Pancreatic Cancer Biomarkers Predictors of Chronic Pancreatitis: Analysis using Machine Learning

Introduction: Several studies have investigatedbiomarkers of pancreatic cancer, which would be useful for ruling out chronic pancreatic pathologies. The objective was to analyze and compare concentrations of plasma and urinary biomarkers of pancreatic cancer in patients with chronic pancreatitis and healthy people through supervised and unsupervised learning. Materials and methods: Analytical and cross-sectional study based on a secondary database. The variables were: chronic pancreatitis, REG1B, REG1A, TFF1 LYVE1, creatinine, and CA19.9. Student t tests, Spearman correlation, principal component analysis (pca), multilayer perceptron and decision tree were used through automatic detection of interactions by Chi-square. Results: The average of biomarkers was higher in patients with chronic pancreatitis. Using multilayer perceptron, the most predictive biomarkers were: CA19.9, REG1B and TFF-1 (efficiency 91%). The decision tree classified TFF-1 and REG1B as predictors (REG1B equal to or less than 12.74ng/ml and TFF1 between 33.11-339ng/ml) and another node with TFF1 greater than 339ng/ml. The principal component analysis generated a component of biomarkers REG1B, REG1A, TFF1 LYVE1 and creatinine, creating a cut-off point of –0.55 for the presence or absence of chronic pancreatitis (sensitivity: 26%, and specificity: 93%) Conclusions: The biomarkers for pancreatic cancer REG1B, REG1A, TFF1 LYVE1, creatinine and CA19.9 are increased in chronic pancreatitis. Supervised learning tools allow efficient prediction and classification of biomarkers, with TFF1 and REG1B being the most relevant. Through unsupervised learning, the combination of REG1B, REG1A, TFF1 LYVE1 and creatinine is highly specific to rule out chronic pancreatic inflammation in healthy people.