Predicting the risk of pancreatic cancer in patients with new-onset diabetes based on CT image.

Abstract

e16320 Background: While long-standing diabetes mellitus (LsDM) is an etiologic factor for PC, new-onset diabetes (NoDM) has been considered as a manifestation and harbinger of this cancer as well. Current evidence support that the PC-associated NoDM was mainly caused by disease of the exocrine pancreas, which was different from the typical type 2 DM. Therefore, we hypothesize that, in the NoDM patients, CT image features of “normal” pancreas from PC-associated NoDM patients might be different from that of the new-onset type 2 DM. We sought to develop and validate a risk prediction model to facilitate the distinction between NoDM vs potential early PC-associated NoDM on pancreatic CT imaging. Methods: We retrospectively collected CT imaging from three types of patients at Guangdong Provincial People's Hospital between January 2015 and December 2022: PC patients with pathological diagnosis of pancreatic ductal adenocarcinoma (n=249: 139 patients without DM, 63 patients with NoDM, 47 patients with LsDM), NoDM patients diagnosed with PC within 3 years (n=12), NoDM patients without development of cancer within the next 5 years (n=151). We used four different machine learning algorithms, including support vector machine (SVM), random forest (RF), eXtreme Gradient Boosting (XGB), and generalized linear model (GLM), and gradient boost methods, to build prediction models based on PC pateints and cancer-free NoDM patients. The algorithm with best ROC was used to build a final model, which was validated in the “normal” CT images of NoDM patients prior to their diagnosis of PC within 3 years. Results: In PC cases, the radiomics features exclusive to NoDM-PC cases were selected using a non-parametric test, and 134 radiomics features were retained (86 from arterial and 48 from venous phases, respectively). The RF algorithm showed best prediction skill (AUC of ROC: 0.952). Based on the variable importance measures in the RF algorithm, a risk-scoring model was generated for NoDM patients as a tool to differentiate between PC patients and cancer-free DM patients. As a result of validation, the scores calculated based on the “normal” CT images of NoDM patients prior to their diagnosis of PC within 3 years, were significantly higher than the scores from CT images of NoDM patients who were cancer-free within the next 5 years (P < 0.01). Conclusions: In NoDM patients, even the PC lesion was nonvisible on CT, the radiomics features of “normal” pancreas of PC patients were different from that of the real DM patients. Our model has the potential for identifying individuals with PC in the whole NoDM population.