Feature Fusion for Multi-Class Arrhythmia Detection Using Focalbased Deep Learning Architecture

Cardiac arrhythmia is a substantial health condition that can result in serious consequences, ranging from discomfort to life-threatening events. To diagnose this heart disease, healthcare practitioners manually analyze the ECG recordings, which can be inefficient given that it is timeconsuming, especially for long ECG signals, and prone to error due to noise. Therefore, machine learning has shown remarkable results in arrhythmia diagnosis, but it also has some limitations such as the hand-crafted feature selection step, and the extended training time. Additionally, the lack of labeled ECG data reduces the accuracy of the classification task. To overcome these limitations, a Deep Learning architecture for the classification of abnormal heartbeats is proposed in this paper. The data is first augmented using the SMOTE technique. For feature extraction, we use Residual Network to extract spatial features, and the Orthogonal Matching Pursuit (OMP) algorithm to extract time-frequency features. The proposed method was tested on two open-access databases: the MIT-BIH arrhythmia database and the MIT-BIH NSR database with leads MLII and V5, for the classification of eight types of heartbeats. The results indicate an average classification recall, precision, F1-score, and overall accuracy of 94.04%, 95.55%, 94.77%, and 98.99%, respectively.