Polygenic Risk Scores for Pancreatic Cancer Risk Prediction and Early Detection: A Systematic Evidence Review of Clinical Utility

1. Introduction

Pancreatic cancer remains among the most lethal malignancies worldwide, with a five-year relative survival rate of 13%.[1] This poor survival figure is primarily driven by late stage diagnosis, with most patients presenting with advanced, unresectable disease, which eliminates curative resection and restricts treatment to palliative options.[1] The incidence continues to rise across high-income countries and among younger adults, with particularly notable increases among younger women.[2,3] Established risk factors include heritable predisposition, notably pathogenic variants in BRCA1/2, CDKN2A, and STK11, together with smoking, obesity, chronic pancreatitis, and diabetes.[3,4,5,6,7,8,9] High-penetrance variants, however, account for a small fraction of cases, leaving most population-level risk distributed across common variants with individually small effects. In current practice, risk stratification for pancreatic cancer primarily relies on clinical factors, family history, imaging, and testing for high-penetrance germline mutations.[7,10,11] These efforts identify the extremes of the risk distribution but miss individuals with moderate polygenic predisposition, and no approach performs well enough to support population-wide screening.[12,13]

Polygenic risk scores (PRS) aggregate many common variants identified through genome-wide association studies (GWAS) to estimate inherited susceptibility.[14] In addition, combining PRS with established clinical predictors offers a route to stratify risk among people carrying no high-penetrance mutation.[15] Recent evaluations suggest that such integration yields small but meaningful gains, whereas PRS used alone provides only modest discrimination.[16,17,18,19,20] Verras et al. meta-analyzed 27 studies and reported pooled standalone discrimination of Area Under the Receiver Operating Characteristic Curve (AUC) of 0.61 (95% confidence interval [CI], 0.58 to 0.65), rising to 0.70 (95% CI, 0.61 to 0.80) for mixed models that added clinical risk factors. The odds ratio per standard deviation reached 1.42 (95% CI, 1.30 to 1.56) in populations of European ancestry but 1.21 (95% CI, 0.77 to 1.90) in Asian populations, where the association did not reach statistical significance.[16] Wang et al. scoped 21 studies published through March 2023 and found that only seven evaluated risk factors beyond age and sex, and only three quantified the discrimination gain attributable to adding PRS to a clinical model.[17]

Recent reviews have concentrated on association and discrimination endpoints. Clinical utility depends on a different class of evidence, including calibrated absolute risk estimates, explicit thresholds that separate action from observation, and decision analytic measures demonstrating that the resulting management changes deliver net benefit at an acceptable cost.[21] Because absolute risk remains low even in enriched strata, PRS is most plausibly valuable as an enrichment layer within multistep pathways, for example, PRS combined with clinical factors, followed by biomarker triage and targeted imaging, delivered through decision support integrated into the Electronic Health Record (EHR).[22] Evidence across these steps has not been synthesized, and the literature has expanded substantially since the most recent scoping review closed its search in March 2023, including studies conducted in enriched populations, such as those with new-onset diabetes.

Whether PRS-informed risk stratification for pancreatic cancer risk prediction is ready for clinical application remains unresolved.[23] In this systematic evidence review, we synthesized evidence on the clinical utility of PRS for pancreatic cancer risk assessment across five prespecified domains: threshold based discrimination, incremental value beyond established clinical predictors, calibrated absolute risk, decision analytic outputs, and economic or downstream resource impact. We further characterized PRS construction, external validation, and transportability across ancestry groups and appraised risk of bias and applicability using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Our objective was to establish the current state of evidence in each domain and specify the studies required to move PRS from association to clinical practice.

2. Materials and Methods

2.1. Protocol and Reporting

This systematic review was conducted in accordance with a published protocol (PROSPERO number: CRD42024511326) and adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Supplementary Table 1).[24,25] We amended the PROSPERO record after registration to extend the search end date to May 11, 2026, so the prespecified update captured studies published after the initial search.

2.2. Eligibility Criteria

We included peer-reviewed, English-language studies published through May 11, 2026, that evaluated adults (≥18 years) and reported development, validation, or performance of PRS for primary pancreatic cancer risk prediction, either as standalone models or combined with non-genetic predictors (e.g., clinical risk factors). Studies using observational designs (cohort, case-control, cross-sectional) were eligible, as were clinical trials and validation studies (internal and/or external) when PRS was evaluated for risk prediction or intervention guidance. We excluded studies focused on pediatric populations (<18 years), simulated populations, or monogenic/syndromic pancreatic cancer. We also excluded editorials, commentaries, reviews, case reports/series, book chapters, and studies without primary data.

To maintain a clinical utility focus, we excluded prognostic models in diagnosed patients (e.g., progression/metastasis prediction) and studies that did not report outcomes relevant to clinical application (e.g., risk stratification, screening guidance, or patient-centered/decision relevant outcomes), as well as studies lacking adequate methodological detail for PRS development/validation. We distinguished redundant reporting from distinct analyses of shared data. When multiple reports described the same PRS analysis in the same cohort, we retained the most complete and most recent report and excluded the earlier duplicates. When separate reports analyzed overlapping cohorts (e.g., the UK Biobank) but addressed different predictors, outcomes, or clinical-utility questions, we retained each report as an independent study and recorded the cohort overlap in the risk of bias appraisal.

