Endoscopic ultrasound and clinical features for differentiating autoimmune pancreatitis and pancreatic ductal...

Differentiating autoimmune pancreatitis (AIP) from pancreatic ductal adenocarcinoma (PDAC) remains a diagnostic challenge. Although there is limited evidence that EUS is useful for diagnosing AIP, its full potential has yet to be leveraged in practice. This study aimed to develop an interpretable, accurate, and clinically applicable multimodal AIP diagnostic model by combining EUS features with clinical parameters.

This was a retrospective study of 577 patients attending 6 centers in China who underwent EUS and were diagnosed with AIP or PDAC between January 2013 and April 2024. AIP and PDAC cases were randomly divided 1:1 into primary (n = 289) and validation (n = 288) cohorts. LASSO regression with 5-fold cross-validation was used to determine the optimal clinical (n = 5) and EUS (n = 17) features discriminating AIP from PDAC. Multivariable logistic regression analysis was performed to develop a predictive model, which was assessed in the primary and validation cohorts using receiver operating characteristic area under curve (AUC) and calibration curves. Comparator models based on clinical and EUS features alone were also evaluated using the same framework. Logistic regression coefficients were scaled to develop a clinical nomogram.

Most clinical and EUS features differed significantly between patients with AIP and PDAC. Patients with AIP were slightly younger, were much more likely to have elevated immunoglobulin G4 (IgG4), and their ultrasound features were more consistent with a diffuse pancreatic parenchymal pathology. LASSO regression identified 16 nonzero coefficients, from which the top 4 features with the largest, noncollinear coefficients (elevated IgG4, elevated CA19-9, vascular involvement, and diffuse hypoechogenicity) produced a highly discriminative and well-calibrated model for AIP, with an AUC of 0.994 (95% CI, 0.986-0.999; 97.2% accuracy, 96.7% sensitivity, 97.8% specificity) for the primary cohort and an AUC of 0.984 (95% CI, 0.972-0.994; 95.5% accuracy, 93.5% sensitivity, 97.8% specificity) for the validation cohort. The integrated model showed more balanced diagnostic performance than clinical-only or EUS-only models, particularly with respect to validation sensitivity.

Combining EUS features with clinical and serological variables yields more balanced diagnostic performance for differentiating AIP from PDAC than either modality alone. Given the retrospective design, risk of overfitting, and lack of independent external validation, prospective validation is required before routine clinical implementation.