Fig 1.
Three types of signal fusion strategies are categorized based on the different stages of signal fusion: Early Fusion, Intermediate Fusion, and Late Fusion.
Fig 2.
Outline of the proposed method.
Outline of the proposed approaches: 1. respiratory modulated components (RMCs) extraction: three modes of RMCs, i.e., amplitude modulation - AM, frequency modulation - FM, and baseline wander - BW, were extracted from ECG and PPG, which gives the ECG / PPG AM (EAM, PAM), ECG / PPG FM (EFM, PFM) and ECG / PPG BW (EBW, PBW); 2. RMCs Preprocessing: denoising of the RMCs with variational mode decomposition (VMD) and screening of the RMCs with respiratory quality index (RQI), where T represents the average RQI of all 6 RMCs; 3. principal component analysis (PCA) fusion of the screened RMCs; 4. Respiratory rate(RR) estimation from the fused respiratory waveform.
Fig 3.
Signal denoising by variational mode decomposition (VMD).
Left panel: input original signal and the decomposed intrinsic mode functions (IMFs); Right panel: the corrresponding power spectrum of the signals.
Fig 4.
Respiratory modulated components extraction.
Representative electrocardiogram (ECG) and photoplethysmogram (PPG) signals alongside the corresponding respiratory modulated components extracted. ECG Amplitude Modulation (EAM), ECG Frequency Modulation (EFM), ECG Baseline Wander (EBW), PPG Amplitude Modulation (PAM), PPG Frequency Modulation (PFM), and PPG Baseline Wander (PBW).
Fig 5.
Respiratory modulated components filtering and fusion.
Comparison of respiratory modulated components (RMCs) before and after variational mode decomposition (VMD), as well as the comparison of estimated respiratory signal and reference respiratory signal: (a) Time domain waveform of the RMC before VMD; (b) Frequency domain power spectrum of the RMC signal before VMD; (c) Time domain waveform of the RMC after VMD; (d) Frequency domain power spectrum of the RMC after VMD; (e) Estimated respiratory waveform obtained by fusing multiple RMCs and the reference respiratory waveform.
Fig 6.
Comparison of estimated respiratory rates and reference values.
Scatter plot (a) and Bland-Altman plot (b) of respiratory rate (RR) estimation with the proposed method (RQI-PCA) versus reference RR in Capnobase dataset.
Table 1.
Respiratory rate estimation performance in Capnobase.
Fig 7.
Time-series of respiratory rate obtained using different methods.
Time-series of respiratory rate (RR) estimates for two representative subjects in Capnobase dataset with the methods including Chung et al. (blue line with diamond dots), Motin et al. (purple line with asterisk dots), Langley et al. (green line with square dots) and the proposed method (RQI-PCA) (yellow line with triangle dots) versus the reference values (red line with hexagonal dots): (a) subject 0147 in controlled breathing state and (b) subject 0332 in spontaneous breathing state.
Table 2.
Comparison between the proposed method (RQI-PCA) and methods in the previous literature with different signals.
Fig 8.
Variation of error with window length.
The mean absolute error (MAE) of two datasets - BIDMC (purple and blue lines) and Capnobase (yellow and orange lines) - at different window sizes.