A model for predicting degree of malignancy in patients with intraductal papillary mucinous neoplasm

There is no predictive model available to address early stage malignant intraductal papillary mucinous neoplasm (IPMN) including high grade dysplasia (HGD) and pT1a (invasive component≤0.5 cm). The aim of this study was to establish an objective and sufficient model to predict the degree of malignancy in patients with IPMN, which can be easily applied in daily practice and adopted for any type of lesion.

A retrospective cohort study of 309 patients who underwent surgical resection for IPMN was performed. Members of the cohort were randomly allocated to the training or testing set. A detection tree model and random forest model were used for a 3-class classification to distinguish low grade dysplasia (LGD), HGD/pT1a IPMN, and invasive intraductal papillary mucinous cancer (I-IPMC) beyond pT1a.

Of the 309 patients, 54 (17.4%) had early stage malignancy (19 HGD, 35 pT1a), 49 (15.9%) had I-IPMC beyond pT1a, and 206 (66.7%) had LGD IPMN. We proposed a 3-class classification model using a random forest algorithm, and the model had an accuracy of 99.5% with the training set, and displayed an accuracy of 96.0% with the testing set. We used SHAP for interpretation of the model and showed the top five factors (mural nodule size, main pancreatic duct diameter, CA19-9 levels, lesion edge and common bile duct dilation) were most likely to influence the 3-class classification results in terms of interpretation of the random forest model.

This predictive model will help assess an individual's risk for different stages of IPMN malignancy and may help identify patients with IPMN who require surgery.