Development and Validation of a Parsimonious Risk Stratification Model for Pancreatic Cancer

Key Points

Question

Can pancreatic ductal adenocarcinoma (PDAC) risk be predicted using routinely available clinical data in diverse patient populations?

Findings

In this cohort study of more than 11 million adults across 54 US health systems and the UK Biobank, a parsimonious model with 19 standard demographic, clinical, and laboratory predictors demonstrated strong discrimination and calibration for incident PDAC.

Meaning

A transparent and generalizable PDAC risk model enabled population-level identification of individuals at elevated risk using routinely collected clinical data, providing a foundation for targeted biomarker testing and multistage early detection strategies.

This cohort study validates a novel parsimonious, interpretable, and generalizable model for predicting incident pancreatic ductal adenocarcinoma.

Abstract

Importance

Pancreatic ductal adenocarcinoma (PDAC) is a leading cause of cancer deaths in the US. Although early detection improves survival, the rarity of the disease has rendered population screening a difficult approach.

Objective

To develop and validate a parsimonious, interpretable, and generalizable model predicting incident PDAC—termed PRIME (PDAC Risk Model for Earlier Detection)—using routinely available electronic health record (EHR) data.

Design, Setting, and Participants

This cohort study used the Optum Labs Data Warehouse, a longitudinal, deidentified US EHR and claims database. Adults 40 years or older with an outpatient clinical encounter between 2016 and 2018 were included. Participants from 23 health systems (n = 4 859 833) comprised the training cohort; 31 additional systems (n = 5 619 091) served as validation. International validation was conducted in the UK Biobank (n = 498 754). Data analysis occurred July 2025 to January 2026.

Exposures

Demographics, diagnosis codes, and routinely measured laboratory values were evaluated. Elastic-net regularization with 10-fold cross-validation selected the predictor set.

Main Outcomes and Measures

Incident PDAC was identified by International Classification of Diseases, Ninth and Tenth Revisions (ICD-9/10) codes. Model performance was assessed using time-dependent area under the curve (AUC) and calibration metrics.

Results

Overall, the study included more than 11 million adults (2.1% Asian individuals, 8.4% Black individuals, 4.3% Hispanic/Latino individuals, 82.7% White individuals, and 2.4% other race/ethnicity by EHR reporting). In the training cohort (mean [SD] age, 60.4 [11] years), 14 405 individuals were diagnosed with PDAC (incidence 55 per 100 000 person-years) over a mean (SD) of 5.4 (2.5) years; in the validation cohort, 11 693 individuals were diagnosed with PDAC (54 per 100 000 person-years) over a mean (SD) of 3.9 (2.5) years. PRIME retained 19 predictors including history of pancreatitis, gastrointestinal disorders, prior cancers, type 2 diabetes, elevated aspartate aminotransferase levels, smoking, non–type-O blood, and male sex. Discrimination was strong at the 36-month time horizon (AUC = 0.75 in both the training and validation cohorts) with good calibration. In the validation cohort, patients in the top 1% of predicted risk had substantially higher PDAC risk (HR, 7.63; 95% CI, 6.85-8.49) compared with average-risk patients. In the UK Biobank, PRIME achieved a 36-month AUC of 0.71 with good calibration.

Conclusions and Relevance

In this validation cohort study, PRIME was a transparent EHR-based model that effectively stratified PDAC risk across diverse US health systems and generalized internationally. Prospective studies should evaluate for EHR-guided PDAC case-finding and integration with blood-based early-detection assays.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is an uncommon cancer with an outsized public health burden. In 2025, it was the third leading cause of cancer-related mortality in the US, and its incidence is rising.^1,2 Despite incremental survival gains over recent decades, outcomes remain poor, largely because most patients are asymptomatic until late stages and present with locally advanced or metastatic disease.^1,2,3 Consequently, most patients are ineligible for operative resection, which remains the only treatment with established curative potential.³

