Sparse spectrum based swarm decomposition for robust nonstationary signal analysis with application to sleep apnea detection from EEG. (August 2022)

• Record Type: Journal Article
• Authors: Bhalerao, Shailesh Vitthalrao; Pachori, Ram Bilas
• Is Part Of: Biomedical signal processing and control. Volume 77(2022)
• Publication Date: 2022-08
• Subjects: Sparse spectrum -- Swarm decomposition -- Least absolute shrinkage and selection operator -- Hilbert-Huang transform -- Nonstationary signal -- Time–frequency representation -- Sleep apnea disorder; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103792
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: A new decomposition technique, the swarm-sparse decomposition method, is proposed. The feature fusion and TFR image features are proposed for sleep apnea detection. The proposed method has also been tested on synthetic and real-life EEG signals. The developed framework outperforms existing methods for detecting sleep apnea. Abstract: Background and motivation: Time–frequency representation (TFR) of a signal finds its application in numerous fields for non-stationary multicomponent signal analysis. Due to underlying difficulties and improvement scope in the current methodology, developing a new time–frequency method can improve spectral analysis of real-life signals and further can be extended to practical applications. Materials and methods: The proposed new method swarm-sparse decomposition method (SSDM) is an advanced version of swarm decomposition (SWD) for decomposing nonstationary multicomponent signals into a finite number of oscillatory components (OCs). Benefiting from sparse spectrum and SWD, the proposed SSDM method delivers optimal estimation of boundary frequencies in the sparse spectrum, resulting in improved filter banks. In addition to SSDM, we have also proposed the spectrum approximator function, i.e., fused least absolute shrinkage and selection operator to modify sparse spectrum and get significant OCs. The performance of the proposed SSDM has been evaluated by TFR analysis and compared to SWD and Hilbert-Huang transform methods. Also, it has been tested for automated sleep apnea classification using a convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) on the publicly available EEG database. Results: The proposed SSDM-TFR-CNN and SSDM-feature-fusion-BiLSTM frameworks outperformed all the compared methods used for sleep apnea detection and achieved the highest classification accuracy of 96.24% and 95.86%, respectively, in the subject-independent cross-validation scheme. Conclusion: Simulation result shows that the proposed SSDM method delivers substantial improvement in time–frequency analysis. Our developed sleep apnea detection model could be a vital aid in clinical solutions.