Atrial fibrillation detection using heart rate variability and atrial activity: A hybrid approach

• Published in: Expert systems with applications Vol. 169; p. 114452
• Main Authors: Hirsch, Gerald, Jensen, Søren H., Poulsen, Erik S., Puthusserypady, Sadasivan
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 01.05.2021; Elsevier BV
• Subjects: Algorithms; Atrial activity; Atrial fibrillation; Automatic detection; Cardiac arrhythmia; Classifiers; Discriminant analysis; Electrocardiography; Empirical mode decomposition; Ensemble classifier; Feature extraction; Fibrillation; Heart rate; Heart rate variability; Performance evaluation; Subspace methods; Time series; Empirical mode decomposition; Atrial activity; Heart rate variability; Atrial fibrillation; Ensemble classifier; Automatic detection
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2020.114452

•A hybrid model for low resolution, real-time detection of atrial fibrillation.•Subject wise cross validation reveals over optimism of conventional evaluation.•A hybrid classification approach leading to improvements on all metrics.•A total of 24 statistical features and three classifiers were used and evaluated.•Empirical mode decomposition combines filtering and feature pre-processing. Goal: Develop a real-time hybrid scheme for the automatic detection of atrial fibrillation (AF), based on the RR interval (RRI) time series and the atrial activity (AA) derived from the electrocardiogram (ECG) signals. Method: The whole scheme was developed and tested on the MIT-BIH AF database (AFDB). First the R-peak detection and the filtering was performed. Following, all features regarding the RRI time series and AA were extracted. These features were then fed into three popular classifiers (boosted trees (BoT), random forest (RF), and linear discriminant analysis (LDA) with random subspace method (RSM)). Sampling training and test data from the same subject (23 overall) was strictly avoided. Furthermore, for each ECG, individual performance statistics were analyzed to elaborate on the subject-wise performance dependencies. Results: From a 4-fold cross validation (CV) analysis, the RF classifier provided the best results with a sensitivity (Sn), specificity (Sp), accuracy (Acc), and F1 score of 98.0%, 97.4%, 97.6%, and 97.1%, respectively for the AF prediction. Test results on individual ECG’s however, have slightly reduced these performances to 95.9%, 96.1%, 97.4% and 88.4%, respectively. Conclusion: Using the RRI features alone were found to provide satisfying prediction performance of the model. The addition of AA features to the model enhanced the model performance by up to 3%. Overall, the results obtained in this study are comparable or even superior to the state-of-the-art algorithms using RRI and AA based features. Significance The hybrid model allows us to detect AF even with regular RRI. The performance was evaluated under real-world conditions, and no manual labelling, exclusion, or pre-processing was performed. Furthermore, we evaluated the performance for each ECG individually and kept the subjects strictly unknown for the classifier. Finally, we show that the overall performance on a data set, especially from a standard CV, results in an over-optimistic estimation.