Population-wide screening for PDAC is generally considered infeasible due to low disease incidence.^3,4,5 Guidelines therefore recommend surveillance for selected high-risk groups, including individuals with hereditary cancer syndromes, familial pancreatic cancer, and certain mucinous pancreatic cysts.^4,5,6 These groups have a 5- to 20-fold increase in disease risk and account for up to 20% of PDAC cases.^7,8,9,10 Screening with magnetic resonance imaging (MRI) or endoscopic ultrasonography (EUS) in these populations is associated with increased detection of early-stage disease and improved lead time–adjusted survival.^{5,11,12,13,14}

Early detection of the approximately 80% of pancreatic cancers that arise sporadically remains a critical challenge.¹⁵ Despite promising advances in blood-based biomarkers, existing tests are not sufficiently specific for use in unselected populations.^16,17,18 Recently, a PDAC multistage case-finding approach termed heuriskance has been proposed.¹⁵ The first stage uses an electronic health record (EHR) algorithm to identify individuals at elevated PDAC risk. Biomarker testing is then performed in high-risk patients. Finally, patients with positive biomarker results undergo abdominal imaging. The strategy minimizes false-positive results but provides the gold standard, most costly screening procedure (imaging) to the highest-risk population (identified by clinical and biomarker screening).

A major barrier to the clinical implementation of multistage case-finding strategies has been the absence of validated, deployment-ready risk stratification models. Although several EHR-based models have been proposed,^{19,20,21,22,23,24,25} most were developed in relatively small or homogenous populations and have struggled to generalize to diverse cohorts. Other challenges include poor interpretability of black box machine learning models, a focus on single risk factors (eg, new-onset diabetes), and the inclusion of predictor variables suggestive of an active diagnostic workup (eg, neoplasm of uncertain origin).

To address these limitations, we developed and externally validated the PDAC Risk Model for Earlier Detection (PRIME) model, an interpretable EHR-based model incorporating routinely available PDAC risk factors, using data from 54 US health systems and the UK Biobank.

Methods

This cohort study used deidentified data that is not linkable to identifiable living individuals; per NYU Grossman School of Medicine institutional review board policy, this project was not considered human subjects research and informed consent was not required. This study was reported in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines.²⁶

Data Sources and Setting

We conducted a retrospective cohort study using the Optum Labs Data Warehouse (OLDW), a longitudinal US deidentified EHR and administrative claims database including more than 300 million enrollees from across the US.²⁷ Eligible participants were aged 40 years or older, had at least 1 year of prior EHR observation, and had an index encounter between January 1, 2016, and December 31, 2018, with a recorded body mass index (BMI).

Participants and Cohort Assembly

Health systems in OLDW were randomly assigned to development (23 systems) or US validation (31 systems) cohorts. Individuals with PDAC diagnosed before the index date were excluded.

Outcome

Incident PDAC was defined using International Classification of Diseases, Ninth and Tenth Revisions (ICD-9/10) diagnosis codes recorded after the index date (eTable 1 in Supplement 1). Follow-up extended from the index encounter until PDAC diagnosis, last encounter, or death within 90 days of the last encounter

Variable Selection

Candidate predictor variables were identified from a PDAC phenome-wide association study conducted at NYU Langone Health²⁸ and a focused literature review,^29,30 emphasizing routinely available data, defined as data captured during clinical encounters and documented in structured EHR fields. Predictors were specified a priori to increase clinician trust and interpretability, limit overfitting, and avoid selection of variables reflecting downstream diagnostic evaluation. Candidate predictors included the most recently recorded demographics, health behaviors, cancer history, metabolic and gastrointestinal conditions, laboratory values, and BMI recorded on or before the index date (definitions in eTable 1 in Supplement 1). Race and ethnicity were derived from structured EHR fields. These variables were included to avoid systematic underprediction of risk in Black individuals, who experience higher PDAC incidence.³¹ To minimize redundancy and prevent information leakage, highly collinear variables were collapsed and variables suggestive of an active diagnostic workup or existing indications for screening were removed prior to elastic net regularization, which selected the final predictor set without a prespecified limit on model size (workflow in eTable 2 in Supplement 1).

