AF episodes recognition using optimized time-frequency features and cost-sensitive SVM

• Published in: Australasian physical & engineering sciences in medicine Vol. 44; no. 3; pp. 613 - 624
• Main Authors: Hamil, Hocine, Zidelmal, Zahia, Azzaz, Mohamed Salah, Sakhi, Samir, Kaibou, Redouane, Ould Abdeslam, Djaffar
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.09.2021; Springer Nature B.V
• Subjects: Algorithms; Arrhythmia; Biological and Medical Physics; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Biophysics; Cardiac arrhythmia; Electrocardiography; Fibrillation; Heart rate; Mathematical analysis; Medical and Radiation Physics; P waves; Scientific Paper; Segmentation; Support vector machines; Time-frequency analysis; RR interval; S-transform with CSK; SVM; Atrial fibrillation; Sliding window; Time-frequency features
• ISSN: 2662-4729; 0158-9938; 2662-4737; 1879-5447
• DOI: 10.1007/s13246-021-01005-2

Although atrial fibrillation (AF) Arrhythmia is highly prevalent within a wide range of populations with major associated risks and due to its episodic occurrence, its recognition remains a challenge for doctors. This paper aims to present and experimentally validate a new efficient approach for the detection and classification of this cardiac anomaly using multiple Electrocardiogram (ECG) signals. This work consists of applying Stockwell transform (ST) with compact support kernel (ST-CSK) for ECG time-frequency analysis. The estimation of the atrial activity (AA) is then achieved after analyzing P-waves of the ECG signals for each heartbeat. ECG signals segmentation allows characterizing the AA by making use of its ( t ,  f ) flatness, ( t ,  f ) flux, energy concentration and heart rate variability. The features matrix is employed as an input of the support vector machines (SVM) working in binary and asymmetrical mode with an embedded reject option. The proposed algorithm is trained and then tested using different ECG sources namely two databases provided by PhysionNet (MIT-BIH Arrhythmia, MIT-BIH Atrial Fibrillation) and recorded ECG signals using MySignals HW development platform with raspberry Pi 3 model B + . The used method has achieved 98.46 % and 97.81 % as sensitivity and specificity, respectively. The obtained results confirm that the proposed approach represents a promising tool for Atrial Fibrillation Episodes (AFE) recognition with significant separability between Normal atrial activity and atrial activity with AF even under real and clinical conditions.