2.3. Information Sources and Search Strategy

A health sciences librarian developed and executed searches in PubMed, Embase, Cochrane Library, and Web of Science. The search combined controlled vocabulary and keywords for pancreatic cancer and polygenic/genetic risk scores. The initial search covered database inception through December 31, 2023. A prespecified update, executed with the same controlled vocabulary and keyword strategy, extended coverage from January 1, 2024, through May 11, 2026. Records from both search periods were combined and processed using a single workflow for deduplication, screening, and synthesis. Gray literature sources were screened, and reference lists of included studies were hand-searched to identify additional records. Full search strategies for both search periods are provided in Supplementary Table 2. The lead author (M.E.) and the study team members (O.B., Y.A., Y.K.) reviewed the strategy for completeness.

2.4. Study Selection

Records were imported into Covidence for deduplication, screening, and data management. Each title, abstract, and full-text was screened independently by reviewers (O.B., Y.A., Y.K.), with discrepancies resolved by consensus and adjudication by a fourth reviewer (M.E.) when needed.

2.5. Data Extraction, Synthesis, and Risk of Bias Assessment

We used the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies (CHARMS) to guide standardized extraction.[26] Data were extracted in Covidence using a structured template customized for this review. The extracted data included study-level information, such as details on risk models, PRS estimation and construction methods, participant-level characteristics, intervention types, and assessed outcomes. Two reviewers extracted data from each report [independently / one extracting and one verifying] (O.B., Y.A., Y.K.), with discrepancies resolved by M.E. Study investigators were not contacted. Metrics not reported in the main text or supplement were recorded as not reported.

2.6. Clinical Utility assessment

We prespecified a clinical utility framework focused on decision relevant performance and actionability rather than association alone. For each eligible study, we abstracted the measures listed in Table 1 and organized clinical utility outcomes across five domains: (i) threshold based predictive accuracy at study defined risk thresholds; (ii) incremental value/model comparison when adding a PRS to clinical models; (iii) calibration and absolute risk characterization required for actionability; (iv) decision analytic outputs; and (v) economic and downstream impact measures. Discrimination establishes clinical validity rather than clinical utility, and we treated it as the prerequisite against which the five utility domains are read, not as evidence of actionability in itself. We therefore abstracted discrimination separately for every eligible study and report it alongside the utility domains. External validation was likewise recorded separately, as a marker of transportability rather than a utility domain.

Discrimination metrics were recorded as reported and were not converted between scales. For case-control and other binary-outcome designs, AUC and the concordance statistic are numerically identical. For time to event analyses subject to censoring, Harrell's C estimates concordance over follow-up and is not the same estimand as a cross-sectional AUC. Reported association estimates were also non-comparable across studies, spanning per-standard-deviation effects, per-allele effects, and contrasts between extreme quantiles with differing cut-points. We therefore did not display them on a common forest plot axis. Instead, we used a structured narrative synthesis supported by evidence tables and figures aligned to the framework above.

2.7. Risk of Bias and Applicability

Because no quantitative pooling was undertaken, publication bias assessments (e.g., funnel plots) were not conducted, and Grading of Recommendations Assessment, Development and Evaluation (GRADE) certainty ratings[27] were not applied. Instead, the PROBAST[28] was used to appraise the risk of bias and applicability across four domains: participants, predictors, outcome, and analysis. Two reviewers applied PROBAST to each study independently, with disagreements resolved by M.E. We rated each domain as low, high, or unclear risk of bias and assigned an overall judgment for each included study. PROBAST risk of bias and applicability judgments for each included study are summarized in Table 4.

2.8. Outcome and predictor variables

The primary outcome was pancreatic cancer risk, defined by pancreatic cancer case status or time to diagnosis as specified in each included study. The primary predictor was a pancreatic cancer PRS, evaluated either as a continuous standardized score (e.g., per 1 standard deviation [SD]) alone or in combination with clinical and demographic risk factors. Outcomes of interest included PRS association estimates and the prespecified model performance and clinical utility measures listed in Table 2 and Table 3.

2.9. Statistical Analysis and Evaluation

Given the substantial heterogeneity in PRS construction (e.g., from a small number of genome-wide-significant variants to genome-wide polygenic scores; weighted vs. unweighted), covariate sets, ancestry, validation approaches, and analytic windows, we did not perform a meta-analysis. Instead, we employed a narrative synthesis, displaying Single-Nucleotide Polymorphism (SNP) compositions and reuse across studies in Figure 2, translational milestone coverage across all included studies in Figure 3, discrimination and incremental value in Figure 4, and study and model characteristics in Table 2 and Table 3. Associations were abstracted as odds ratios (ORs) or hazard ratios (HRs) with 95% CIs. Where available, we preferentially recorded OR/HR per 1-SD increase in PRS. When studies reported quantiles (e.g., top vs. bottom quintiles or deciles), we captured those contrasts and noted the cut points used. For prospective cohorts, we recorded follow-up windows and time to event specifications. Decision analytic and economic outputs (decision curve net benefit, clinical impact curves, and cost-effectiveness summaries such as incremental Cost-Effectiveness Ratio (ICER) per Quality-Adjusted Life Year (QALY) and willingness-to-pay thresholds) were extracted when available and summarized narratively rather than quantitatively, given their dependence on local thresholds and resource assumptions. Validation and transportability were characterized as internal versus external (temporal/geographic/ancestry-external), including whether recalibration was required prior to reporting absolute risk.

3. Results

3.1. Study Selection

The PRISMA flow diagram (Figure 1) summarizes the study-selection process. We identified 874 records from databases and registers, plus 6 additional records from other sources. After removing 278 duplicates (277 via Covidence, 1 manually), 602 unique records underwent title and abstract screening. Of the screened records, 537 were excluded. Of the 65 reports sought for retrieval, 3 were not retrieved, leaving 62 full-text reports assessed for eligibility. Nineteen reports were excluded: abstract only (n = 10), repeated or duplicate publication (n = 2), review paper (n = 1), wrong outcomes (n = 3), wrong comparator (n = 1), and wrong intervention (n = 2). Accordingly, 43 studies met all inclusion criteria and were included in the final synthesis.

