Abstract
BACKGROUND AND AIMS: Accurate evaluation of intraductal papillary mucinous neoplasm (IPMN) is necessary to inform clinical decision-making. But it is still difficult to distinguish benign and malignant IPMN preoperatively. This study aims to evaluate the utility of EUS to predict the pathology of IPMN. METHODS: Patients with IPMN who underwent endoscopic ultrasound within 3 months before surgery were collected from six centers. Logistic regression model and random forest model were used to determine risk factors associated with malignant IPMN. In both models, 70% and 30% of patients were randomly assigned to the exploratory group and validation group, respectively. Sensitivity, specificity, and ROC were used in model assessment. RESULTS: Of the 115 patients, 56 (48.7%) had low-grade dysplasia (LGD), 25 (21.7%) had high-grade dysplasia (HGD), and 34 (29.6%) had invasive cancer (IC). Smoking history (OR = 6.95, 95%CI: 1.98-24.44, p = 0.002), lymphadenopathy (OR = 7.91, 95%CI: 1.60-39.07, p = 0.011), MPD > 7 mm (OR = 4.75, 95%CI: 1.56-14.47, p = 0.006) and mural nodules > 5 mm (OR = 8.79, 95%CI: 2.40-32.24, p = 0.001) were independent risk factors predicting malignant IPMN according to the logistic regression model. The sensitivity, specificity, and AUC were 0.895, 0.571, and 0.795 in the validation group. In the random forest model, the sensitivity, specificity, and AUC were 0.722, 0.823, and 0.773, respectively. In patients with mural nodules, random forest model could reach a sensitivity of 0.905 and a specificity of 0.900. CONCLUSIONS: Using random forest model based on EUS data is effective to differentiate benign and malignant IPMN in this cohort, especially in patients with mural nodules.