Considerations for Spectral Tracking of Respiration in Photoplethysmography

(1)

Mikko Pirhonen

CONSIDERATIONS FOR SPECTRAL

TRACKING OF RESPIRATION IN PHOTOPLETHYSMOGRAPHY

Approach by Bayesian logic

Faculty of Medicine and Health Technology

Master of Science Thesis

January 2020

(2)

(This page is intentionally left blank)

(3)

ABSTRACT

Mikko Pirhonen:

"Considerations for spectral tracking of respiration in photoplethysmography" – 208 pages Master of Science Thesis

Tampere University

Degree Program in Bioengineering, MSc (Tech) January 2020

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 surro- gate 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. Fur- thermore, 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.

Keywords: Photoplethysmography, respiration, algorithm construction, time-frequency representation, synchrosqueezing, particle filtering, spectral fusion

The originality of this thesis has been checked using the Turnitin OriginalityCheck service.

(4)

TIIVISTELMÄ

Mikko Pirhonen:

"Näkökohtia hengitysspektrin seurantaan fotopletysmografiassa" – 208 sivua Diplomi-insinöörin opinnäytetyö

Tampereen yliopisto

Biotekniikan DI-tutkinto-ohjelma Tammikuu 2020

Hengityksen fysiologisen sääntelyn muutokset ovat osoittautuneet pääosin primääreiksi ennus- merkeiksi tarkkailtaessa terveydentilaa äkillisesti heikentävien poikkeamien syntyä. Tämän johdosta hengitystaajuutta (RR, engl. respiratory rate), yhtä ihmisen vitaaleista parametreista, on esitetty mahdollisena muuttujana fysiologisten häiriötekijöiden luokittelussa ja havainnoinnissa jatkuvan monitoroinnin aikana. Tästä huolimatta, nykyiset hengitysmuuttujien sovellukset ovat ra- jallisessa käytössä sairaalalaitteistoihin liittyvien käyttökulujen ja obtrusiivisten mittausmenetel- mien johdosta. Optisesti perifeerisen veritilavuuden vaihtelusta tietoa antavaa menetelmää, fo- topletysmografiaa (PPG, engl. photoplethysmography), on esitetty yhdeksi vaihtoehdoksi RR:n monitorointiin. Tämä menetelmä tukeutuu huomioon siitä, että hengityksen piirteet kytkeytyvät erinäisinä signaalikomponentteina verenkiertoelimistön toimintaan ja siten PPG-signaalista ero- tettaviin piirteisiin. Menetelmässä voidaan hyödyntää laajalti kliinisessä ympäristössä saatavilla olevia pulssioksimetri-laitteita. Kuitenkin, PPG-pohjaisen RR-estimoinnin yleistymisen hidasteena on erityisesti signaalinkäsittelyalgoritmeihin liittyviä haasteita.

Tämän opinnäytetyön jako on kaksitahoinen; hengityksen fysiologian ja fotopletysmografian syventävä osio korostaa biolääketieteellistä teoriatuntemusta; toisaalta matemaattinen osio RR:n erotuksesta mahdollistaa syventävän teorian esityksen bayesiläisestä seurannasta ja signaalin- käsittelystä. Avainasioina toimivat kehitysaskeleet uudesta, monivaiheisesta algoritmirakenteesta sekä PPG-signaalin ominaisuuksista. Työssä seurataan erityisesti hengitykseen kytkeytyneitä modulaatiokomponentteja, jotka tunnetaan kootusti joukkona hengityksen indusoimia muunteluja (RIV, engl. respiratory induced variability). Yksi työn tuloksista on avoimesti julkaistu tietoaineisto 'MARSH', joka sisältää mittausdataa 29 nuoresta aikuisesta, mahdollistaen kehityksen PPG-poh- jaisissa hengitystunnistusalgoritmeissa sekä kehitettyjen menetelmien suorituskyvyn arvioinnin verrattuna impedanssipneumografi- ja termistorihengitysmaskireferensseihin. Myöhemmissä kappaleissa painotetaan tilastollista päättelyä ja tukeudutaan bayesiläisen logiikan esittelyyn.