3.2. Study Characteristics

Table 2 reports the main characteristics of the 43 included studies, published between 2012 and 2026. Case-control designs predominated (n = 24), followed by cohort studies (n = 16) and nested case-control studies (n = 3). Study populations and data sources were heterogeneous, spanning population biobanks, cancer cohorts, and international consortia. UK Biobank was the most frequently used single data source (25 studies), including analyses focused on PRS performance alone and those integrating PRS with clinical risk factors.[18,19,20,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50] Analytic sample sizes varied considerably, from 235 participants (103 cases and 132 controls) in a multi-ethnic Kuwaiti case-control study[51] to biobank-based analyses with hundreds of thousands of participants.[18,29,31,32,36,52,53] Several studies also used large external GWAS consortia for PRS derivation or genetic instrument construction, including datasets comprising more than 1.4 million participants.[37] Ancestry representation was limited. Thirty-four studies were conducted in European ancestry populations.[18,19,20,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,52,54,55,56,57,58,59,60,61] Three studies focused primarily on East Asian populations.[62,63,64] Multi-ancestry or diverse multiethnic samples were uncommon and included a U.S. multiethnic sample[65], stratified analyses within African and European ancestry groups[66], consortium-based datasets with mixed ancestry representation,[53] and a Brazilian case-control study.[67]

3.3. Model Characteristics

Model characteristics, PRS composition, covariate adjustment, and reported outcomes are summarized in Table 3. The number of genetic variants included in PRS models ranged from 4[51] to 49[20] genotyped variants after linkage disequilibrium pruning, and one study derived a genome-wide score using AnnoPred.[40] Two studies carried non-pancreatic comparator scores, an 87-variant transaminase score[48] and a 424-variant type 2 diabetes score.[49]

Figure 2 shows the number of times SNP(s) were used across synthesized studies, with rs6971499 recurring in 24 of the 29 studies reporting individual variants, and rs16986825 and rs9543325 in 23. Most PRSs were derived from common-variant GWAS results and analyzed either as continuous measures (i.e., often per SD increase or similar scaling) or as risk strata based on quantiles (e.g., quintiles, deciles, percentile cut points). Continuous PRS modeling appeared more commonly in cohort-based analyses reporting hazard ratios per PRS increment.[18,31,32] Extreme-strata comparisons included highest-versus-lowest quintile analyses[20,29,30] and top-versus-bottom decile analyses.[53]

Several studies evaluated approaches beyond conventional GWAS-based PRS. These included telomere-length polygenic scores ("teloscores")[55,68], transcriptome-informed or Transcriptome-Wide Association Study (TWAS) augmented models[56], cross-validation and threshold based SNP inclusion strategies[57], and models embedding PRS within broader multi-exposure or multi-cancer frameworks.[33,35,52] Clinical and environmental covariates varied across studies but most commonly included age and sex, with frequent adjustment for smoking, Body Mass Index (BMI) or anthropometrics, diabetes, family history, and pancreatitis (Table 3). A small number of studies integrated PRS into models that explicitly combined multiple domains of predictors, including biomarkers or EHR-derived risk features.[19,30,34]

Figure 2. SNP composition and reuse across pancreatic cancer PRS studies. Rows show the 29 studies that reported individual variants, grouped by panel family. Columns show the 64 variants carried by two or more studies, ordered by frequency; the upper bars give the number of studies carrying each variant, and the bracket marks the 22 variants present in at least 12 studies. Right-hand bars give candidate variants per study before linkage disequilibrium pruning (dark, shared; light, study-unique); analytic PRS size after pruning is given in Table 3 and is smaller for Sharma 2022 (49 of 54) and Zeng 2023 (44 of 54 for PRS54). Al-Qahtani 2024 shows the four GWAS variants that entered its score. Yuan 2025 is omitted because its 87-variant transaminase score is not a pancreatic PRS.

Figure 2. SNP composition and reuse across pancreatic cancer PRS studies. Rows show the 29 studies that reported individual variants, grouped by panel family. Columns show the 64 variants carried by two or more studies, ordered by frequency; the upper bars give the number of studies carrying each variant, and the bracket marks the 22 variants present in at least 12 studies. Right-hand bars give candidate variants per study before linkage disequilibrium pruning (dark, shared; light, study-unique); analytic PRS size after pruning is given in Table 3 and is smaller for Sharma 2022 (49 of 54) and Zeng 2023 (44 of 54 for PRS54). Al-Qahtani 2024 shows the four GWAS variants that entered its score. Yuan 2025 is omitted because its 87-variant transaminase score is not a pancreatic PRS.

3.4. PRS Effect Sizes

PRS associations with pancreatic cancer risk ranged from near-null to strong gradients depending on PRS construction, population, and the risk contrast used (Table 3). On the odds ratio scale, effect sizes spanned from OR = 0.55 to OR = 14.37. The lower bound reflected a referent-category contrast in one allele-count implementation[60], while the upper bound reflected a combined multifactorial score incorporating PRS plus clinical risk factors.[54] Continuous PRS associations were generally modest. Kim et al. (2020) reported ORs of 1.37 for the clinical + genetic model and 1.36 after adding biomarkers to the model. Telomere-length scores were close to null in at least one analysis (Antwi et al., 2017; multivariable-adjusted OR = 1.03). Stronger gradients emerged when comparing extreme PRS strata. For example, Ke et al. (2023) reported OR = 2.03 for the highest versus lowest PRS quintile (Q5 vs. Q1) in the UK Biobank, and Sharma et al. (2022) reported OR = 2.74 for Q5 versus Q1 in a diabetes-focused prediction context.

