Assessing time-dependent performance of a feature selection method using correlation sharing t-statistics (corT) for heart failure data classification.

Abstract

World Health Organization (WHO) has announced that cardiovascular disease is one of the leading causes of death with approximately 18 million people died per year and this specifically because of heart failure (HF). Many predictive models of HF were proposed using various feature selection methods, but none was found performing the appropriate external validation. The performance of the models in the literature was assessed considering disease status as fixed over time. However, disease status and biomarker values of an individual may change over time thus time-dependent accuracy measures is more appropriate. This paper aims to propose a predictive model of HF mortality using a feature selection method that accounting the relationship between variables (biomarkers or factors) and assess the performance of the model at different time points. A published dataset consists of 299 HF patients collected from Faisalabad Institute of Cardiology and the Allied Hospital in Faisalabad (Punjab, Pakistan), during April-December 2015 is used to illustrate our proposed model. The most important factors are determined using corT statistics on the 80% of the dataset, as training set, and the selected factors are used to fit a logistic regression model as the classifier on the remaining 20% as testing set. The performance of the model is then assessed by calculating the accuracy measures at several prediction times. The dynamic model performance may guide medical decision in collecting the best measurements along the follow-up time period.