Predicted Prognosis of Patients with Pancreatic Cancer by Machine Learning.

Abstract

Pancreatic cancer remains a disease of high mortality despite advanced diagnostic techniques. Mucins (MUC) play crucial roles in carcinogenesis and tumor invasion in pancreatic cancers. MUC1 and MUC4 expression are related to the aggressive behavior of human neoplasms and a poor patient outcome. In contrast, MUC2 is a tumor suppressor, and we have previously reported that MUC2 is a favorable prognostic factor in pancreatic neoplasia. This study investigates whether the methylation status of three <i>mucin</i> genes from postoperative tissue specimens from patients with pancreatic neoplasms could serve as a predictive biomarker for outcome after surgery. We evaluated the methylation status of MUC1, MUC2, and MUC4 promoter regions in pancreatic tissue samples from 191 patients with various pancreatic lesions using methylation-specific electrophoresis. Then, integrating these results and clinicopathologic features, we used support vector machine-, neural network-, and multinomial-based methods to develop a prognostic classifier. Significant differences were identified between the positive- and negative-prediction classifiers of patients in 5-year overall survival (OS) in the cross-validation test. Multivariate analysis revealed that these prognostic classifiers were independent prognostic factors analyzed by not only neoplastic tissues but also nonneoplastic tissues. These classifiers had higher predictive accuracy for OS than tumor size, lymph node metastasis, distant metastasis, and age and can complement the prognostic value of the TNM staging system. Analysis of epigenetic changes in <i>mucin</i> genes may be of diagnostic utility and one of the prognostic predictors for patients with pancreatic ductal adenocarcinoma.