Risk stratification of cardiovascular patients using a novel classification tree induction algorithm with non-symmetric entropy measures

Abstract

Risk stratification allows clinicians to choose treatments consistent with a patient's risk profile. Risk stratification models that integrate information from several risk attributes can aid clinical decision making. One of the technical challenges in developing risk stratification models from medical data is the class imbalance problem. Typically the number of patients that experience a serious medical event is a small subset of the entire population. The goal of my thesis work is to develop automated tools to build risk stratification models that can handle unbalanced datasets and improve risk stratification. We propose a novel classification tree induction algorithm that uses non-symmetric entropy measures to construct classification trees. We apply our methods to the application of identifying patients at high risk of cardiovascular mortality. We tested our approach on a set of 4200 patients who had recently suffered from a non-ST-elevation acute coronary syndrome. When compared to classification tree models generated using other measures proposed in the literature, the tree models constructed using non-symmetric entropy had higher recall and precision. Our models significantly outperformed models generated using logistic regression - a standard method of developing multivariate risk stratification models in the literature.