ConvNet: 1D-Convolutional Neural Networks for Cardiac Arrhythmia Recognition Using ECG Signals Amine Ben Slama 1* , Hanene Sahli 2 , Ramzi Maalmi 1 , Hedi Trabelsi 1 1 Laboratory of Biophysics and Medical Technologies, LR13ES07, University of Tunis El Manar, ISTMT, Tunis 1006, Tunisia 2 Laboratory of Signal Image and Energy Mastery, SIME, LR13ES03, University of Tunis, ENSIT, Tunis 1008, Tunisia Corresponding Author Email: amine.slama@istmt.utm.tn https://doi.org/10.18280/ts.380617 ABSTRACT Received: 10 April 2021 Accepted: 25 August 2021 In healthcare, diagnostic tools of cardiac diseases are commonly known by the electrocardiogram (ECG) analysis. Atypical electrical activity can produce a cardiac arrhythmia. Various difficulties can be imposed to clinicians e.g., myocardial infarction arrhythmia via the non-stationarity and irregularity heart beat signals. Through the assistance of computer-aided diagnosis methods, timely specification of arrhythmia diseases reduces the mortality rate of affected patients. In this study, a 1 Lead QRS complex -layer deep convolutional neural network is proposed for the recognition of arrhythmia datasets. By the use of this CNN model, we planned a complete structure of the classification architecture after a pre-processing stage counting the denoising and QRS complex signals detection procedure. The chief benefit of the new proposed methodology is that the automatically training the QRS complexes without requiring all original extracted ECG signals. The proposed model was trained on the increased ECG database and separated into five classes. Experimental results display that the established CNN method has improved performance when compared to the state-of-the-art studies. Keywords: cardiac arrhythmia disease, ECG data, QRS complex signals, classification, conventional neural network 1. INTRODUCTION Cardiac arrhythmia (Ca) presents the major causes of human death. In 2020, Ca still accounts for 20-31% of all deaths of the worldwide. Due to its increased frequency and incidence, sudden cardiac (SC) death is the greatest shared cause of death [1] accounting for approximately 25million deaths every year with SC. Via ventricular fibrillation being the final underlying mechanism, the majority of deaths are unwitnessed. The routine tool-focused analysis paradigm presents inefficient method due to dealing with large amount of heterogeneous information, and lacks serious examination and medical judgment to attain appropriate precision in diagnosis. This analysis would be better consistent, instinctive, and mostly low-cost scheme for screening and control. Thus, suitable medical appraisal can be related to the use of computer aided diagnosis systems. Being a non-stationary electrophysiological signal, electrocardiogram (ECG) remains the electrical activity of heart. This technique is employed for the heart-beats’ regularity measure for pathological patterns, and even other conditions such mental stress to control the heart. Generally, ECG cycle consists of the P-QRS-T waves as shown in Figure 1. Depending the ECG signal, the main characteristics is involved in the frequency, latency and amplitude ranges of 0.5-45Hz, 0.06-0.08s and 0.1-0.2 mV respectively [2, 3]. The QRS complex presents a ventricular depolarization of amplitude (1 mV) and duration time (0.06-0.12 s). Evaluation of the ECG signal act is focused on a subjective judgment of different aspects: noise, heart beat irregularities... Advanced approaches can be used to classify ECG data according to different pathological levels. The aim of such procedures is to distinguish different Ca quality through significant measures for expert support in the diagnosis method [4]. Global categorization accuracy depends on filtering and characterization strategies in ECG analysis process. Feature extraction from ECG signals is a complex task due to the representation variability: peak amplitudes distribution, P- QRS-T complex, peak intervals. Supra ventricular Tachycardia (SVT) and Atrial Fibrillation (AF) cardiac- diseases are confused with atrial flutter in final decision for atrial arrhythmias identifying. Deep learning methods (DL) has been widely applied for classification and prediction purposes in diverse domains which, feature extraction step can automatically made. Recently, DL algorithms are being established sharply with an important effect on the classification accuracy for a wide variety of medical tasks. Modern computer aided systems such DLs to recognize arrhythmia of measured ECG signal for continuous heart cost reduction to improve the identification quality. Many studies have been presented in literature using ECG data processing [5]. Analysis, filtering process [6], and classification scheme [7, 8] of ECG samples exhibit some limits of the algorithms that cannot permit clinician to recognize different cardiac diseases. Due to diversity of amplitudes in the ECG signals, different techniques do not present efficient detection of these anomalies using both time and frequency domains [7]. The CNN is considered a state-of- the-art tool for the classification of arrhythmia, and it has been studied with several variations such as 1-dimensional, 2- dimensional, or the combination of both [8, 9]. According to Xiao et al. [8], a novel arrhythmia classification method Traitement du Signal Vol. 38, No. 6, December, 2021, pp. 1737-1745 Journal homepage: http://iieta.org/journals/ts 1737