes alone. In a case-control study evaluating fasting glucose patterns before cancer diagnosis, Chari et al[8] reported that diabetes was more frequent in patients with pancreatic cancer than in controls and was more often new onset among cancer cases. Meta-analytic evidence has also shown that pancreatic cancer risk is highest closer to the time of diabetes diagnosis and declines with longer diabetes duration, consistent with a bidirectional relationship in which diabetes may act both as a risk factor and as an early manifestation of PDAC[9].

More recent prospective validation has strengthened this epidemiologic signal. In a prospective cohort of 18838 adults aged 50 years or older with glycemically defined NOD, Chari et al[12] diagnosed 82 pancreatic cancers during follow-up. The overall race-adjusted 3-year incidence was 0.62%, and the mean interval between glycemically defined NOD and clinical cancer diagnosis was 8 months[12]. Importantly, while standardized incidence ratios were elevated across racial and ethnic groups, the absolute incidence of PDAC remained low in real-world settings, reinforcing the discrepancy between epidemiologic signal strength and clinical actionability[12].

Collectively, these studies establish NOD as a reproducible epidemiologic marker associated with PDAC, characterized by a strong temporal association but modest absolute event rates. To contextualize these observations, Table 1 summarizes pancreatic cancer risk across the general population, glycemically defined NOD, and progressively enriched high-risk subgroups. This comparison highlights the discrepancy between relative risk enrichment and low absolute event rates, and underscores the rationale for stepwise risk stratification rather than indiscriminate diagnostic evaluation. This duality has important implications: NOD identifies a population enriched for PDAC risk compared with the general population, yet the majority of individuals with NOD will not develop cancer. Consequently, epidemiologic association alone is insufficient to justify routine screening and instead necessitates additional risk stratification to identify subgroups in whom diagnostic evaluation may offer a favorable balance of benefit and harm[7-9,12]. Key studies informing this interpretation are summarized in Table 2.

Table 1 Contextualizing pancreatic cancer risk across screening-relevant populations.

Population	Approximate risk context	Implication for screening
General population	Very low background incidence; population-level screening is not justified	No routine screening
Adults with glycemically defined NOD	Relative risk is increased, but absolute PDAC risk remains low, approximately around 1% over 3 years in older adults with NOD	Indiscriminate imaging of all NOD patients is inefficient
Enriched NOD subgroup	Risk increases with older age, weight loss, rapid glycemic worsening, or high END-PAC score	Candidate group for selective imaging or biomarker-based enrichment
Very high-risk clinical subgroup	NOD plus symptoms, abnormal biomarkers/imaging, hereditary risk, or strong clinical concern	Diagnostic evaluation rather than screening

END-PAC: Enriching new-onset diabetes for pancreatic cancer; NOD: New-onset diabetes; PDAC: Pancreatic ductal adenocarcinoma.

Table 2 Key studies informing the clinical interpretation of new-onset diabetes in pancreatic cancer detection.

Ref.	Design and population	Key numerical findings	Main implication
Sharma et al[6]	Development and validation study of adults older than 50 years with glycemically defined NOD	Approximately 1% developed PDAC within 3 years; END-PAC score ≥ 3 yielded 78% sensitivity, 85% specificity, and 3.6% PDAC prevalence in the high-risk subgroup	NOD enriches risk, and simple clinical variables can improve enrichment
Chari et al[12]	Prospective cohort of 18838 adults aged 50 years or older with glycemically defined NOD	Eighty-two PDACs diagnosed; race-adjusted 3-year incidence 0.62%; mean lead time 8 months	Prospective validation confirms enrichment but also underscores low absolute risk
Mellenthin et al[16]	United Kingdom primary care cohort of 197092 patients with NOD	END-PAC AUC 0.69 after imputation and 0.71 after recalibration; stand-alone use judged insufficient for diagnostic workup selection	Real-world implementation attenuates performance of clinical scores
Cichosz et al[17]	Danish registry-based machine-learning model using routine biochemical trajectories	AUC 0.78; top 1% risk stratum had 12% 3-year PDAC risk, with 20% sensitivity	Longitudinal routine data can identify a very high-risk subgroup, but sensitivity remains limited
Khan and Bhushan[18]	Multisystem United States cohort using XGBoost	AUC 0.80; positive predictive value 12% at the Youden cutoff and ≥ 2.5% when sensitivity fell to 38%	Model discrimination can improve, but clinically acceptable positive predictive value remains difficult

NOD: New-onset diabetes; PDAC: Pancreatic ductal adenocarcinoma; END-PAC: Enriching new-onset diabetes for pancreatic cancer; AUC: Area under the curve.

ABSOLUTE RISK AND THE LIMITS OF SCREENING FEASIBILITY

Clinical translation requires more than a statistically significant association; it requires an absolute event rate high enough to justify downstream testing. Relative risk describes how much more frequently PDAC occurs in one group than in a reference group, whereas absolute risk describes the actual probability that an individual will be diagnosed with PDAC over a defined time interval.

