Deep Belief Networks and Bayesian Networks for Prognosis of Acute Lymphoblastic Leukemia

Abstract

Cancer is one of main non-communicable diseases. Acute Lymphoblastic Leukemia (ALL), a type of white blood cancer, is one of the most common pediatric cancers. Analysis of cancer prognosis is necessary to determine the proper treatment for each patient. However, cancer data analysis is challenging because multiple risk factors may influence the prognosis of cancer, including gene and clinical condition of patient. This study aims to develop prediction model for cancer prognosis using clinical and gene expression (microarray) data. In this research, manifold learning is applied to microarray data to reduce its dimension, then two Deep Belief Network (DBN) models for both clinical and microarray data are trained separately. Probabilities obtained from Clinical DBN model and Microarray DBN model are integrated using softmax nodes on Bayesian Network structure. Based on various experiments, the best integration model obtained is DBN+BN 32 with prediction accuracy 84.2% for 2-years survival, 70.2% for 3-years, 68.4% for 4-years, and 73.7% for 5-years. This prediction model can be used in cancer analysis and help doctor to decide proper treatment for patient.