A Multistage Deep Residual Network for Biomedical Cyber-Physical Systems

In this paper, we propose a novel biomedical cyberphysical system for automated and efficient arrhythmia and seizure detection in the time-series biomedical signals such as electrocardiogram (ECG) and electroencephalography (EEG). We use a novel multilayer, automated, and multistage deep residual network for the anomaly detection in the biomedical signals. Generally, the biomedical datasets have class imbalance problem; hence, we leverage the concepts of undersampling techniques to address this issue. The proposed algorithm is validated on the publicly available benchmark MIT-BIH Arrhythmia and CHB-MIT Scalp databases. The results show a significant improvement in terms of the sensitivity of 90% and 97.1% for supraventricular and ventricular beats for best fold, respectively. The accuracy obtained is at par with most of the state-of-the-art methods, and in particular, for the supraventricular beats, the proposed method outperforms all but one state-of-the-art method. The advantage of the proposed method is that it gives reliable results with EEG samples of small duration and, as opposed to other state-of-the-art methods, it does not involve any preprocessing, hence computationally efficient. Additionally, the proposed algorithm provides 81% sensitivity for seizure detection in EEG signals, which is comparable to existing deep learning methods.