Elevated risk estimates were observed in selected settings and subgroups. Aoki et al. (2021) reported OR = 8.13 for Q5 versus Q1 in a Brazilian cohort, although this estimate was based on a relatively small case-control sample. In an age-stratified analysis, Yuan et al. (2022) reported OR = 6.91 for the top versus bottom 10% of PRS among individuals aged 60 years or younger (and OR = 4.12 among those older than 70 years). For early-onset pancreatic cancer, Nodari et al. (2023) reported OR = 4.93 for age 50 years or younger in Pancreatic Cancer Cohort Consortium (PanScan) plus Pancreatic Cancer Case-Control Consortium (PanC4) (and OR = 2.90 in Pancreatic Disease Research consortium [PANDoRA]), with smaller but still elevated estimates among individuals older than 50 years. In cohort-based analyses reporting hazard ratios, continuous PRS associations were generally modest (HR approximately 1.10 to 1.58), with stronger gradients for extreme strata contrasts. Zeng et al. (2023) reported HR = 1.58 for PRS modeled continuously and HR = 1.89 for the highest versus middle PRS tertile among individuals aged 60 years or younger. Kachuri et al. (2020) reported HR = 1.49 in UK Biobank. Jia et al. (2020) reported HR = 3.37 for Q5 versus Q1, with HR = 2.31 for the top 5% versus an average risk group.

3.5. Model Performance

Twenty-two of the 43 studies reported discrimination, and 20 contributed interpretable model-level estimates to Figure 3. Standalone PRS discrimination clustered tightly. Ten of the 16 standalone estimates fell between 0.56 and 0.64, with the full range spanning 0.53 in the Michigan Genomics Initiative validation cohort to 0.88 in a new-onset diabetes cohort. The lowest values were reported in studies with early cross-validation approaches.[19,57] The highest standalone estimate came from Sun et al. (2024), whose genetic-only logistic regression model achieved an AUC of 0.879 (95% CI, 0.842 to 0.915) among patients with new-onset diabetes. That model was classified on 24 study-selected variants rather than applying a published weighted PRS, which sets it apart from the remaining estimates and places it well above the range reported elsewhere. Mid-range PRS-only AUC values included those from Jia et al. (2020) (AUC = 0.639) and Kim et al. (2023) (AUC = 0.58 in females and 0.59 in males).

Adding PRS to clinical predictors typically produced statistically significant but numerically modest improvements in discrimination (Figure 3). In an analysis of the UK Biobank, Kachuri et al. (2020) reported AUC increases from 0.715 (age, sex, family history, and other predictors) to 0.745 after adding PRS, with parallel C-index improvements (from 0.714 to 0.743). Similarly, adding risk factors to PRS models improved the AUC from 0.72 to 0.78 in the UK Biobank[30], and Hu et al. (2024) reported a larger gain, from 0.549 for a genome-wide PRS alone to 0.689 after adding clinical variables.[40] Kim et al. (2020) reported AUC increases from 0.61 (clinical model) to 0.67 (clinical plus genetic) and 0.69 (clinical plus genetic plus biomarker) in the 0- to 10-year follow-up window. Dite et al. (2023) reported a small gain, with a Harrell's C of 0.702 for the Jia et al. (2020) 22-SNP PRS, 0.703 for the Klein clinical score[69], and 0.714 (95% CI, 0.698 to 0.730) for the combined 10-year risk score. This amounted to a difference of 0.011 that reached significance against both components in 376,462 participants.

Zhou et al. (2025) illustrated how much the standalone estimate depends on how the score is derived. Their PRS re-derived within the new-onset diabetes setting reached AUC = 0.719 (95% CI, 0.687 to 0.751), whereas a standard pancreatic cancer PRS applied to the same cohort reached 0.611 (95% CI, 0.575 to 0.646) and a type 2 diabetes PRS reached 0.592 (95% CI, 0.555 to 0.628).[49] Sharma et al. (2022) reported an increase in AUC from 0.791 (clinical risk model) to 0.83 (combined PRS plus clinical risk model; p = 0.0002). Zeng et al. (2023) reported a PRS-only AUC of 0.616 and an age-specific AUC of 0.623 among individuals aged 60 years or younger when PRS was stratified by age. In several multi-domain models, non-genetic predictors drove most of the discrimination, with PRS contributing only incremental change. This pattern was evident in an ancestry-stratified evaluation in which adding PRS yielded no AUC gain over a baseline model including principal components, age, and sex (Wang et al., 2022; African ancestry ΔAUC = 0, p = 0.75; European ancestry ΔAUC = -0.003, p = 0.44).

3.6. Further Assessment of PRS Performance and Clinical Utility Measures

Thirteen studies reported at least one measure beyond discrimination that informs clinical utility or model reliability (Table 3; Figure 4), including threshold based predictive accuracy metrics, calibration or goodness-of-fit outputs, reclassification metrics, and error-based measures. In a diabetes-stratified prediction setting, Sharma et al. (2022) reported sensitivity of 31.8% overall (range 28.9% to 41.4% across diabetes strata), specificity of 81.2% overall (up to 91.4% in new-onset diabetes), and Positive Predictive Value (PPV) ranging from 11.9% (no diabetes) to 86.7% (new-onset diabetes) when focusing on the highest PRS quintile. For reclassification, Kachuri et al. (2020) reported a Net Reclassification Improvement (NRI) of 0.228 (95% CI, 0.174 to 0.284) when PRS was added to a comprehensive conventional risk factor model. Zhou et al. (2025) reported a continuous NRI of 0.319 (95% CI, 0.207 to 0.436) in cases and 0.309 (95% CI, 0.293 to 0.325) in controls with an Integrated Discrimination Improvement (IDI) of 0.021 (95% CI, 0.015 to 0.028)[49], and Dite et al. (2023) reported an NRI of 0.066 (95% CI, 0.035 to 0.098) for the combined score against the PRS and 0.060 (95% CI, 0.027 to 0.094) against the clinical score, using prespecified 10-year risk thresholds of 0.36% and 0.64%.[50]

