Evaluating the Long-Term Malignancy Risk Prediction Accuracy of Major Guidelines for Pancreatic Cystic Lesions Using Radiology Features.
Authors
Affiliations (6)
Affiliations (6)
- NYU Grossman School of Medicine, 550 First Ave, New York, NY 10016. United States.
- Center for Data Science, New York University, 60 5th Ave, New York, NY 10011United States.
- Perlmutter Cancer Center, New York University Langone Health, New York, 10016, NY, USA.
- Department of Medicine, NYU Langone Health, 550 First Avenue, New York, NY 10016. United States; Perlmutter Cancer Center, New York University Langone Health, New York, 10016, NY, USA.
- Department of Radiology, NYU Langone Health, 550 First Avenue, New York, NY 10016. United States.
- Department of Radiology, NYU Langone Health, 550 First Avenue, New York, NY 10016. United States; Center for Data Science, New York University, 60 5th Ave, New York, NY 10011United States. Electronic address: [email protected].
Abstract
To compare major pancreatic cystic lesion (PCL) guidelines using routinely reported radiology features for long-term malignancy risk prediction. This single-center retrospective study analyzed abdominal radiology reports from 2000 to 2025. A previously validated large language model extracted PCL features from radiology reports to assign risk categories per ACR, Kyoto, and European guidelines. The outcome of interest was malignancy, including pancreatic ductal adenocarcinoma and high-grade dysplasia. Patients with concurrent malignancy within 6 months were excluded . Five-year predictive performance was assessed using AUROC, AUPRC, and Harrell's C-index. Inter-guideline concordance and malignancy risk following 5 years of imaging stability (per Kyoto criteria) were evaluated. The final cohort included 13,269 patients (median age, 69 years; 62% female); 209 (1.6%) developed malignancy over 5 years. Kyoto guidelines demonstrated slightly higher AUROC (0.655) compared with the ACR (0.639) and European (0.623) guidelines (p < 0.001). High-risk classification was fully concordant across guidelines. Worrisome feature assignment varied substantially (ranging from 308 to 507 patients), though 5-year cancer incidence remained similar (10.3-11.7%). Following 5 years of imaging stability (n = 1,830), 8-year cancer incidence was 0.71% overall, 0.46% among patients meeting Kyoto discontinuation criteria, and 1.57% among those who did not. Major pancreatic cyst guidelines demonstrated modest long-term predictive performance using radiologic features alone. All guidelines showed strong concordance for high-risk lesion identification. Among patients with low-risk cysts demonstrating 5 years of imaging stability, subsequent malignancy risk remained low, informing future individualized surveillance strategies.