A Robust Fusion Model for Estimating Respiratory Rate From Photoplethysmography and Electrocardiography

• Published in: IEEE transactions on biomedical engineering 2018-09, Vol.65 (9), p.2033-2041
• Main Authors: Birrenkott, Drew A., Pimentel, Marco A.F., Watkinson, Peter J., Clifton, David A.
• Format: Article
• Language: English
• Subjects: Algorithms; Amplitude modulation; Biomedical monitoring; Correlation; Data processing; EKG; Electrocardiography; electrocardiography (ECG); Electrocardiography - methods; Estimation; Fast Fourier transformations; Fourier transforms; Frequency dependence; Frequency modulation; Humans; photoplethysmography (PPG); Photoplethysmography - methods; Quality assessment; Respiration; respiratory quality index (RQI); Respiratory rate; Respiratory rate (RR); Respiratory Rate - physiology; Robustness; signal fusion; Signal Processing, Computer-Assisted; Signal quality; Waveforms; Weight; Windows (intervals)
• ISSN: 0018-9294; 1558-2531
• DOI: 10.1109/TBME.2017.2778265

Objective: Respiratory rate (RR) estimation algorithms based on the photoplethymogram (PPG) and electrocardiogram (ECG) lack clinical robustness. This is because the PPG and ECG respiratory modulations are dependent on patient physiology, regardless of general signal quality. The present work describes an RR estimation algorithm using respiratory quality indices (RQIs) that assess the presence or absence of the PPGand ECG-derived respiratory modulations. Methods: Six respiratory waveforms are derived from the amplitude modulation, frequency modulation, and baseline wander of the PPG and ECG. The respiratory quality of each modulation is assessed by using RQIs based on the fast Fourier transform, autoregression, and autocorrelation. The individual RQIs are fused to obtain a single RQI per modulation per time window. Based on a tunable threshold, the RQIs are used to discard poor modulations and weight the remaining modulations to provide a single RR estimation per time window. Results: The proposed method was tested on two independent datasets and found that using a conservative threshold, the mean absolute error was 0.71 ± 0.89 and 3.12 ± 4.39 brpm while discarding only 1.3% and 23.2% of all time windows, for each dataset, respectively. Conclusion: These errors are either better than or comparable to current methods, and the number of windows discarded is far lower demonstrating improved robustness. Significance: This work describes a novel preprocessing algorithm that can be implemented in conjunction with other RR estimation techniques to improve robustness by specifically considering the quality of the respiratory information.