Model Development

Time-to-event models were trained in the 23 development health systems using Cox proportional hazards with elastic-net regularization (eMethods in Supplement 1). Continuous predictors were scaled in clinically interpretable units, and missing continuous data were median imputed with accompanying binary missingness indicators; ABO blood type was similarly assigned a missingness indicator. Models were developed using 10-fold cross-validation. To support clinical deployment, we refit a conventional Cox model using the selected predictors and missingness indicators, yielding interpretable coefficients and a baseline survival function for absolute risk estimation. Proportional hazards assumptions were assessed using Schoenfeld residual tests. The final model was applied unchanged to all validation cohorts.

Validation and Performance Assessment (OLDW Data)

The model was evaluated without recalibration in the 31 validation health systems. Discrimination was assessed using time-dependent area under the receiver operating characteristic curve (AUC) at 24 and 36 months, and time-independent concordance index (C-index). Calibration was summarized by regression slope and intercept of observed vs predicted risk across deciles at 24 and 36 months, with observed risk calculated using inverse probability of censoring weighting to model censoring.³² PRIME performance was further assessed using hazard ratios and enrichment metrics, including positive predictive value (PPV) and case enrichment (1/PPV), across pragmatic risk strata (top 1%, 5%, and 10% predicted risk) at clinically relevant time horizons (24 and 36 months). Hypothetical downstream imaging volume and PDAC case detection were estimated by applying representative blood-based assay sensitivity and specificity values to the observed PDAC incidence within each PRIME risk stratum, assuming a population of 1 000 000 individuals assessed by PRIME.

Sensitivity Analyses

Sensitivity analyses evaluated model performance after excluding PDAC diagnosed within 6 months of index, excluding individuals with prior abdominal imaging, restricting to participants aged 50 years or older, and stratifying by sex. Additional models excluded race and ethnicity variables, excluded missingness indicators, and incorporated log(time) interaction terms.

UK Biobank Validation

PRIME was further validated in the UK Biobank (UKB). All predictors except lipase were broadly available. Identical predictor definitions and transformations were applied to this cohort when possible; however, medical history was largely derived from a baseline questionnaire rather than ICD codes. Absolute risk was calculated using development-set coefficients and baseline hazard. Performance was assessed using time-dependent AUC and calibration metrics.

Prevalence of Pathogenic Variants Among Model-Defined High-Risk Patients

To identify whether the PRIME model identified individuals with established genetic indications for PDAC screening rather than sporadic risk, we tested for enrichment of pathogenic or likely pathogenic variants, as defined by ClinVar, in National Comprehensive Cancer Network (NCCN) guideline–designated PDAC predisposition genes (ATM, BRCA1, BRCA2, CDKN2A, EPCAM, MLH1, MSH2, MSH6, PMS2, PALB2, STK11, and TP53). Carriers were defined by the presence of a pathogenic or likely pathogenic variant, and enrichment was compared between high-risk (top percentile) and median-risk (45th-55th percentile) groups among UKB participants. The Haldane–Anscombe correction method was used to calculate odds ratios (ORs) whenever a cell count was 0.

Statistical Analysis and Reporting

All analyses were conducted in R (version 4.5.2; R Foundation). Two-sided statistical tests were used with a P value of .05 considered statistically significant; 95% CIs are reported. Data analysis occurred July 2025 to January 2026.

Results

Study Population and Follow-Up

Among 10 478 924 adults from 54 US health systems, 4 859 833 from 23 systems were assigned to the training cohort and 5 619 091 from 31 systems to the validation cohort (Table 1). Mean (SD) follow-up was 5.4 (2.5) years in the training cohort and 3.9 (2.5) years in the validation cohort. Incident PDAC occurred in 14 405 participants in the training cohort and 11 693 participants in the validation cohort, corresponding to incidence rates (IRs) of 54.9 and 53.9 per 100 000 person-years, respectively.

Table 1. Cohort Baseline Characteristics and Clinical Outcomes^a.