Figure 3. Discrimination and incremental value of pancreatic cancer PRS in the 20 studies reporting interpretable model-level estimates. Diamonds, PRS alone; open circles, base model without PRS; filled circles, model with PRS. The shaded band marks 0.56 to 0.64, where 10 of 16 standalone estimates fell. Grey lines connect ROC AUC estimates; purple lines connect Harrell's C estimates from time-to-event analyses (Zhou 2025 model estimates; Zhu 2025; Kachuri 2020; Dite 2023). Δ is the gain from adding PRS to a base model where one was reported; for Zhou 2025, Kachuri 2020, and Dite 2023 it is a difference in Harrell's C. Shaded rows and red Δ mark new-onset diabetes cohorts (Sun 2024, Zhou 2025). Zhou 2025 plots the re-derived new-onset-diabetes PRS; its standard pancreatic cancer PRS (0.611) and type 2 diabetes PRS (0.592) are reported in the text. Sun 2024's standalone estimate derives from a logistic model on 24 study-selected variants. Wang 2022 values come from a secondary analysis with 46 cases per ancestry stratum. Peduzzi 2025 (UEG and DLD) are omitted because they report only combined machine-learning AUCs.

Figure 3. Discrimination and incremental value of pancreatic cancer PRS in the 20 studies reporting interpretable model-level estimates. Diamonds, PRS alone; open circles, base model without PRS; filled circles, model with PRS. The shaded band marks 0.56 to 0.64, where 10 of 16 standalone estimates fell. Grey lines connect ROC AUC estimates; purple lines connect Harrell's C estimates from time-to-event analyses (Zhou 2025 model estimates; Zhu 2025; Kachuri 2020; Dite 2023). Δ is the gain from adding PRS to a base model where one was reported; for Zhou 2025, Kachuri 2020, and Dite 2023 it is a difference in Harrell's C. Shaded rows and red Δ mark new-onset diabetes cohorts (Sun 2024, Zhou 2025). Zhou 2025 plots the re-derived new-onset-diabetes PRS; its standard pancreatic cancer PRS (0.611) and type 2 diabetes PRS (0.592) are reported in the text. Sun 2024's standalone estimate derives from a logistic model on 24 study-selected variants. Wang 2022 values come from a secondary analysis with 46 cases per ancestry stratum. Peduzzi 2025 (UEG and DLD) are omitted because they report only combined machine-learning AUCs.

Calibration was reported in six studies.[19,30,40,45,49,50] Sun et al. (2024) reported adequate calibration for the logistic regression model and for combined models in four of five algorithms, and Zhou et al. (2025) reported a calibration slope of 0.994 (95% CI, 0.885 to 1.102) with calibration-in-the-large of -0.006 (95% CI, -0.115 to 0.102). Dite et al. (2023) reported a calibration slope of 0.952 (95% CI, 0.865 to 1.039) with slight overestimation in the calibration in the large (-0.089; 95% CI, -0.156 to -0.021), consistent with the cohort's lower than expected incidence. Salvatore et al. (2021) reported Brier scores of 0.22370 for PRS-only models and 0.21610 for combined models (phenotype risk score [PheRS] plus PRS plus covariates/risk factors) in the Michigan Genomics Initiative cohort. The Hosmer-Lemeshow test suggested poor calibration for PRS-only models (p < 0.05) but improved calibration in combined models. In a UK Biobank nested case-control analysis, Ke et al. (2023) reported a Hosmer-Lemeshow test p of 0.1543, consistent with an acceptable fit of the evaluated model specification. Model-fit comparisons were also occasionally reported. Nakatochi et al. (2018) reported Akaike Information Criterion (AIC) values to compare model fit in a Japanese risk model (AIC 1764 for risk factors plus PRS versus 1796 for PRS-only). Pierce et al. (2012) reported low variance explained (R² of 0.005 to 0.018) and limited incremental improvements in sensitivity and specificity of the fitted PRS models relative to baseline prediction.

In a screening-adjacent context, Kim et al. (2023) evaluated PRS-stratified PPV for multi-cancer early detection (MCED) tests, reporting PPV differences by PRS percentile and age. For the Galleri test, PPV was 70.3% for high-risk females aged 75 years at the 90th to 95th percentile, compared with 48.2% for low-risk females at the 5th to 10th percentile, a 22.1 percentage-point difference. Evidence on gene-environment interaction was limited. Rothwell et al. (2022) reported that positive associations between metabolic syndrome and pancreatic cancer risk were maintained across PRS strata, with no evidence of interaction (p = 0.69).

Two studies reported decision curve analyses, both in new-onset diabetes cohorts.[45,49] Sun et al. (2024) went further and anchored two explicit action thresholds, reporting that a 1.28% cut-off restricts definitive testing to 13% of the cohort while capturing 76% of cases, and that a 5.26% cut-off restricts testing to 2% while capturing nearly half. Three studies reported external validation.[19,44,49] Calibrated absolute risk anchored to explicit action thresholds appeared in two studies, Sun et al. (2024) and Dite et al. (2023). Kim et al. (2020) reported absolute risk by percentile, with cumulative risk to age 80 of 0.4% at the 10^th percentile and 3.8% at the 99^th, but derived it from population incidence rates without calibration or an action threshold.[34] Dite et al. (2023) went further, computing competing mortality-adjusted 10-year absolute risk and anchoring two thresholds to existing screening criteria. At the high-risk threshold, the combined score identified 17% of cases, compared with 11% for the clinical score, and the number needed to screen fell from 1 in 442 unselected adults to 1 in 145. No study reported a cost-effectiveness analysis of a PRS-guided pancreatic cancer screening pathway.

