Abstract 1644: Pancreatic cancer survival prediction using machine learning and comparing its performance with TNM staging system and prognostic nomograms

Abstract Background: Pancreatic cancer is one of the deadliest cancers with a 5-year survival rate of only 6%. Patients with cancer often want to know how long they have left to live and it’s the first question patients ask after diagnosis. Reliable predictions are very important for achieving more personalized care and better management. However, the accuracy of physicians' clinical predictions of survival is poor. Predicting pancreatic cancer survival is challenging due to different tumor characteristics, treatments and patient populations. In this study, we evaluate the performance of Machine Learning (ML) in survival prediction compared to TNM system and published nomograms. Methods: Pancreatic cancer patients were identified through the Surveillance, Epidemiology and End Results database (SEER). Clinical data of the patients were extracted including: age, sex, race, marital status, tumor site, tumor histology, grade, tumor size, extension, Lymph Node (L.N) involvement, metastasis, cancer sequence, TNM stage, surgery, radiotherapy, chemotherapy, examined (L.N), positive (L.N) and survival months. Patients’ records were randomly divided into a training set (80%) and a validation set (20%). ML algorithms were tested to predict survival at 6, 12 and 24 months as well as TNM staging system and two nomograms. Results: A total number of 22,700 patients were identified through SEER from 2004 – 2013 with median survival of 12 months. The most common primary sites were head of pancreas (62.8%), tail (13.2%), and body (10.7%). Random Forests algorithm achieved better results compared to other models. The trained model yielded Area under the Receiver Operating Characteristic Curve (AUCs) of 86.6% at 6 months, 83.4% at 12 months and 82.2% at 24 months. Compared to TNM staging system which achieved AUCs of 66.6% at 6 months, 65.5% at 12 months and 57% at 24 months. The nomograms achieved AUCs of 71.1% and 55.2% in predicting 1-year-survival. Sensitivity of the trained model were 80%, 75% and 79% at 6, 12 and 24 months, respectively. The most important characteristics which influenced the model prediction performance were: age at diagnosis, surgery and tumor size. Conclusion: Machine learning achieved a much better performance compared to TNM staging system and prognostic nomograms. Improved survival prediction can help in in making better treatment decisions and planning social and care needs. Citation Format: Mohamed H. Osman. Pancreatic cancer survival prediction using machine learning and comparing its performance with TNM staging system and prognostic nomograms [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1644.