Characteristic	No (%)
	Training cohort	Validation cohort
Demographics
No.	4 859 833	5 619 091
Age, mean (SD), y	60.4 (11.7)	60.2 (11.6)
Sex
Female	2 744 558 (56.4)	3 188 414 (56.7)
Male	2 115 275 (43.6)	2 430 677 (43.3)
Race and ethnicity
Asian	103 862 (2.1)	107 774 (1.9)
Black	406 007 (8.4)	585 510 (10.4)
Hispanic/Latino	210 747 (4.3)	258 978 (4.6)
White	4 011 886 (82.7)	4 420 733 (78.7)
Otherb	108 331 (2.4)	246 096 (4.4)
Clinical
BMI, mean (SD)	30.1 (7.02)	30.3 (7.1)
Current smoker	785 331 (16.2)	1835 090 (14.9)
Former smoker	1 254 296 (25.9)	1 385 653 (24.7)
Total diabetes	980 163 (20.2)	1 069 569 (19.0)
O blood type	246 701 (5.1)	290 600 (5.2)
History of acute pancreatitis	43 955 (0.9)	50 202 (0.9)
History of chronic pancreatitis	13 028 (0.3)	16 313 (0.3)
History of any cancer (nonpancreatic)	460 925 (9.5)	514 407 (9.2)
New GI pain <12 mo	195 113 (4.0)	217 889 (3.9)
History of abnormal weight loss	176 997 (3.6)	207 642 (3.7)
Newly recorded diabetes <36 mo	427 858 (8.8)	472 277 (8.4)
Laboratory test results, mean (SD)
HbA1c, %	6.36 (1.45)	6.39 (1.48)
Glucose random, mg/dL	108 (40)	108 (40)
Lipase, U/L	64.5 (93)	75.8 (103)
AST, U/L	24.2 (18.4)	24.5 (18.4)
Missingness
O blood type missing	4 282 263 (88.3)	4 913 971 (87.5)
HbA1c missing	3 122 003 (64.4)	3 542 578 (63.0)
Glucose random missing	1 058 241 (21.8)	982 611 (17.5)
Lipase missing	4 535 863 (93.5)	5 223 167 (93.0)
AST missing	1 366 241 (93.5)	1 333 217 (93.7)
Utilization
Recent abdominal imaging <1 y	269 780 (5.6)	405 977 (5.4)
Outcomes
Follow-up time, mean (SD), y	5.4 (2.5)	3.9 (2.5)
PDAC events	14 405 (0.3)	11 693 (0.2)
PDAC incidence rate (per 100 000 person-years)	54.9 (0.5)	53.9 (0.5)

Abbreviations: AST, aspartate aminotransferase; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); GI, gastrointestinal; HbA_1c, hemoglobin A_1c; PDAC, pancreatic ductal adenocarcinoma.

To convert glucose to mmol/L, multiply by 0.0555.

^a Baseline characteristics and clinical outcomes for the development (training) and validation cohorts. Values are mean (SD) or No. (%); follow-up in years; incidence rate per 100 000 person-years.

^b Other includes all categories in the electronic health record race field not classified as Asian, Black, Hispanic/Latino, or White.

Population Characteristics

In the training cohort, mean (SD) age was 60.4 (11.7) years, with 2 744 558 female individuals (56.4%) and 2 115 275 male individuals (43.6%). The race and ethnicity distribution was 2.1%, Asian, 8.4% Black, 4.3% Hispanic/Latino, 82.7% White, and 2.4% other race and ethnicity. Current smoking was reported by 16.2% and former smoking by 25.9%. Diabetes was present in 20.2%. Prior pancreatitis was uncommon (<1%). Mean (SD) BMI (calculated as weight in kilograms divided by height in meters squared) was 30.1 (7.02) and the mean (SD) HbA_1c levels was 6.36% (1.45%). Overall, 5.6% of patients had abdominal imaging in the year before index. The validation cohort was similar in age (mean [SD], 60.2 [11.6] years, sex (56.7% female, 43.3% male), and clinical characteristics, with slightly greater racial/ethnic diversity (Table 1).

