New EHR-Based Model PRIME Achieves 0.75 AUC in Predicting Pancreatic Cancer Risk Across 11 Million Patients

Background

Pancreatic ductal adenocarcinoma (PDAC) represents one of the most lethal malignancies in the United States, consistently ranking among the leading causes of cancer-related deaths. The grim prognosis associated with PDAC—five-year survival rates remain below 12%—is largely attributable to the disease’s aggressive biology and the challenge of early detection. Most patients present with advanced-stage disease, when curative surgical resection is no longer feasible. While early detection has been shown to meaningfully improve survival outcomes, the relative rarity of PDAC in the general population has historically rendered population-wide screening programs impractical and cost-prohibitive. This inherent tension between the致命 nature of the disease and the low base rate of incidence has created a significant unmet need in oncology: identifying individuals at elevated risk who would benefit most from intensive surveillance or early-detection strategies.

In this context, the development of validated, generalizable risk prediction tools that leverage readily available clinical data offers a compelling pathway forward. Electronic health record (EHR) systems contain a wealth of structured and unstructured information—demographics, diagnostic codes, laboratory values, medication histories—that can be harnessed to build models capable of flagging individuals at higher-than-average PDAC risk. Such tools, if integrated into clinical workflows, could guide targeted case-finding efforts, optimize resource allocation, and ultimately shift the diagnostic paradigm toward earlier-stage disease identification.

Study Design

The research team conducted a large-scale cohort study using the Optum Labs Data Warehouse, a longitudinal, deidentified US EHR and claims database encompassing adults 40 years of age or older with at least one outpatient clinical encounter between 2016 and 2018. The study design incorporated three distinct cohorts to assess model generalizability: a training cohort drawn from 23 health systems (n = 4,859,833) and a validation cohort from 31 additional health systems (n = 5,619,091). For international validation, the model was further tested in the UK Biobank (n = 498,754).

Predictor variables included demographic characteristics, International Classification of Diseases (Ninth and Tenth Revisions; ICD-9/10) diagnostic codes, and routinely measured laboratory values. The modeling approach employed elastic-net regularization with 10-fold cross-validation to select the most parsimonious yet predictive set of features. Incident PDAC served as the primary outcome, identified through ICD-9/10 diagnostic codes. Model performance was evaluated using time-dependent area under the receiver operating characteristic curve (AUC) and calibration metrics. Data analysis was conducted between July 2025 and January 2026.

Key Findings

The combined study population included more than 11 million adults across the US cohorts. Racial and ethnic distribution reflected the broader EHR landscape: 82.7% White individuals, 8.4% Black individuals, 4.3% Hispanic/Latino individuals, 2.1% Asian individuals, and 2.4% from other racial and ethnic groups.

In the training cohort, with a mean (SD) age of 60.4 (11) years, 14,405 individuals were diagnosed with PDAC over a mean (SD) follow-up of 5.4 (2.5) years, corresponding to an incidence rate of 55 per 100,000 person-years. The validation cohort demonstrated comparable incidence at 54 per 100,000 person-years, with 11,693 PDAC diagnoses over a mean (SD) follow-up of 3.9 (2.5) years.

The resulting model—termed PRIME (PDAC Risk Model for Earlier Detection)—retained 19 predictors following elastic-net regularization. These predictors encompassed clinical, laboratory, and demographic variables. Notably, the retained features included a history of pancreatitis, gastrointestinal disorders, prior cancers, type 2 diabetes, elevated aspartate aminotransferase (AST) levels, smoking status, non-type-O blood group, and male sex. The transparency of the model—its reliance on interpretable, routinely available variables—distinguishes it from more complex black-box approaches and facilitates clinical adoption.

Discrimination performance was strong at the 36-month prediction horizon, with an AUC of 0.75 in both the training and validation cohorts. Calibration was reported as good across risk strata. In the validation cohort, individuals in the top 1% of predicted risk demonstrated a substantially elevated hazard of PDAC compared with average-risk patients (HR, 7.63; 95% CI, 6.85–8.49), underscoring the model’s ability to identify a high-risk subgroup with clinically meaningful risk elevation.

International validation in the UK Biobank yielded a 36-month AUC of 0.71 with acceptable calibration, indicating that PRIME generalizes beyond the US healthcare context despite differences in coding systems, patient populations, and data collection methodologies.

Expert Commentary

The PRIME study represents a notable advancement in the application of machine learning and EHR-derived data to oncology risk stratification. Several methodological strengths merit attention. The use of a parsimonious model with 19 interpretable predictors strikes a pragmatic balance between predictive performance and clinical implementability—a critical consideration for any tool intended for real-world deployment. The multi-cohort validation strategy, spanning both US health systems and an independent international dataset, provides a level of generalizability evidence that is infrequently achieved in predictive modeling studies.

The retention of readily available variables—including basic demographics, diagnostic codes, and common laboratory values—means that PRIME could potentially be calculated at the point of care without requiring specialized testing or invasive procedures. This contrasts with emerging blood-based early-detection assays, which, while promising, face challenges related to cost, accessibility, and false-positive rates in low-prevalence populations.

However, several considerations warrant careful interpretation. First, while an AUC of 0.75 represents meaningful discrimination, it falls short of thresholds typically considered sufficient for population screening. The authors appropriately acknowledge that PRIME is not intended to replace population screening but rather to identify high-risk subgroups for targeted surveillance or further workup. Second, the reliance on ICD diagnostic codes introduces potential misclassification bias, as coding accuracy varies across institutions and clinical settings. Third, the EHR-derived race and ethnicity data may be incomplete or inconsistently recorded, which could affect model performance in underrepresented populations. Fourth, the UK Biobank, while valuable for external validation, represents a volunteer cohort with inherent selection bias, and performance metrics in this cohort may not fully translate to routine clinical populations.

Future directions should include prospective evaluation of EHR-guided case-finding workflows, integration with emerging blood-based early-detection biomarkers (such as circulating tumor DNA or protein panels), and assessment of whether PRIME-driven identification leads to meaningful stage migration and improved survival outcomes in clinical practice.

Conclusion

The PRIME model represents a transparent, interpretable, and generalizable approach to PDAC risk stratification using routinely collected EHR data. Its ability to achieve strong discrimination in both US validation cohorts and the UK Biobank, combined with the clinically meaningful risk elevation observed in the top 1% of predicted risk, positions it as a promising tool for targeted early-detection strategies. While prospective studies are needed to establish the clinical impact of EHR-guided case-finding, PRIME offers a scalable framework that could meaningfully shift the paradigm of pancreatic cancer diagnosis from symptomatic presentation toward earlier, more treatable disease stages.

Funding

The study did not report specific funding information in the available abstract. Readers are referred to the full publication in JAMA Oncology for complete conflict-of-interest and funding disclosures.

References

Mavromatis LA, Zlatanic V, Agarunov E, Sanoba SA, Kluger MD, Horwitz LI, Razavian N, Maitra A, Gonda TA, Grams ME. Development and Validation of a Parsimonious Risk Stratification Model for Pancreatic Cancer. JAMA oncology. 2026-03-26. PMID: 41885821.