Avaliação do ritmo cardíaco em eletrocardiogramas de curta duração utilizando análise dos intervalos RR e aprendizado supervisionado

Abstract

Atrial fibrillation is a condition that often does not show itself through symptoms and is strongly related to infarction and sudden cardiac death. This work aims at developing an algorithm that differentiates atrial fibrillation rhythm from noise, normal and other rhythms in single short ECG leads collected by a mobile device. A total of 36 features were collected mostly from the sequence of beat-to-beat intervals. Neighborhood component analysis (NCA) feature selection technique was applied, and several supervised learning algorithms were compared and optimized using cross validation approach. Performances were compared with an index F1 that considers both sensitivity and specificity. NCA allowed selecting 11 features. The classifier based on support vector machines gave the best overall result (F1 = 72,9%), were the best performance occurred for the atrial fibrillation class (F1 = 82,5)