Model and Predictors

Nineteen predictors and associated missingness indicators were selected for the final model (Figure 1; eTable 3 in Supplement 1) with stable feature selection across cross-validation folds (eTable 4 in Supplement 1). Higher PDAC risk was associated with age, male sex, Black race, smoking, pancreatitis, prior cancer, gastrointestinal diagnoses, and markers of hyperglycemia, whereas type O blood was associated with lower risk. The absolute risk equation, baseline hazard values, and predictor coefficients needed for implementation are reported in eTables 5 to 7 in the Supplement.

Figure 1. Forest Plot of Effect Sizes of PRIME Predictor Variables.

Adjusted HRs with 95% CIs for all predictors in the EHR-based Cox model for incident PDAC. Points denote HRs, horizontal bars show 95% CIs on a log scale, and the dashed line indicates HR = 1. Continuous predictors are scaled per model specification (age per 10 years; body mass index per 5 kg/m²; random glucose per 10 mg/dL; HbA_1c levels per 1%; lipase and AST per log-transformed unit), and missing indicator terms reflect coded missingness. AST indicates aspartate aminotransferase; EHR, indicates electronic health record; GI, gastrointestinal; HbA_1c, hemoglobin A_1c; HR, hazard ratio; PDAC, pancreatic ductal adenocarcinoma; PRIME, the PDAC Risk Model for Earlier Detection; T2D, type 2 diabetes.

Training Cohort Performance

In the training cohort, the model yielded an AUC of 0.76 (95% CI, 0.76-0.76) at 24 months and an AUC of 0.75 (95% CI, 0.75-0.75) at 36 months; the time-independent C-index was 0.73 (95% CI, 0.73-0.73). Calibration indicated close agreement between observed and predicted risks (36-month slope of 1.04; 95% CI, 0.96-1.12), intercept of −0.0001 (95% CI, −0.0002 to 0.0001).

Validation Cohort Performance and Risk Stratification

In the US validation cohort, time-dependent AUC was 0.76 (95% CI, 0.76-0.76) at 24 months and 0.75 (95% CI, 0.75-0.75) at 36 months (Table 2); time-independent C-index was 0.74 (95% CI, 0.74-0.74). Calibration was good (36-month slope, 1.02; 95% CI, 0.95-1.09), intercept, −0.0001 (95% CI, −0.0002 to 0]) (Figure 2). Observed PDAC incidence increased with higher risk percentiles (Figure 3). Individuals in the top 1% and 95th to 99th percentile had 7.63-fold (95% CI, 6.85–8.49) and 4.40-fold (95% CI, 4.04–4.79) higher risk of PDAC, respectively, compared with those at average risk.

Table 2. PRIME Model Discrimination, Projected Imaging Volume and PDAC Detection with Hypothetical Biomarker Testing^a.

Time horizon	AUC (95% CI)	Predicted risk percentile, %	Predicted risk threshold, %	PRIME PPV,%	Flagged by PRIME for receipt of biomarker test, no. per 1 000 000	No.
						Flagged by high-sensitivity biomarker for receipt of imaging	PDAC cases found through imaging	Flagged by high-specificity biomarker for receipt of imaging	PDAC cases found through imaging
24 m	0.76 (0.76-0.76)	Top 1	≥0.42	0.69 (0.63-0.77)	10 000	1048	55	538	41
		Top 5	≥0.26	0.42 (0.40-0.45)	50 000	5150	170	2615	125
		Top 10	≥0.20	0.34 (0.33-0.36)	100 000	10 240	270	5180	200
36 m	0.75 (0.75-0.75)	Top 1	≥0.61	0.78 (0.71-0.86)	10 000	1054	62	543	47
		Top 5	≥0.38	0.50 (0.48-0.53)	50 000	5175	200	2640	150
		Top 10	≥0.29	0.42 (0.41-0.44)	100 000	10 030	340	5230	250

Abbreviations: AUC, area under the curve; PDAC, pancreatic ductal adenocarcinoma; PPV, positive predictive value; PRIME, the PDAC Risk Model for Earlier Detection.

