Adaptive learning and cross training improves R-wave detection in ECG. (March 2021)

• Record Type: Journal Article
• Authors: Ganapathy, Nagarajan; Swaminathan, Ramakrishnan; Deserno, Thomas M.
• Is Part Of: Computer methods and programs in biomedicine. Volume 200(2021)
• Publication Date: 2021-03
• Subjects: Electrocardiography -- R-wave detection -- Deep learning -- Convolutional neural network -- Cross-database training -- Cross-lead training -- Adaptive training -- Adaptive model; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.105931
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: Continuous ECG monitoring requires robust R-wave detection. Convolutional neural networks (CNN) outperform other techniques. Adaptive cross data and cross lead training further increases performance. Evaluation is based on 4 ECG datasets with 4.5 million annotated beats. Abstract: Background and Objective: Automated R-wave detection plays a vital role in electrocardiography (ECG) and ECG-based computer-aided diagnosis. Recently, a multi-level one-dimensional (1D) deep learning approach was presented that shows good performance as compared to traditional methods. Methods: In this paper, we present several improvements of the multi-level 1D convolutional neural network (CNN)-based deep learning approach using: (i) adaptive deep learning, (ii) cross-database training, and (iii) cross-lead training. For this, we consider ECG signals from four publicly available databases: MIT-BIH, INCART, TELE, and SDDB, having 109, 404, 175, 660, 6, 708, and 1, 684, 447 annotated beats, respectively. Except for TELE, all databases provide at least two-lead recordings. To evaluate the improvements, experiments are performed with adaptive k-times cross-trained databases validation scheme (k = 5). The hypothesis tested are: (i) the improvements outperform the state-of-the-art, (ii) cross-database training and adaptive deep learning contribute, and (iii) additional databases or cross-lead training further improves the results. Results: Our proposed approach outperforms the state-of-the-art. In terms of F-measure, F = 99.75% and F = 95.25% is obtained for the MIT-BIH and TELE databases, respectively. Further, cross-database training (F = 98.02%) is found to be more effective than training on individual databases (F = 97.33%). The performance of our approach further improves when additional databases and different leads are used for training. Conclusion: Existing state-of-the-art methods perform low on noisy and pathological signals. Adaptive cross-data training identifies the optimal model. Using multiple datasets and leads allows analyzing noisy, pathological and mobile-recorded long-term ECG signals without ground truths. These conclusions are based on the comprehensive evaluation of four different databases, and in total, about 4.5 million annotated beats.