Automated classification of five arrhythmias and normal sinus rhythm based on RR interval signals

• Published in: Expert systems with applications Vol. 181; p. 115031
• Main Authors: Faust, Oliver, Acharya, U. Rajendra
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.11.2021; Elsevier BV
• Subjects: Algorithms; Arrhythmia; Arrhythmia detection; Bradycardia; Cardiac arrhythmia; Computer aided diagnosis; Deep learning; Fibrillation; Flutter; M-health; Machine learning; Residual Neural Network; Rhythm; Signal processing; Sinuses; Tachycardia; Wearable technology; Deep learning; Arrhythmia detection; Residual Neural Network; Computer aided diagnosis; M-health
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115031

•Automated arrhythmia detection with high accuracy.•Residual Neural Network for time series data classification.•Six class problem: five arrhythmias and normal.•RR intervals: Convenient data acquisition and cost efficient data handling.•Algorithmic foundation for smart m-health applications. Arrhythmias are abnormal heart rhythms that can be life-threatening. Atrial Fibrillation (AFIB), Atrial Flutter (AFL), Supraventricular Tachycardia (SVT), Sinus Tachycardia (ST), and Sinus Bradycardia (SB) are common arrhythmias that affect a growing number of patients. In this paper we describe a method to detect these arrhythmias in RR interval signals. We propose a deep learning algorithm to discriminate these fife arrhythmias and Normal Sinus Rhythm (NSR). The deep learning model was trained and tested with data from 10093 subjects. We used 10-fold cross-validation to establish the performance results. The overall accuracy for the six-class problem was 98.37%. When considering the binary problem of arrhythmia versus NSR, where the arrhythmia group is formed by combining the data from all fife arrythmias, the performance results are: Accuracy (ACC) = 98.55%, Sensitivity (SEN) = 99.40%, Specificity (SPE) = 94.30%. These results indicate that it is possible to discriminate RR interval sequences from SVT, ST, SB, AFIB, AFL, and NSR subjects with minimal error. Furthermore, the proposed model can provide a robust and independent second opinion when it comes to a decision if arrhythmia is present or not. Another positive aspect of the proposed arrhythmia detection algorithm is economic viability. RR interval signals are cost-effective to measure, communicate, and process. The discriminate powers of the proposed algorithm together with the advent of wearable technology and m-health infrastructure might lead to pervasive long-term arrhythmia monitoring. The detection results can support early diagnosis which helps to reduce the burden of the disease.