For screening-related decisions, absolute risk is the more actionable measure because it determines expected yield, the approximate number of patients who must undergo evaluation to detect one cancer, and the burden of downstream harms. In NOD, these measures diverge in clinically important ways. Although PDAC risk is elevated relative to the general population[7], the 3-year absolute incidence remains only 0.62% in prospective glycemically defined NOD cohorts[12] and approximately 1% in earlier population-based studies of adults aged 50 years or older[7]. In practical terms, these values imply that approximately 100-160 unselected patients with NOD would need to undergo initial evaluation to identify one pancreatic cancer, before accounting for false-positive findings, incidental lesions, repeat testing, or invasive follow-up procedures.

Absolute risk also determines screening yield and positive predictive value. Even when a test has acceptable sensitivity and specificity, applying it to a low-pretest-probability population produces a low positive predictive value, meaning that many positive or indeterminate findings will not represent clinically relevant PDAC. This creates downstream harms, including additional imaging, detection of incidental pancreatic or extrapancreatic lesions, psychological burden, endoscopic procedures, and potential overtreatment with limited overall yield[2]. These concerns are consistent with current recommendations: The United States Preventive Services Task Force recommends against screening asymptomatic adults for pancreatic cancer, citing low incidence, uncertain accuracy of candidate screening tests, and at least moderate harms from false-positive results and treatment of screen-detected disease[3]. The International Cancer of the Pancreas Screening Consortium restricts surveillance to selected individuals with familial or genetic susceptibility, ideally in expert centers and research-oriented settings[4]. This concern is further illustrated by high-risk surveillance cohorts, in which a meta-analysis estimated that 135 individuals at hereditary or familial high risk must be screened to identify one high-risk pancreatic lesion. If screening yield is modest even in such enriched populations, unselected evaluation of all patients with NOD is unlikely to be clinically efficient[13].

Accordingly, treating NOD as a screening indication would likely trigger large numbers of low-yield imaging evaluations, false-positive findings, and downstream harm without meaningful improvement in cancer detection yield. Its most defensible clinical role is as an entry point for further risk refinement. Any attempt to translate NOD into early detection practice must therefore emphasize absolute risk, not relative risk alone, when balancing benefit, harm, and resource utilization. A practical challenge is defining when enriched absolute risk becomes clinically actionable, because no universally accepted threshold exists for pancreatic imaging in asymptomatic NOD. Conceptually, a predicted 3-year PDAC risk of approximately 1% has been used in risk-modeling studies as a potential threshold for selective screening, because it narrows the population requiring imaging while preserving the opportunity for early detection[14]. In practice, the threshold should depend on the intended intervention: Non-invasive cross-sectional imaging may be reasonable at lower enriched-risk thresholds, whereas invasive testing such as endoscopic ultrasound should generally be reserved for higher-risk strata or for patients with additional clinical warning features.

RISK STRATIFICATION BEYOND NOD: FROM CLINICAL VARIABLES TO PREDICTIVE MODELS

As NOD is heterogeneous, current efforts have focused on identifying modifiers that distinguish pancreatic cancer-associated diabetes from more common type 2 diabetes. A systematic review and meta-analysis of 22 studies including 576210 patients with NOD found that older age, family history of PDAC, pancreatitis or gallstone-related disease, weight loss, and rapidly increasing glycemia were the most informative enrichers[15]. These features provide the biological and clinical rationale for multivariable risk modeling.

The most widely studied clinical model is Enriching NOD for Pancreatic Cancer (END-PAC). Clinically, the END-PAC model is useful because it distinguishes the phenotype of ordinary type 2 diabetes from that of pancreatic cancer-associated diabetes. The score integrates age at diabetes onset, change in body weight, and change in blood glucose[6]. In this context, Δblood glucose refers to the trajectory of glycemic worsening-namely, the change in blood glucose from the preceding year to diabetes onset-rather than a single static glucose value. Older age, weight loss, and rapid glycemic deterioration increase the likelihood that NOD represents a cancer-associated metabolic phenotype.

In the original validation cohort, an END-PAC score of at least 3 identified 7 of 9 pancreatic cancers with 78% sensitivity and 85% specificity, increasing cancer prevalence to 3.6% in the high-score subgroup[6]. This represented a meaningful enrichment over the unselected NOD population and established the principle that NOD can be subdivided into clinically distinct risk strata.

However, real-world implementation has exposed important limitations. In a British primary care cohort of 197092 patients with NOD, Mellenthin et al[16] found that complete data required to calculate the original END-PAC score were available in only 9.2% of patients. After imputation, the area under the receiver operating characteristic curve was 0.69 and improved only to 0.71 after recalibration. The authors concluded that END-PAC alone remained insufficient to select patients for diagnostic workup[16]. These findings are important because they shift the discussion from theoretical enrichment to practical deployment in routine care.

Contemporary machine-learning models extend the same concept by using longitudinal clinical and biochemical trajectories. In a Danish population-based study, a random forest model based on age, sex, and routine biochemical trajectories achieved an area under the curve of 0.78; the top 1% of predicted risk had a 3-year pancreatic cancer risk of 12%, but sensitivity was only 20%[17]. In a multisystem United States cohort, an XGBoost model achieved an area under the curve of 0.80, but the positive predictive value was only 1.2% at the Youden cutoff and increased to ≥ 2.5% only when sensitivity decreased to 38%, illustrating the difficulty of achieving clinically actionable predictive value in low-incidence populations[18]. These studies are encouraging because they show that routine clinical data can further enrich risk, but they also show that discrimination alone does not solve the central problem of low prevalence.

