Considerations for the prediction of survival time in pancreatic cancer based on registry data

Semi-parametric and parametric survival models in patients with pancreatic adenocarcinoma (PC) using data from Surveillance, Epidemiology, and End Result (SEER) registry were developed to identify relevant covariates affecting survival, verify against external patient data and predict disease outcome. Data from 82,251 patients was extracted using site and histology codes for PC in the SEER database and refined based on specific cause of death. Predictors affecting survival were selected from SEER database; the analysis dataset included 2,437 patients. Survival models were developed using both semi-parametric and parametric approaches, evaluated using Cox-Snell and deviance residuals, and predictions were assessed using an external dataset from Saint Louis University (SLU). Prediction error curves (PECs) were used to evaluate prediction performance of these models compared to Kaplan-Meier response. Median overall survival time of patients from SEER data was 5 months. Our analysis shows that the PC data from SEER was best fitted by both semi-parametric and the parametric model with log-logistic distribution. Predictors that influence survival included disease stage, grade, histology, tumor size, radiation, chemotherapy, surgery, and lymph node status. Survival time predictions from the SLU dataset were comparable and PECs show that both semi-parametric and parametric models exhibit similar predictive performance. PC survival models constructed from registry data can provide a means to classify patients into risk-based subgroups, to predict disease outcome and aide in the design of future prospective randomized trials. These models can evolve to incorporate predictive biomarker and pharmacogenetic correlates once adequate causal data is established.