Automated atrial arrhythmia classification using 1D-CNN-BiLSTM: A deep network ensemble model

Abstract

Atrial arrhythmias are more frequently encountered arrhythmias with a significant impact on mortality and morbidity. Electrocardiogram (ECG) consisting of P, Q, R, S, and T waves is a novel, cost-effective diagnostic tool for detecting atrial arrhythmias. Discrimination of atrial arrhythmias from one another is quite difficult and time-consuming in clinical practice, which leads to false positive diagnosis and treatment. Furthermore, atrial anomalies like Atrial Fibrillation (AF), Atrial Flutter (AFL), and Atrial Tachycardia (AT) often co-occur in hospitalized individuals with cardiac conditions. Therefore, the present work proposed an automatic classification technique for differentiating atrial arrhythmias such as AT, AF, AFL, and Sinus Tachycardia (ST) from normal Sinus Rhythm (SR). The automated classification uses a one-dimensional Convolutional Neural Network (1D-CNN) and Bidirectional Long Short-Term Memory (BiLSTM) model ensemble. While the Bi-LSTM network handles the classification of atrial arrhythmias, the 1D-CNN architecture is responsible for automated feature extraction. Data from Lead-II obtained from Chapman University and Shaoxing People’s Hospital (CUSPH) were utilized for dataset preparation. The dataset undergoes preprocessing to address missing values, segment the data, and augment it to ensure balance across all classes. We used a 10-fold cross-validation methodology to evaluate the model’s efficacy. Our model attained a 94% accuracy across cross-validation and testing datasets when classifying atrial arrhythmias. The current study includes all the arrhythmias originating in the atria and shows the best performance in the state-of-the-art methods, considering atrial arrhythmias as one of the classes. Thus, this method is reliable for precisely diagnosing atrial arrhythmias in real-time clinical applications.