Machine learning-based identification of paroxysmal atrial fibrillation from Sinus Rhythm ECG

Abstract

Atrial fibrillation (AF) is a common cardiac arrhythmia which is associated with impaired quality of life and increased mortality. Paroxysmal AF (PAF) is a subtype of AF which is presented spontaneously and terminated usually within 48 hours. Because of the nature of this arrhythmia many patients are not aware of it and sometimes they don’t present any symptom. AF is detected when presenting an absence of P-waves for longer than 30s or with irregular heart rhythm. In this work, we aim to detect subtle signs of atrial abnormality using P-waves metrics that can indicate the patients with PAF history during the sinus rhythm (SR) without presenting any visible change on ECG. We collected 10 min ECG recordings in lead X and Y during the sinus rhythm of 70 patients with PAF history and 59 healthy. For each subject we extracted beat-to-beat P-waves and we calculated some conventional metrics as well as novel ones related with P-wave’s integral and slope before and after applying a time scale factor to eliminate heart rate dependence. Due to the plethora of extracted features we tried to reduce set’s dimension using feature selection (FS) methods. We observed statistical differences among the examined features of cohorts mostly in lead X and we tested several combinations based on FS methods. We achieved a maximum classification accuracy of 95% which is state of the art to the best of our knowledge using a feature set of integral and slope features of the lead-X signal. Our results were achieved using Random Forests to identify patients with PAF history from healthy ones during the SR. As a result of this work, the medical assessment in a routine clinical examination can be facilitated in a non invasive and inexpensive way as well as the quality life of patients minimising the stroke risk due to the early AF detection.