Gaurav Bajaj (1), Erin Dombrowsky (1), Qilu Yu (2), Shashank Rohatagi (3), Jeffrey S. Barrett (1)
(1) Laboratory of Applied PK/PD, Department of Clinical Pharmacology and Therapeutics, The Children’s Hospital of Philadelphia, Philadelphia, PA, USA, (2) Westat Inc., Rockville, MD, USA, (3) Piramal Life Sciences, Mumbai, India
Objectives: Pancreatic cancer is the fourth leading cause of cancer related deaths in United States. With most patients diagnosed with pancreatic adenocarcinoma (AC) in advanced stages and few months to survive, the disease is considered largely incurable due to minimal effects of treatment. Sample size in pancreatic cancer trials is usually small and poses a significant challenge for comparing treatment effects across trials. Our objective is to develop a parametric survival model using data from Surveillance, Epidemiology, and End Result (SEER) registry [1], identify relevant covariates to stratify patients in future trials and predict disease outcome.
Methods: Data from 82,251 patients was extracted using site (pancreas) and histology codes (duct, mucinous and monopormphic AC) in the SEER database and refined based on the specific cause of death. Predictors effecting disease outcome were chosen from previous studies and in consultation with clinical expertise. Categorical predictors influencing survival were tested using non-parametric analysis. Both semi-parametric and parametric approaches were used for testing continuous predictors, and multivariate modeling was done with variable reduction based on a manual backward elimination using significance level of α=0.05. Model evaluation was done using Cox-Snell, Martingale, and Score residuals [2]. Analyses were performed using SAS 9.2.
Results: Median overall survival in pancreatic cancer patients was 4 months which is similar to previous studies. Predictors that influence survival were tumor characteristics, therapy, and LN status. Treatments that improved survival outcome were LN removal, surgery of the tumor, chemotherapy and radiotherapy with hazard ratios (HR) 0.63, 0.46, 0.52, and 0.90 respectively. HR for tumor size and LN status were 1.01 and 1.05. SEER data was best fitted by the log-logistic parametric distribution and model selection was based on -2 Log likelihood, Akaike information criterion and residual analysis.
Conclusions: While the SEER dataset lacks granularity in terms of time dependent information, it provides valuable information on tumor size, LN status, and treatment. Parametric analysis shows that the pancreatic cancer data from SEER follows log-logistic regression model. Future efforts include model validation against a dataset from prospective clinical trial and the development of a trial simulation model informed by this and previously published trials.
References:
[1] The Surveillance, Epidemiology, and End Results (SEER) Program of the National Cancer Institute: http://seer.cancer.gov/.
[2] Hosmer, D. W. and S. Lemeshow (1999). Applied survival analysis : regression modeling of time to event data. New York, Wiley.
Reference: PAGE 21 () Abstr 2399 [www.page-meeting.org/?abstract=2399]
Poster: Oncology