An approach to predict the survival time of Acute Lymphoblastic Leukemia (ALL) patients

Abstract

Survival time prediction is the task of predicting the length of time that a particular patient will survive. Accurate survival time predictions can support doctors on selecting treatments and planning futures. This task di ers from the standard survival analysis methods, which focus on population-based studies and tries to identify the prognostic factors and/or analyze the average survival times of di erent groups of patients. The objective of our research work, survival time prediction, is di erent: to nd the most accurate model to predict the survival times for each individual childhood acute lymphoblastic leukemia cancer patient. We view this problem as a regression problem, where we try to map the features of each individual patient to his/her survival time. As the relationship between survival time and features is not understood, we consider various ways to learn these prediction models from patents' historical records. This is a challenging task due to the presence of out- liers, irrelevant features, and missing class labels. This dissertation describes our approach for overcoming these challenges and producing techniques that can pre- dict the survival times. We conduct our experiments on a data set of 512 patients, including 421 censored patients (i.e., patients whose actual survival time is not known). Our approach consists of two major steps. In the rst step, we apply various grouping methods to segregate the patients into smaller populations. In the second step, we apply various regression methods to each sub-population we obtained from the rst step. Our experiments show that the multiple linear regression and the support vector regression are e ective: each predictor can achieve an average cross vali- dated relative absolute error lower than 0.30. We also use our prediction models to classify each patient into \short survivor" versus \long survivor" where the classi - cation boundary is the average survival time of the entire population; here, we show that most of the prediction models can achieve at least 70% accuracy. These exper- imental results verify that we can e ectively predict childhood acute lymphoblastic leukemia cancer patients' survival times using machine learning approaches.