Diagnóstico clínico automatizado a partir do uso de métodos de análise multivariada aplicados a sinais de eletrocardiograma

Abstract

This work presents a study of methods applied to pattern recognition of heart rate variability parameters obtained from electrocardiogram (ECG) for the aid of automated clinical diagnosis of various diseases associated with the heart, using multivariate statistical methods and computational machine learning. For this purpose, the electrocardiograms signs of 137 volunteers clinically diagnosed with normal sinus rhythm (NSR), with n = 54 individuals, which will represent the control group, and two clinical conditions formed by individuals with congestive heart failure (CHF), with n = 29 individuals or suppression of cardiac arrhythmia (CAST), with n = 54 individuals, considering these two clinical conditions, such as the case groups. All these signals were obtained from the PhysioNet, which covers a set of real biomedical signals, open source software and from studies consolidated in the literature. A procedure for obtaining characteristic variables of ECG tachograms was described, these variables were modeled by classification approaches of Discriminant Analysis data by Partial Least Squares regression (PLS-DA) and Artificial Neural Networks (ANN), aiming at the diagnosis of two clinical conditions when compared with a control group. Data matrices of variables associated with the time domain, frequency domain and obtained by non-linear methods were considered separately, each one, and all of these in a single data matrix, of statistical parameters associated with heart rate variability. The figures of merit showed that there is a pattern in the behavior of the tachogram parameters that may be used for clinical diagnostic aid. Both congestive heart failure and the classification and prediction of samples belonging to the cardiac arrhythmia suppression were satisfactorily obtained, with an area under the ROC curve close to 0.9. The PLS-DA model demonstrated the best data classification results, where congestive heart failure was diagnosed with rates of 90.9% of sensitivity and selectivity of 85.7% and suppression of cardiac arrhythmia was predicted with rates of 75.0% of sensitivity and 100.0 % of selectivity, suggesting that clinical diagnosis assisting real time and a personalized prognosis can become a reality that will contribute positively to medical practice.