^a Time-dependent AUCs for PRIME risk stratification at 24 and 36 months in the US validation cohort, with projected downstream imaging volume and PDAC case detection following application of a hypothetical blood-based biomarker assay after risk stratification with PRIME. All downstream yields are reported per 1 000 000 individuals assessed by PRIME, with the number flagged as high risk determined by the specified predicted risk percentile. For each PRIME risk stratum, projected case detection and imaging volume were calculated as follows: true-positive results (TP) = prevalence × sensitivity × N and false-positive results (FP) = (1 − prevalence) × (1 − specificity) × N, where N denotes the number flagged by PRIME at the specified risk percentile and prevalence represents the observed cumulative incidence of PDAC among PRIME-positive individuals at the given threshold and time horizon. Imaging volume represents TP + FP, corresponding to the total number of individuals with a positive biomarker result. Hypothetical biomarker performance characteristics were 80% sensitivity and 90% specificity for the high-sensitivity assay and 60% sensitivity and 95% specificity for the high-specificity assay.

Figure 2. Scatter Plots of the Pancreatic Ductal Adenocarcinoma Risk Model for Earlier Detection (PRIME) Calibration Over Clinically Relevant Time Horizons in US Validation Cohort.

Points show per-decile observed risks (inverse probability of censoring weighting–adjusted) vs predicted risk; the dashed line denotes perfect calibration. Calibration intercepts and slopes are: 24 months (slope, 1.21; 95% CI, 1.15-1.26; intercept −0.0001; 95% CI, −0.0002 to 0) and 36 months (slope, 1.02; 95% CI, 0.95-1.09; intercept, −0.0001; 95% CI, −0.0002 to 0). Ideal calibration metrics: intercept = 0, slope = 1.

Figure 3. Cumulative Incidence Curves of PDAC by PRIME Predicted Risk Percentile.

Kaplan-Meier cumulative incidence of PDAC in the US validation cohort stratified by baseline predicted-risk percentiles (<45th, 45th-55th [reference], 55th-90th, 90th-95th, 95th-99th, top 1%). Curves are labeled with HRs vs the reference. HR indicates hazard ratio; PDAC, pancreatic ductal adenocarcinoma; PRIME, the PDAC Risk Model for Earlier Detection.

At 24 months, PPV in the top 1% of predicted risk (24-month predicted risk ≥0.42%) was 0.69% (95% CI, 0.63%-0.77%), or 1 case per 144 patients. In the top 5% (≥0.26%) and top 10% (≥0.20%), PPV was 0.42% and 0.34%, or 1 case per 235 and 292 patients, respectively. At 36 months, PPV was 0.78%, 0.50%, and 0.42% in the top 1%, 5%, and 10%, or 1 case per 128, 198, and 236 patients (Table 2). Patients in the top 1% and 10% of predicted risk comprised 6.8% and 33.3% of PDAC cases at 24 months, and 5.8% and 31.5% of cases at 36 months.

Hypothetical Imaging Volume and PDAC Case Detection With Subsequent Biomarker Testing

The hypothetical downstream yield varied by PRIME risk threshold and biomarker assay performance (Table 2). For example, if PRIME were applied to 1 000 000 individuals in a population with PDAC incidence similar to OLDW and the top 1% were flagged as high risk, 10 000 would undergo biomarker testing. A high-sensitivity biomarker assay (hypothetical sensitivity, 80% and specificity, 90%) would be projected to result in 1050 individuals proceeding to abdominal imaging, among whom about 60 would be expected to be diagnosed with PDAC. Alternatively, use of a higher-specificity biomarker assay (hypothetical sensitivity, 60% and specificity, 95%) would be projected to result in 540 individuals undergoing imaging, among whom about 45 would be expected to be diagnosed with PDAC.

International Validation (UK Biobank Data)

We further validated the model in the UK Biobank (N = 498 754). In this younger (mean [SD] age, 56.5 [8.1] years), less racially diverse (2.2% Asian, 1.6% Black, 94.7% White, and 1.5% other race and ethnicity), and lower-risk British population (IR, 44.1 per 100 000 person-years) (eTable 8 in Supplement 1), the model achieved an AUC of 0.71 (95% CI, 0.71-0.71) at 36 months without recalibration, with modest overprediction of risk (36-month slope, 0.87; 95% CI, 0.71-1.03; intercept, −0.0003; 95% CI, −0.0004 to –0.0001).

