Electrocardiogram signal classification for automated delineation using bidirectional long short-term memory. (2021)

• Record Type: Journal Article
• Authors: Nurmaini, Siti; Tondas, Alexander Edo; Darmawahyuni, Annisa; Rachmatullah, Muhammad Naufal; Effendi, Jannes; Firdaus, Firdaus; Tutuko, Bambang
• Is Part Of: Informatics in medicine unlocked. Volume 22(2021)
• Publication Date: 2021
• Subjects: Automated detection -- Deep learning -- ECG delineation -- Long short-term memory (LSTM) -- Recurrent neural networks (RNNs); Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2020.100507
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: Analysis of electrocardiogram (ECG) signals is challenging due to the complexity of their signal morphology. Any irregularity in a cardiac rhythm can change the ECG waveform. A reliable machine learning model is developed here to provide substantial input to cardiologists and help confirm their diagnoses. To achieve high diagnostic accuracy, nearly all ECG analytics tools require records of the positions and morphologies of various segments of P-waves, QRS complexes, and T-waves to be kept in ECG records. However, analyzing such vast amounts of ECG records is not always easy. In most cases, it is challenging and highly time-consuming. Hence, an in-depth investigation regarding automatic ECG signal delineation is necessary. This paper proposes an automated delineation algorithm for ECG waveform signals that utilizes recurrent neural networks (RNNs) with bidirectional long short-term memory (LSTM) architecture. This delineation process consists of four steps: noise cancellation, ECG waveform segmentation, ECG signal classification in four classes (P-wave, QRS complex, T-wave, and isoelectric line) and model evaluation. The classification is conducted based on time duration, and each waveform is determined by using annotated data from the well-known QT database (QTDB). The results show that the proposed model produces satisfactory performance for the four classes in terms of average accuracy, sensitivity, specificity, precision, and F1-score, with values of 99.64%, 98.74%, 99.75%, 98.81%, and 98.78%, respectively. The proposed model is validated with abnormal ECG signals from the QTDB, i.e., MIT-BIH Arrhythmia, MIT-BIH ST Change, MIT-BIH Supraventricular Arrhythmia, European ST-T, and MIT-BIH Long-Term ECG. The results show that bidirectional LSTM can delineate ECG signals from QTDB in both normal and abnormal conditions. The proposed delineation method could be utilized in potential applications following further investigation.