Development and validation of two nomograms to predict prognostic factors for patients with pancreatic cancer: A population-based retrospective study

Background: Pancreatic cancer (PC) patients’ overall survival rate (OS) is associated with multiple factors. The prognosis is generally poor, and the overall survival rate is extremely low. Although there have been numerous prognostic studies on the clinical or genomic characteristics of pancreatic cancer patients, currently there is no research combining clinical features of patients obtained from the SEER database with tumor genomic features obtained from the TCGA and GTEx database to explore the prognostic factors of PC patients. Method: We obtained 2697 patients from the Surveillance, Epidemiology, and End Results (SEER) database between 2010 and 2015, and also downloaded RNA expression and clinical information of PC from TCGA and GTEx. We used LASSO and one-way and multivariate Cox regression analysis to screen out independent prognostic factors and candidate genes for PC to construct nomograms for PC patients. Receiver operating curve (ROC), calibration curve, decision curve analysis (DCA) and Kaplan-meier (K-M) survival curves were used to evaluate the predictive accuracy to evaluate the predictive accuracy, discrimination, and clinical validity of the nomogram. Results: By Lasso and multifactorial Cox regression analysis, we selected age, AJCC N stage, chemotherapy, tumor stage, lymph node dissection, and primary site surgery information to develop the first prognostic model for PC patients. In the development and validation test set, the AUC and C-index of the OS prediction model were greater than 0.7, indicating good predictability of the model. The calibration curves of the two models overlapped with the diagonal. Decision analysis curves (DCA) indicated that the clinical benefit of these models was higher than any single factor. We analyzed three gene datasets for enrichment and found that the gene enrichment was mainly focused on genes related to metabolism and cell division. Then we performed univariate and LASSO analysis on the related genes to obtain significance for ANLN, TGM2, ADAMTS12, and SERPINB5 to build a prognostic model. We used KM survival curves for risk stratification of ANLN and SERPINB5. To assess the prognostic power of the models, we calculated the area under the curve (AUCs) of the ROC curves over time, which were all greater than 0.67 in TCGA and 0.7 in GSE183795. Conclusion: It is verified that our OS prediction curves for PC patients are reliable and can effectively predict the future OS of PC patients at 1, 3, and 5 years, which is a guide for clinical prognosis estimation and treatment of PC patients. Finally, by comparing the two nomograms, we found that the nomogram built for the clinical features of pancreatic cancer had better predictive accuracy than the gene expression nomogram.