Multiclass Classification of Cardiac Rhythms on Short Single Lead ECG Recordings using Bidirectional Long Short-Term Memory Networks

The recognition of cardiac rhythms is a topic of great relevance, particularly using short single-lead ECG recordings, due to its potential to early detect cardiovascular diseases and take actions quickly to preserve peoples wellbeing. Among cardiac arrhythmias, atrial fibrillation is the most common sustained cardiac arrhythmia, with significant mortality and morbidity rates. Several approaches have been conceived to identify cardiac rhythms, from the comparison of heart rate with adaptive and fixed thresholds to the application of deep and machine learning techniques. In this work, the classification of three cardiac rhythms (normal, atrial fibrillation, and other arrhythmias), as well as the identification of noise recordings, were performed using bidirectional LSTM networks that exploit the ECG signal (a representation of the cardiac electric activity) and time series containing information about the auricular and ventricular activities. A Monte Carlo 10-fold cross-validation of 10 iterations was performed to assure the generalization of the classifiers and the replicability of the results. An average accuracy of 77.97% was obtained to recognize the four classes but increase up to 85.95% when noise recordings were left out of the classification process. Moreover, micro F1 scores of 89.96%, 79.23%, and 79.77% were obtained for normal rhythm, atrial fibrillation, and other arrhythmias, respectively. The imbalance of classes and the characteristic patterns of normal rhythm and atrial fibrillation were the main factors associated with these performances.