Figure 3. Evidence gap map of clinical-translation milestones across the 43 included studies. Filled squares indicate a reported milestone; rows are ordered by milestone count. Twenty-three studies reported at least one milestone and 20 reported none. No study reported a cost-effectiveness analysis.

Figure 3. Evidence gap map of clinical-translation milestones across the 43 included studies. Filled squares indicate a reported milestone; rows are ordered by milestone count. Twenty-three studies reported at least one milestone and 20 reported none. No study reported a cost-effectiveness analysis.

3.7. Quality Assessment

Risk of bias was assessed for all 43 studies (Table 4). Overall risk of bias was low in 16 studies, high in 13, and unclear in 14. Most adverse judgments came from the analysis domain, which was rated low in 20 studies, high in 11, and unclear in 12. That domain alone drove 11 of the 13 high overall ratings, with the remaining two studies, Byrne et al. (2023) and Nodari et al. (2023), rated high for participant selection.[32,61] Internal validation and correction for optimism were often absent. Calibration was rarely reported, and several studies tested a score in the same sample that had discovered its variants. The smallest studies fit more predictors than their event counts could support. In contrast, the participants, predictors, and outcome domains were rated low in 36, 41, and 39 studies. No study was rated high risk on predictors or outcomes. Applicability raised no high concern in any domain. Thirty-eight studies were judged to be of low concern overall, and the five judged unclear had uncertain population representativeness or incomplete reporting of predictors rather than any mismatch with the review question.[18,32,51,64,66] As such, the weakness in this evidence base lies in how the models were built and tested, not in the populations they were built for.

Table 4. PROBAST Risk of Bias and Applicability Judgments for the 43 Included Studies.

Author, Year	Risk of Bias				Applicability			Overall
	1. Participants	2. Predictors	3. Outcome	4. Analysis	1. Participants	2. Predictors	3. Outcome	Risk of Bias	Applicability
Al-Qahtani, 2024	?	+	+	-	?	+	+	-	?
Antwi, 2017	+	+	+	+	+	+	+	+	+
Aoki, 2021	+	+	+	-	+	+	+	-	+
Bogumil, 2020	+	+	?	+	+	+	+	?	+
Byrne, 2023	-	+	+	?	?	+	+	-	?
Dite, 2023	+	+	+	+	+	+	+	+	+
Felici, 2024	+	+	+	-	+	+	+	-	+
Galeotti, 2021	+	+	+	+	+	+	+	+	+
Gentiluomo, 2022	?	+	?	-	+	+	+	-	+
Graff, 2021	+	+	+	+	+	+	+	+	+
Hu, 2024	+	+	+	?	+	+	+	?	+
Jia, 2020	+	+	+	+	+	+	+	+	+
Jiang, 2026	+	+	+	?	+	+	+	?	+
Kachuri, 2020	+	?	+	+	+	?	+	?	?
Ke, 2023	+	?	+	+	+	+	+	?	+
Kim, 2023	+	+	+	?	+	+	+	?	+
Kim, 2020	+	+	+	?	+	+	+	?	+
Klein, 2018	+	+	+	-	+	+	+	-	+
Liu, 2020	+	+	+	+	+	+	+	+	+
Meisner, 2020	+	+	+	-	+	+	+	-	+
Nakatochi, 2018	+	+	+	+	+	+	+	+	+
Nodari, 2023	-	+	+	+	+	+	+	-	+
Peduzzi, 2024	+	+	+	+	+	+	+	+	+
Peduzzi (DLD), 2025	+	+	+	-	+	+	+	-	+
Peduzzi (UEG), 2025	+	+	+	+	+	+	+	+	+
Pierce, 2012	+	+	+	+	+	+	+	+	+
Prassas, 2025	+	+	+	?	+	+	+	?	+
Rothwell, 2022	+	+	+	+	+	+	+	+	+
Salvatore, 2021	+	+	+	?	+	+	+	?	+
Sharma, 2022	+	+	+	+	+	+	+	+	+
Shi, 2019	-	+	+	-	+	+	+	-	+
Sun, 2024	+	+	+	?	+	+	+	?	+
Wang, 2020	+	+	+	+	+	+	+	+	+
Wang, 2022	-	+	?	-	?	+	+	-	?
Wolpin, 2014	+	+	+	-	+	+	+	-	+
Wu D, 2025	+	+	+	?	+	+	+	?	+
Wu Z, 2025	+	+	+	-	+	+	+	-	+
Yuan, 2022	+	+	+	+	+	+	+	+	+
Yuan, 2025	+	+	+	?	+	+	+	?	+
Zeng, 2023	+	+	+	?	+	+	+	?	+
Zhou, 2025	+	+	+	+	+	+	+	+	+
Zhu, 2025	?	+	?	?	?	+	?	?	?
Zhang, 2022	+	+	+	+	+	+	+	+	+

Table 4. PROBAST Risk of Bias and Applicability Judgments for the 43 Included Studies.

