Therapeutic Research

Abstract

Background：Pancreatic cancer（PC）is one of the most difficult cancers to detect and diagnose early. 　　Objective：This study aimed to identify factors that can serve as predictors for early diagnosis of PC. 　　Methods：This retrospective cohort study of PC and other gastrointestinal cancers used data from the electronic medical record database of the medical data analysis web service“DATuM IDEA®”（TOPPAN INC.）. A neural network model was created with about 500 items；items such as diagnosis and clinical laboratory test results were used as explanatory variables and cancer diagnosis, i.e., PC or non‒PC, was used as the objective variable. As a result, the top 30 items from the 500 items were extracted as relatively important variables affecting the diagnosis of PC. Cox proportional hazard model analyses were performed to assess whether these variables are risk factors for PC. 　　Results：Data of 2951 patients were extracted（PC, n＝465；non‒PC, n＝2486）. The background characteristics of the two groups were very similar. Cox regression analysis identified four factors as significantly related to a diagnosis of PC：acute tonsillitis（hazard ratio［HR］, 4.51）；type 1 insulin‒dependent diabetes（HR, 1.55）；neoplasm of uncertain or unknown behavior of female genital organs（HR, 1.52）；and other respiratory disorders （HR, 3.63）. 　　Conclusion：The factors acute tonsillitis, type 1 insulin‒dependent diabetes, neoplasm of uncertain or unknown behavior of female genital organs, and other respiratory disorders appear to be related to a diagnosis of PC. However, the clinical applicability of these predictors needs to be verified in future studies.