Improved heart disease detection from ECG signal using deep learning based ensemble model

• Published in: Sustainable computing informatics and systems Vol. 35; p. 100732
• Main Authors: Rath, Adyasha, Mishra, Debahuti, Panda, Ganapati, Satapathy, Suresh Chandra, Xia, Kaijian
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.09.2022
• Subjects: CAD; DL based CAD classification; Early detection; Ensemble CAD detection; Imbalanced data; Imbalanced data; DL based CAD classification; Early detection; CAD; Ensemble CAD detection
• ISSN: 2210-5379
• DOI: 10.1016/j.suscom.2022.100732

The heart disease (HD) is very fatal in nature and comparatively takes more number of lives across the world. To save lives from the HD, early and robust detection method is essential. The HD of a subject can be diagnosed by clinical test attributes, the electrocardiogram (ECG) signal, heart sound signal, impedance cardiography (ICG), magnetic resonance imaging and computerized tomography (CT). In this paper, the problem of detection of coronary artery disease (CAD) using ECG signal as the prime source has been undertaken by developing four different deep learning (DL) models such as autoencoder (AE), radial basis function network (RBFN), self-organizing map (SOM) and restricted Boltzmann machine (RBM). Two public arrhythmia datasets: PTB-ECG and MIT-BIH have been used for training and validation of the proposed models. Further, an ensemble classification model has been developed by combining two best performing AE and SOM models using the principle of majority voting. Simulation based experiments using two standard datasets demonstrate that the AE model providing accuracy, F1-score and area under the curve (AUC) values of 0.974, 0.932 and 0.922 for MIT-BIH dataset and 0.984, 0.967 and 0.932 for PTB-ECG datasets respectively outperforms the other three models. It is further observed that the proposed SOM - AE ensemble model exhibits the best performance compared to the AE model with accuracy, F1-score and AUC values of 0.984, 0.971, 0.997 for MIT-BIH as well as 0.992, 0.986 and 0.995 for PTB-ECG datasets respectively. The detection models presented in this work can further be employed for other different diseases. The robustness and other performance measures can be obtained by applying larger and more imbalanced datasets to the DL models. In addition, internet of things (IoT) based diagnosis platform can be developed using proposed approach for online detection of CAD which would be beneficial for detecting remote CAD patients. •Detection of coronary artery disease (CAD) using ECG signals as the prime source has been undertaken by developing four different deep learning (DL) models such as autoencoder (AE).•Two standard arrhythmia datasets: PTB-ECG and MIT-BIH have been used for training and validation.•An ensemble classification model has been developed by combining two best performing AE and SOM models using the principle of majority voting.