Machine learning for post-acute pancreatitis diabetes mellitus prediction and personalized treatment recommendations

Abstract

Abstract Post-acute pancreatitis diabetes mellitus (PPDM-A) is the main component of pancreatic exocrine diabetes mellitus. Timely diagnosis of PPDM-A improves patient outcomes and the mitigation of burdens and costs. We aimed to determine risk factors prospectively and predictors of PPDM-A in China, focusing on giving personalized treatment recommendations. Here, we identify and evaluate the best set of predictors of PPDM-A prospectively using retrospective data from 820 patients with acute pancreatitis at four centers by machine learning approaches. We used the L1 regularized logistic regression model to diagnose early PPDM-A via nine clinical variables identified as the best predictors. The model performed well, obtaining the best AUC = 0.819 and F1 = 0.357 in the test set. We interpreted and personalized the model through nomograms and Shapley values. Our model can accurately predict the occurrence of PPDM-A based on just nine clinical pieces of information and allows for early intervention in potential PPDM-A patients through personalized analysis. Future retrospective and prospective studies with multicentre, large sample populations are needed to assess the actual clinical value of the model.