Author, Year	Risk of Bias				Applicability			Overall
	1. Participants	2. Predictors	3. Outcome	4. Analysis	1. Participants	2. Predictors	3. Outcome	Risk of Bias	Applicability
Al-Qahtani, 2024	?	+	+	-	?	+	+	-	?
Antwi, 2017	+	+	+	+	+	+	+	+	+
Aoki, 2021	+	+	+	-	+	+	+	-	+
Bogumil, 2020	+	+	?	+	+	+	+	?	+
Byrne, 2023	-	+	+	?	?	+	+	-	?
Dite, 2023	+	+	+	+	+	+	+	+	+
Felici, 2024	+	+	+	-	+	+	+	-	+
Galeotti, 2021	+	+	+	+	+	+	+	+	+
Gentiluomo, 2022	?	+	?	-	+	+	+	-	+
Graff, 2021	+	+	+	+	+	+	+	+	+
Hu, 2024	+	+	+	?	+	+	+	?	+
Jia, 2020	+	+	+	+	+	+	+	+	+
Jiang, 2026	+	+	+	?	+	+	+	?	+
Kachuri, 2020	+	?	+	+	+	?	+	?	?
Ke, 2023	+	?	+	+	+	+	+	?	+
Kim, 2023	+	+	+	?	+	+	+	?	+
Kim, 2020	+	+	+	?	+	+	+	?	+
Klein, 2018	+	+	+	-	+	+	+	-	+
Liu, 2020	+	+	+	+	+	+	+	+	+
Meisner, 2020	+	+	+	-	+	+	+	-	+
Nakatochi, 2018	+	+	+	+	+	+	+	+	+
Nodari, 2023	-	+	+	+	+	+	+	-	+
Peduzzi, 2024	+	+	+	+	+	+	+	+	+
Peduzzi (DLD), 2025	+	+	+	-	+	+	+	-	+
Peduzzi (UEG), 2025	+	+	+	+	+	+	+	+	+
Pierce, 2012	+	+	+	+	+	+	+	+	+
Prassas, 2025	+	+	+	?	+	+	+	?	+
Rothwell, 2022	+	+	+	+	+	+	+	+	+
Salvatore, 2021	+	+	+	?	+	+	+	?	+
Sharma, 2022	+	+	+	+	+	+	+	+	+
Shi, 2019	-	+	+	-	+	+	+	-	+
Sun, 2024	+	+	+	?	+	+	+	?	+
Wang, 2020	+	+	+	+	+	+	+	+	+
Wang, 2022	-	+	?	-	?	+	+	-	?
Wolpin, 2014	+	+	+	-	+	+	+	-	+
Wu D, 2025	+	+	+	?	+	+	+	?	+
Wu Z, 2025	+	+	+	-	+	+	+	-	+
Yuan, 2022	+	+	+	+	+	+	+	+	+
Yuan, 2025	+	+	+	?	+	+	+	?	+
Zeng, 2023	+	+	+	?	+	+	+	?	+
Zhou, 2025	+	+	+	+	+	+	+	+	+
Zhu, 2025	?	+	?	?	?	+	?	?	?
Zhang, 2022	+	+	+	+	+	+	+	+	+

4. Discussion

In this systematic evidence review of 43 studies, pancreatic cancer PRS showed reproducible but modest standalone discrimination. While PRS added a small increment to clinical models in general populations, it reached clinically interpretable performance only within enriched populations. Nine studies reported a base model. In the seven general-population cohorts, adding PRS shifted discrimination by at most 0.05, and by nothing at all in the one ancestry-stratified analysis.[66] The two new-onset diabetes cohorts gained 0.074 and 0.188.[45,49] Combined models exceeded 0.80 in three studies, two restricted to new-onset diabetes[45,49] and one built around diabetes as a predictor.[20] This overall pattern is consistent with a recent meta-analysis showing modest pooled standalone discrimination and stronger performance for mixed models, and with a scoping review concluding that relatively few studies have directly evaluated whether PRS improves prediction beyond routine clinical data.[16,17]

Our results add a more explicit translational synthesis to this literature. Rather than stopping at association and discrimination, we evaluated calibration, reclassification, threshold based performance, absolute risk estimation, and implementation readiness. Against those domains, the evidence thinned quickly, with 22 of 43 studies reporting discrimination, 6 calibration, 4 reclassification, 7 an explicit threshold, 2 a decision curve, and none cost-effectiveness, and 20 reported no translational milestone at all. Viewed through that lens, the strongest signal is not that pancreatic cancer PRS is ready for standalone clinical use, but that it may function as an enrichment layer within broader multivariable and multistep risk pathways, particularly in clinically enriched settings such as new-onset diabetes.[18,20,45,49,50]

Standalone PRS discrimination was generally modest in our review, with 10 of the 16 standalone estimates falling between 0.56 and 0.64. These results were comparable to the pooled AUC of 0.61 (95% CI, 0.58 to 0.65) reported by Verras et al. (2025).[16] This level of discrimination suggests that PRS can rank-order inherited susceptibility, particularly when comparing extreme strata, but remains insufficient for independent clinical use in a low-incidence cancer where absolute risk remains small for most individuals. The apparent validity of PRS-only models is further constrained by inconsistent calibration reporting, heterogeneous derivation strategies, and limited validation outside European ancestry datasets. Several studies observed stronger gradients when PRS was analyzed across quantiles or in younger or otherwise enriched subgroups. Outside those settings, the canonical 22-SNP panel and its derivatives produced 0.56 to 0.64 in every cohort that tested them, and 12 of the 29 studies reporting individual variants applied that panel or a subset of it. In turn, part of what the field calls reproducibility is one score re-measured many times. However, these findings do not alter the broader conclusion that PRS alone is better understood as a marker of relative genetic susceptibility rather than as a clinically actionable risk prediction tool.[29,31,57,66]

