Prediction of cardiac arrest recurrence using ensemble classifiers

Abstract

Inability of a heart to contract effectually or its failure to contract prevents blood from circulating efficiently, causing circulatory arrest or cardiac arrest or cardiopulmonary arrest. The unexpected cardiac arrest is medically referred to as sudden cardiac arrest (SCA). Poor survival rate of patients with SCA is one of the most ubiquitous health care problems today. Recent studies show that heart-rate-derived features can act as early predictors of SCA. Addition of angiographic and electrophysiological features can increase the robustness of the prediction system. Early warning has the capability of saving many lives. Risk of recurrent terminal cardiac arrest is high for out-of-hospital survivors. Foregoing studies indicate that recurrent cardiac events are time dependent and, while in clinical follow-up, are highly probable, predominantly in early phase. In this paper, we observe the changing risk of and changing influence of various clinical, angiographic and electrophysiological parameters on subsequent cardiac arrest recurrence with time. Various medical and synthetic datasets such as ECG dataset from PhysioNet, Pima Indian Diabetes dataset from UCI Machine Learning Repository and gene expression dataset from GEO are used, which are unique as compared with related works. Various classifiers such as LogitBoost with simple regression function, random forest and multilayer perceptron are used for recurrence risk prediction. Collection of these classifiers together forms the ensemble classifiers. Classifiers are compared based on various measures like accuracy and precision. Based on the classification, risk scores are calculated using logistic regression with backward elimination. The proposed method is used for final risk estimation. The same datasets are used for risk score calculation model development. Experimental results are found to be encouraging.