Survival Analysis of Young Leukemia Patients

Faculty Advisors: Dr. Keshav P. Pokhrel 4, Dr. Taysseer Sharaf 5 Abstract With cancer as a leading cause of death in the United States, the study of its related data is imperative due to the potential patient benefits. This paper examines the Surveillance, Epidemiology, and End Results program (SEER) research data of reported cancer diagnoses from 1973-2014 for the incidence of leukemia in young (019 years) patients in the United States. The aim is to identify variables, such as prior cancers and treatment, with a unique impact on survival time and five-year survival probabilities using visualizations and different machine learning techniques. This goal culminated in building multiple models to predict the patient's hazard. The two most insightful models constructed were both neural networks. One network used discrete survival time as a covariate to predict one conditional hazard per patient. The prediction rate is nearly 95% for testing datasets. The other network built hazards for discrete time intervals without survival time as a covariate and predicted with lower accuracy, but captured variable effects from initial testing better.