Considerations for Spectral Tracking of Respiration in Photoplethysmography
Pirhonen, Mikko (2020)
2020
Biotekniikan DI-tutkinto-ohjelma - Degree Programme in Bioengineering, MSc (Tech)
Lääketieteen ja terveysteknologian tiedekunta - Faculty of Medicine and Health Technology
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2020-01-10
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202001041038
https://urn.fi/URN:NBN:fi:tuni-202001041038
Tiivistelmä
Respiratory regulation constitutes integral measurands towards the assessment of primary decline in patient health. As a vital parameter, the respiratory rate (RR) facilitates the classification and detection of physiological aberrations in continuous monitoring setting. However, contemporary application of this ventilatory parameter proves limited due to expenses and obtrusive nature with prevailing instrumentation. Accordingly, a well-established optical measurement method known as photoplethysmography (PPG) has been proposed as a prospective alternative for RR monitoring. This technique supports on the notion that respiratory physiology couples as a surrogate signal component into progression of a vascular blood pulse. The method employs an unobtrusive device, the pulse oximeter, for which the application is well-established in the clinical community. As such, the integration of RR estimation algorithms indicates major economic and practical advantages. To date, the main direction in development has accentuated on establishing clinically appropriate accuracy, yet many challenges persist, particularly in algorithm design.
The partitioning of this thesis is two-fold; discussion over respiratory physiology and photoplethysmography emphasizes the biomedical theory over the ventilation mechanism, while mathematical discussion on RR acquisition facilitates advanced topics on Bayesian tracking and signal analysis tools. The main topics comprise the steps in development of a novel multi-layered algorithm and the particulars of PPG signal characteristics thereof. Particularly, we consider the respiratory-coupled modulation sources to PPG signal, the respiratory induced variability family (RIV), including three commonly encountered and two novel formulations of variability signals. Additionally, a dataset, called ‘MARSH’, was constructed from the measurements of 29 young, adult subjects, facilitating data analysis and characterization of algorithm performance. Emphasis on later chapters on statistical inference provides novel insights for presentation in Bayesian logic.
The core structure of a PPG-RR algorithm consists of disparate operational phases. A novel set of features has been proposed, including concepts of preprocessing and smoothing filters by Butterworth and Savitzky-Golay filters, time-frequency representation known as wavelet synchrosqueezing transform, sequential Monte Carlo (sequential importance resampling particle filter), and peak-distance -correlated spectral enhancement method. The algorithm output has been compared with an impedance pneumography reference and the performance has been evaluated with respect to common statistical measures and against blind spectral RR determination. Furthermore, Bayesian logic in statistical inference by computational methods proposes the improvement imposed by the fusion method, particle filter parameters, and the choice of smoothing filters.
Results indicated statistically credible (in Bayesian 95% high density interval (HDI) sense) improvement following the use of the fusion method on different RIV spectra. The main findings indicate fusion enhanced respiratory-induced frequency variability (RIFV) to provide the most accurate readings, with mean absolute error (MAE) of 1.763 breaths per minute (BPM) (SD: 0.655 BPM), root-mean square error (RMSE) of 3.985 BPM (SD: 1.086 BPM), and CP₂ (=% of MAE below 2 BPM) of 81.1% (SD: 10.3%), with the fusion method improving MAE by 0.185 BPM (95% HDI:0.029-0.349 BPM, effect size: 0.548), and RMSE by 0.250 BPM (95% HDI:0.073-0.431 BPM, effect size: 0.653). Other components in RIV family exhibited further pronounced results. We conclude that use of Savitzky-Golay may implicate small improvements, but such results remain statistically inconclusive. We conclude that the fusion of RIV family components proves integral to the improvement of future constructions of PPG-RR algorithms, and that the dataset ‘MARSH’ is a prospective toolbox for further studies on the assessments of PPG-RR algorithm performance.
The partitioning of this thesis is two-fold; discussion over respiratory physiology and photoplethysmography emphasizes the biomedical theory over the ventilation mechanism, while mathematical discussion on RR acquisition facilitates advanced topics on Bayesian tracking and signal analysis tools. The main topics comprise the steps in development of a novel multi-layered algorithm and the particulars of PPG signal characteristics thereof. Particularly, we consider the respiratory-coupled modulation sources to PPG signal, the respiratory induced variability family (RIV), including three commonly encountered and two novel formulations of variability signals. Additionally, a dataset, called ‘MARSH’, was constructed from the measurements of 29 young, adult subjects, facilitating data analysis and characterization of algorithm performance. Emphasis on later chapters on statistical inference provides novel insights for presentation in Bayesian logic.
The core structure of a PPG-RR algorithm consists of disparate operational phases. A novel set of features has been proposed, including concepts of preprocessing and smoothing filters by Butterworth and Savitzky-Golay filters, time-frequency representation known as wavelet synchrosqueezing transform, sequential Monte Carlo (sequential importance resampling particle filter), and peak-distance -correlated spectral enhancement method. The algorithm output has been compared with an impedance pneumography reference and the performance has been evaluated with respect to common statistical measures and against blind spectral RR determination. Furthermore, Bayesian logic in statistical inference by computational methods proposes the improvement imposed by the fusion method, particle filter parameters, and the choice of smoothing filters.
Results indicated statistically credible (in Bayesian 95% high density interval (HDI) sense) improvement following the use of the fusion method on different RIV spectra. The main findings indicate fusion enhanced respiratory-induced frequency variability (RIFV) to provide the most accurate readings, with mean absolute error (MAE) of 1.763 breaths per minute (BPM) (SD: 0.655 BPM), root-mean square error (RMSE) of 3.985 BPM (SD: 1.086 BPM), and CP₂ (=% of MAE below 2 BPM) of 81.1% (SD: 10.3%), with the fusion method improving MAE by 0.185 BPM (95% HDI:0.029-0.349 BPM, effect size: 0.548), and RMSE by 0.250 BPM (95% HDI:0.073-0.431 BPM, effect size: 0.653). Other components in RIV family exhibited further pronounced results. We conclude that use of Savitzky-Golay may implicate small improvements, but such results remain statistically inconclusive. We conclude that the fusion of RIV family components proves integral to the improvement of future constructions of PPG-RR algorithms, and that the dataset ‘MARSH’ is a prospective toolbox for further studies on the assessments of PPG-RR algorithm performance.