Indeterminate main pancreatic duct dilation: an endoscopic ultrasound-based machine learning model to distinguish main duct-intraductal papillary mucinous neoplasm from chronic pancreatitis.

BACKGROUND AND AIMS: Main pancreatic duct dilation occurs in both main-duct intraductal papillary mucinous neoplasm (MD-IPMN) and chronic pancreatitis (CP) but has different clinical implications. We developed an EUS feature-based machine learning model to differentiate MD-IPMN from CP when CT or magnetic resonance imaging is indeterminate. METHODS: We conducted a single-center retrospective study of consecutive patients evaluated at Johns Hopkins Hospital (December 2011-September 2024). Surgically confirmed MD-IPMN and clinically and imaging-diagnosed CP with complete preoperative EUS data were included. Two blinded endoscopists independently scored 21 EUS features. Multiple classifiers were trained with internal validation and class imbalance correction. Performance was assessed using cross-validation and an independent test set, with evaluation of calibration, clinical utility, and feature importance. A prespecified stone-negative subgroup analysis was performed. An online tool was implemented to support point-of-care application and future external validation. RESULTS: Among 123 patients (83 CP; 40 MD-IPMN; mean follow-up 42.67 months), key differentiating features included intraductal nodules (2.4% vs 20.0%; P = .003), ductal stones (77.1% vs 2.5%; P < .001), cyst patterns (P < .001), and parenchymal changes (P = .032). Random Forest test performed best (area under the receiver operating characteristic curve [AUC], 0.946; F1-score, 0.879) with good calibration and net benefit. In the stone-negative subgroup, performance remained robust (F1-score, 0.878; AUC, 0.830), with accuracy, precision, and recall each >0.8. Feature attribution identified ductal stones, diffuse lobularity, diffuse duct dilation, and atrophy as major predictors favoring CP. A web-based interface was developed. CONCLUSIONS: An interpretable EUS-based machine learning model differentiates MD-IPMN from CP-associated duct dilation, including stone-negative cases, and may support standardized evaluation when cross-sectional imaging is inconclusive.