The strongest evidence for clinical utility emerges when PRS is combined with non-genetic predictors rather than evaluated in isolation. Combined models incorporating PRS with clinical and environmental covariates consistently outperformed PRS-only models and usually exceeded routine clinical models by a modest but nontrivial margin. In the UK Biobank, Kachuri et al. (2020) reported an increase in AUC from 0.715 to 0.745 after adding a 22-SNP PRS to conventional risk factors, with an NRI of 0.228.[18] Sharma et al. (2022) reported that adding PRS to a diabetes-focused clinical model increased AUC from 0.791 to 0.83.[20] Those findings were similar to those of Dite et al. (2023), whose combined 10-year risk score raised Harrell's C from 0.703 to 0.714 in 376,462 participants, with an NRI of 0.060 against the clinical score.[50] Zhou et al. (2025) demonstrated that in patients with new-onset diabetes, a PRS-integrated model achieved strong discrimination, good calibration, significant reclassification improvement, and clinically relevant PPV in the highest-risk stratum.[49] Sun et al. (2024) reported the largest increment in the review, from 0.709 to 0.897, also among patients with new-onset diabetes.[45] At the same time, these gains were incremental rather than transformative in most studies. The added value of PRS depended on the strength of the baseline model, the availability of clinically informative non-genetic predictors, and the risk enrichment of the target population. As such, PRS is unlikely to serve as a standalone screening tool but can improve multivariable risk prediction when embedded in integrated models tailored to specific clinical pathways.

Clinical utility remains the least developed aspect of the pancreatic cancer PRS literature. Although several studies reported favorable discrimination or reclassification, only three translated these gains into calibrated absolute risk estimates, explicit action thresholds, or decision analytic evidence that could support clinical implementation. Most pancreatic cancer PRS studies still do not answer the question that matters for clinical practice: at what level of predicted risk should surveillance begin, and which downstream pathway would that estimate trigger? The available evidence suggests that threshold based utility is context dependent and most convincing in clinically enriched settings rather than in the general population. As such, pancreatic cancer PRS cannot yet be considered clinically actionable based on association or AUC improvement alone. Before a PRS-informed pathway can be implemented, the field will need externally validated and well-calibrated absolute risk models, prespecified thresholds linked to surveillance decisions, and formal decision analytic and cost-effectiveness evaluations to establish net benefit under realistic resource constraints. The only general-population absolute-risk model to date was developed and funded by a commercial genetic testing company, which does not invalidate it but does mean its external validation should come from an independent group.[50]

A further challenge is the substantial heterogeneity in how pancreatic cancer PRS has been constructed, reported, and validated. Across the included studies, PRS differed markedly in variant count, weighting strategy, standardization, covariate handling, endpoint definition, and validation approach, ranging from small-allele-count scores and telomere-based scores to broader, weighted GWAS-derived models and hybrid approaches that incorporated biomarkers or EHR-derived data. This heterogeneity limits direct comparability across studies and makes reported performance metrics difficult to interpret. The problem is compounded by sparse reporting of calibration and absolute risk outputs. These gaps matter because strong relative-risk gradients, by themselves, do not establish transportable or actionable risk thresholds. The evidence base is also heavily Eurocentric. Thirty-four studies were European, three East Asian, and five multi-ancestry. The one ancestry stratified evaluation found no discrimination gain from PRS in African-ancestry participants[66], and Verras et al. reported a per-SD odds ratio in Asian populations that crossed the null.[16] Performance summarized here therefore overstates what a European-derived score will deliver elsewhere. Taken together, these findings underscore the need for more standardized PRS reporting and for more rigorous evaluation of prediction models.[70]

This review has several strengths. We conducted this review using a preregistered protocol, followed PRISMA guidance, applied a prespecified clinical-utility framework, and used structured risk of bias and applicability assessments to evaluate a heterogeneous literature on prediction models. By focusing on discrimination, calibration, reclassification, threshold-based performance, and implementation-relevant measures, our review extends prior pancreatic cancer PRS syntheses beyond association alone and provides a more translational appraisal of clinical readiness. These strengths should be interpreted alongside important limitations. We did not perform a meta-analysis because of substantial heterogeneity in PRS construction, covariate selection, outcome definition, validation strategy, and reporting of performance metrics. The evidence was dominated by case-control designs and European ancestry populations, with limited external validation, sparse calibration, and inconsistent reporting of absolute risk and of decision analytic or economic endpoints. Thirteen studies were rated at high risk of bias, 11 in the analysis domain, most often for absent optimism correction, absent calibration, or evaluation of a score in the sample that discovered its variants. Each of those deficits biases reported discrimination upward, so the modest estimates summarized here are, if anything, optimistic. Applicability raised no high concern in any domain, so the weakness of this literature is how the models were built and tested, not whom they were built for.

5. Conclusions

In conclusion, pancreatic cancer PRS demonstrates a reproducible association with pancreatic cancer risk and modest standalone discrimination. The strongest evidence supports PRS as an incremental component of multivariable models rather than as a standalone screening tool. However, direct clinical-utility evidence remains sparse. Two studies tie calibrated absolute risk to explicit action thresholds[45,50], and two report decision-analytic net benefit.[45,49] All three studies are in the UK Biobank, and two are restricted to new-onset diabetes. None of the reviewed studies reported cost-effectiveness. Three study types would resolve what this review could not. External validation of the Dite and Zhou models in a non-UK cohort, with recalibration to local incidence, would establish transportability. A decision-analytic evaluation of the full pathway from PRS-informed risk to imaging, including false-positive burden and cost per case detected, would establish net benefit. A prospective or emulated-trial comparison of PRS-informed against clinical-factor-only referral among patients with new-onset diabetes would establish whether incremental discrimination becomes earlier-stage diagnosis.