An analysis of UK Biobank participants with whole genome sequencing data (N = 488 054) showed that 1.1% of individuals, and 3.2% of individuals who developed PDAC, had a NCCN guideline–defined pathogenic germline variant (pGV) for PDAC. pGVs occurred most frequently in ATM, BRCA2, and PALB2. Among individuals in the top percentile of predicted risk, 1.4% carried a PDAC pGV compared with 1.0% at median risk (OR, 1.4; 95% CI, 1.1-1.9) (eTable 9 in Supplement 1).

Sensitivity and Subgroup Analyses

Model performance was stable across sensitivity cohorts and subgroups. Excluding PDAC diagnosed in the 6 months following index yielded a 36-month AUC of 0.74 (95% CI, 0.74-0.74) in the validation cohort; excluding individuals with recent abdominal imaging yielded a 36-month AUC of 0.75 (95% CI, 0.75-0.75); and restricting to participants 50 years or older yielded a 36-month AUC of 0.71 (95% CI, 0.71-0.71). Discrimination was similar in male and female subgroups, with a 36-month AUC of 0.75 (95% CI, 0.75-0.75) in both groups. PRIME was well-calibrated throughout subgroups, consistent with the primary results. A model with time interaction terms (eTable 10 in Supplement 1) had diminished performance in the validation cohort compared with the primary model (36-month AUC, 0.73 vs 0.75). Models without missingness indicators and race or ethnicity variables both showed 36-month AUCs consistent with the primary model (AUC, 0.75; 95% CI, 0.75-0.75). However, compared with the primary model, the race-free model underpredicted risk among Black individuals at 24 months (calibration slope, 1.33 vs 1.07) and 36 months (calibration slope, 1.12 vs 0.90). Coefficients for the race-free model are provided in eTables 11-13 in Supplement 1.

Discussion

In this cohort study of more than 11 million adults across 54 US health systems and the UK Biobank, we developed and externally validated PRIME, an interpretable EHR-based model for incident PDAC using routinely collected clinical data. PRIME may address key limitations of prior PDAC risk models, including limited generalizability, reliance on complex machine learning, and inclusion of variables reflecting active diagnostic evaluation.^{19,20,21,22,23,24,25} Across diverse populations, PRIME showed strong discrimination and excellent generalizability without fine-tuning or recalibration and identified individuals at substantially elevated PDAC risk.

PRIME incorporated a range of established and plausible risk markers for sporadic PDAC, spanning low-prevalence, high-effect size conditions such as pancreatitis³³ and more common risk factors like non–O blood type and type 2 diabetes.^34,35 Personal history of nonpancreatic cancer may capture underlying genetic and environmental susceptibility, whereas demographic predictors such as age, male sex, and Black race reflect PDAC epidemiology.^36,37 Several predictors likely reflect short-term risk or evolving subclinical disease rather than long-term susceptibility; for example, abnormal weight loss has been described as part of a PDAC-associated metabolic dysregulation syndrome that can significantly precede clinical diagnosis.³⁸ Laboratory tests such as aspartate aminotransferase and lipase levels may capture informative testing patterns in addition to physiology; however, model discrimination was maintained when excluding missingness indicators, supporting that PRIME’s performance was not driven solely by diagnostic workup patterns. A previously identified marker of short-term risk, new-onset diabetes, did not contribute sufficient incremental value for inclusion, likely due to collinearity with related metabolic features (HbA_1c levels, weight loss) and limitations in reliably capturing timing of diagnosis in large, heterogeneous EHR datasets.^39,40

A validated EHR-based risk model is timely given advances in circulating biomarkers for PDAC.^16,18 Although PDAC may be detectable for approximately 2 years prior to clinical diagnosis, biomarkers remain insufficiently specific for population-wide screening, motivating their use in an EHR-defined high-risk cohort as part of a multistage case-finding approach.^{16,18,41,42,43} In this context, individuals in the top 1% of PRIME-predicted risk (ie, ≥0.61% 36-month predicted risk) exhibited a more than 7-fold higher PDAC risk than the population mean, comparable to or exceeding that of pGV carriers,^44,45,46,47 for whom screening is recommended and associated with detection of resectable disease and improved survival.

