Feature Extraction Strategies for Predicting Reduced Left Ventricular Ejection Fraction in Chagas Disease Patients.

Cardiovascular problems are the most important manifestation of the Chagas disease (CD), which can cause arrhythmias, heart failure and thromboembolisms. The present work aims to investigate strategies for extracting ECG parameters for predicting left ventricular systolic dysfunction (LVSD), defined as a left ventricular ejection fraction (LVEF) determined by ECO below a given threshold, which may be 35, 40, 45, 50 or 55%. We process a dataset containing single-lead ECG holter signals from 219 CD patients obtained from University Hospital Clementino Fraga Filho - Federal University of Rio de Janeiro, Rio de Janeiro, Brazil. The approach proposes the segmentation of the original signals in intervals during: 5, 10, 15 and 30 minutes. For each scenario, we obtain statiscal measures related to waveform amplitudes and durations, wavelet decomposition coefficients, heart rate variability parameters and non-linear analysis. Then, a set of Machine Learning (ML) algorithms are applied for each scenario to discriminate between LVSD patients and non-LVSD patients. As results, we obtain the highest performance for 15-minute ECG intervals: recall 72% +- 9% and Area under ROC curve 0,75 +- 0,09 for Gradient Boosting. The results indicate the feasibility of using short-term one-channel ECG signals to predict reduced LVEF within CD patients.