Overall, the evidence suggests that risk stratification can improve substantially on an unselected NOD strategy, but current models are better viewed as gatekeepers for additional evaluation than as stand-alone arbiters of diagnostic intervention.

CLINICAL FEASIBILITY, BIOMARKERS, AND IMPLEMENTATION

The feasibility of broad diagnostic evaluation in NOD is limited less by the intrinsic specificity of cross-sectional imaging than by the low pretest probability of PDAC in unselected NOD populations. In this setting, even accurate imaging can yield a low positive predictive value, with many positive or indeterminate findings representing incidental or clinically irrelevant abnormalities rather than occult PDAC. Endoscopic ultrasound may improve lesion characterization but is invasive, operator-dependent, and difficult to scale[2,4]. Therefore, the practical value of NOD-based detection strategies depends on improving pretest probability before imaging is performed.

Biomarkers are one possible way to achieve that. In a proof-of-concept study, Cohen et al[19] combined circulating tumor DNA analysis with protein biomarkers in patients with resectable pancreatic cancer and controls, reporting 64% sensitivity and 99.5% specificity. This result illustrates the promise of biomarker integration, but it does not eliminate the challenge posed by low baseline incidence in NOD populations[19]. Even highly specific tests can generate limited positive predictive value when applied to broadly defined low-incidence groups.

This problem becomes clear when considered alongside evidence from established high-risk surveillance settings. In a meta-analysis of prospective screening cohorts among asymptomatic individuals at hereditary or familial high risk, Corral et al[13] estimated that 135 individuals needed to be screened to detect 1 high-risk pancreatic lesion. If diagnostic yield is modest even in hereditary high-risk populations, indiscriminate evaluation of all patients with NOD is unlikely to be efficient or clinically sustainable.

The most realistic implementation strategy is therefore sequential rather than binary: Identify NOD, refine risk using dynamic clinical features and prediction models, and reserve biomarkers and targeted pancreatic imaging for the subset in whom absolute risk becomes meaningfully elevated.

TRANSLATING NOD-ASSOCIATED RISK INTO CLINICAL DECISION-MAKING

This review does not argue against current guideline recommendations; rather, it addresses the translational gap between epidemiologic association and screening eligibility. The critical question is no longer whether NOD is associated with PDAC. That relationship is supported by recent prospective validation, population-based epidemiologic studies, temporal case-control analyses, and meta-analytic evidence[7-9,12]. The more relevant question is whether NOD alone defines a population with sufficiently high absolute risk to justify diagnostic intervention. Current evidence suggests that it does not[3,4,6,12].

A more defensible framework is stepwise risk enrichment (Figure 1), integrating epidemiologic signals, clinical modifiers, and biomarkers to guide diagnostic decision-making. Within this framework, the clinical value of NOD lies not in immediate diagnostic escalation, but in structured, sequential risk refinement. First, NOD should be identified in real time using clinical or glycemic criteria. Second, absolute risk should be contextualized by age and time since diabetes onset. Third, patients should be further stratified using dynamic risk features, validated clinical models, and, where available, biomarker panels. Only those crossing a clinically actionable threshold should proceed to targeted pancreatic imaging or endoscopic evaluation.

Figure 1 Stepwise framework for translating new-onset diabetes into pancreatic cancer risk enrichment. New-onset diabetes should be interpreted as an entry signal rather than a screening indication. After identification of new-onset diabetes, risk should be refined sequentially using age, time since diabetes onset, weight change, glycemic trajectory, family history, pancreatitis-related history, and validated clinical models. Biomarkers and targeted pancreatic imaging should be reserved for the highest-risk subgroup in whom absolute risk is clinically actionable. PDAC: Pancreatic ductal adenocarcinoma; EUS: Endoscopic ultrasound; END-PAC: Enriching new-onset diabetes for pancreatic cancer; CT: Computed tomography; MRI: Magnetic resonance imaging; CA19-9: Carbohydrate antigen 19-9.

This framework preserves the value of NOD without overstating its clinical implications. It acknowledges that NOD is a useful signal, but not a sufficient indication for screening or diagnostic workup on its own. It also aligns with current evidence showing that the greatest unmet need is not further demonstration of association, but prospective validation of multicomponent enrichment strategies across diverse populations, with explicit assessment of diagnostic yield, harms, and cost-effectiveness[12,13,16-18].

CONCLUSION

Although NOD is a reproducible clinical signal associated with PDAC, its low absolute risk precludes use as a screening indication. Current evidence supports positioning NOD as a gatekeeper for risk enrichment rather than a trigger for diagnostic evaluation. Future efforts should focus on integrated strategies combining clinical trajectories and biomarkers to identify subgroups in whom absolute risk reaches a clinically actionable threshold, supported by prospective validation and careful evaluation of the harm-benefit balance.