PRIME is designed for implementation in routine clinical care to expand PDAC case-finding to patients at high sporadic risk. The model is intended for passive EHR deployment—for example through automated recalculation every 2 to 3 years using the most recently available clinical data—without requiring clinician input or guiding diagnostic testing. In a multistage case-finding framework, individuals with PRIME risk estimates exceeding a predefined absolute risk threshold would be candidates for downstream biomarker testing. Patients with positive biomarker test results would then undergo imaging, similar to PDAC surveillance strategies for pGV carriers.^48,49 Future studies should assess the impact of this multistage approach on downstream testing burden, stage shift, resectability, and survival.

Strengths and Limitations

This cohort study has several key strengths. PRIME integrates established PDAC risk markers into a deployable risk stratification tool to identify an enriched population for biomarker-based case-finding. To our knowledge, it is the largest PDAC risk prediction study to date, comprising more than 11 million patients and nearly 28 000 patients with PDAC. In contrast to computationally complex, high-dimensional models, PRIME prioritized clinical scalability using only 19 routinely available EHR predictors. The model was developed in a diverse US cohort previously used to develop and validate guideline-endorsed risk equations for cardiovascular disease and kidney failure, and it demonstrated robust international generalizability in the UKB.^50,51 Finally, PRIME identified elevated risk through factors largely independent of germline susceptibility—consistent with low pGV prevalence among high-risk individuals in the UKB—supporting the expansion of case-finding beyond current screening populations.^4,5,6

The PRIME model also has limitations that warrant consideration. Some PDAC risk factors, including family history or timing of type 2 diabetes onset, are poorly captured in EHR data and added insufficient value for inclusion.^39,52 Use of race and ethnicity variables in clinical algorithms remains controversial.⁵³ In PRIME results, race likely reflected social determinants of health rather than biology⁵⁴; however, its removal led to underprediction of risk among Black patients—a population with higher PDAC incidence and more advanced disease at diagnosis—justifying its inclusion.³¹ Despite good model performance, most PDAC cases occurred outside the highest predicted risk strata, underscoring the need for more specific noninvasive biomarkers. Finally, the higher incidence of pancreatic cancer observed in our cohort (approximately 54 per 100 000 person-years compared with approximately 13 per 100 000 person-years in the general population) likely reflected an older, health care–engaged population,¹ consistent with the intended clinical deployment setting.

Conclusions

In this large cohort study, PRIME, an interpretable EHR model for PDAC risk stratification, demonstrated good discrimination and calibration across diverse clinical populations, isolating a high-risk group for prospective evaluation of targeted PDAC case-finding.

Supplement 1.

eMethods. Model Development Details

eTable 1. PRIME Outcome and Predictor Definitions

eTable 2. Variable Selection Workflow

eTable 3. Candidate Predictors and Univariate Coefficients

eTable 4. Feature Selection Stability Across 10-Fold Cross-Validation

eTable 5. PRIME Absolute Risk Equation

eTable 6. Baseline Cumulative Hazard and Survival Estimates

eTable 7. Final PRIME Predictors and Coefficients

eTable 8. Characteristics of the UK Biobank Validation Cohort (N = 498,754)

eTable 9. Germline Pathogenic Variants in UK Biobank Whole-Genome Participants by Risk Percentile (Top Percentile vs 45th–55^th Percentile)

eTable 10. Schoenfeld Residual Tests for Proportional Hazards

eTable 11. PRIME Absolute Risk Equation (Race-Free)

eTable 12. Baseline Cumulative Hazard and Survival Estimates (Race-Free)

eTable 13. Final PRIME (Race-Free) Predictors and Coefficients

Supplement 2.

Data Sharing Statement