PPG-RR -algoritmi koostuu erillisistä esikäsittely- ja signaalintasoitusvaiheista. Työssä esite- tään uusia tapoja algoritmin eri vaiheiden toteutuksesta hyödyntäen Butterworth- ja Savitzky-Go- lay -suotimia, aika-taajuusesitystä nimeltä aallokemuunnoksen synkropuristus (engl. wavelet synchrosqueezing), sekventiaalista Monte Carloa ja aika-taajuuskomponenttien etäisyyden kor- relaatioon spektrillä perustuvaa fuusiota. Algoritmin tuottamia tuloksia on vertailtu impedans- sipneumografiseen referenssiin, ja tuloksia on tarkasteltu yleisillä tilastollisilla menetelmillä ja ver- rattu suoraan spektripohjaiseen RR-tunnistukseen. Lisäksi työssä hyödynnetään bayesiläisiä ti- lastollisia menetelmiä, jotka arvioivat työssä esitetyn fuusiomenetelmän, partikkelisuotimien pa- rametrisoinnin, ja esikäsittelyn suotimien mahdollistamia parannuksia.

Työn tulokset viittaavat tilastollisesti uskottaviin (bayesiläisen 95% tiheysintervalliin (HDI) pe- rustuviin) parannuksiin hyödynnettäessä fuusiomenetelmää erillisiin RIV-spektreihin. Tärkeimmät tulokset viittaavat fuusiomenetelmällä parannetun, hengityksen indusoiman taajuusmuuntelun (RIFV) estimoivan parhaiten hengitystaajuutta, tuottaen keskimääräisen virheen itseisarvon (MAE) 1,763 hengitystä minuutissa (BPM) (SD: 0,655 BPM), keskineliövirheen (RMSE) 3,985 BPM (SD: 1,086 BPM), sekä kattavuusosuuden CP2(=% MAE:sta alle 2 BPM) -arvon 81,12%

(SD: 10,26%), fuusiomenetelmän tällöin parantaen MAE:tä 0,185 BPM (95% HDI:0,029-0,349 BPM, vaikuttavuus [engl. effect size]: 0,548) ja RMSE:tä 0,250 BPM (95% HDI:0,073-0,431 BPM, vaikuttavuus: 0,653). Muut RIV-mekanismit tuottavat vielä merkittävämpiä parannuksia. Yhteen- vetona, RIV-joukon jäsenten fuusio parantaa PPG-RR algoritmin suorituskykyä tilastollisesti mer- kittävästi RR:n monitoroinnissa.

Avainasiat: Fotopletysmografia, hengitys, algoritmikehitys, aikataajuus-esitys, synkropuristus, partikkelisuodin, fuusiospektri

Tämän opinnäytetyön alkuperäisyys on tarkistettu käyttäen Turnitin OriginalityCheck -palvelua.

(5)

FOREWORD

Proicit ampullas et sesquipedalia verba – Horace, Ars Poetica

The circumstances facilitating the writing of this thesis have been multifaceted; it is my sincere pleasure to recognize the great opportunity provided by the Faculty of Medicine and Health Technology, Tampere University (formerly, Tampere University of Technology), and the supervisor, assistant professor (tenure track) Antti Vehkaoja, for their academic excellence and support on directing the research projects related to this work. In addition, university teacher, and the second supervisor of this thesis, Mikko Peltokangas provided invaluable insight for the theory and operation of instrumentation relating to photoplethysmography.

My gratitude goes to the great many individuals who participated in the recordings of the dataset, which proves a core source in promoting scientific transparency and providing the signpost for scientific inference here, in publications, and to come.

Numerous individuals have supported me directly or indirectly through this matura- tion process. I would like to thank my parents and sister who have taught me judge- ment with diligence. Friends and colleagues have given me great comprehension and nurture from outside perspectives and visions.

The scientific community, from which I have had the opportunity to learn from, has shown interest to the topics presented here, and have been supportive towards the novel ideas and improvements over the methodologies discussed here.

The findings preceding the writing of this thesis, including the scientific contributions, have been facilitated along the years by the funding from Business Finland and several companies in projects VitalSens and Buddy and the Smiths 2.0.

Finally, I dedicate this thesis to the years spent at the Tampere University of Tech- nology, and the tradition of the former university.

Tampere, 7.12.2019

(6)

TABLE OF FIGURES

Figure 1. The differing aspects governing the early assessment of a

monitoring device. ... 6 Figure 2. A prospective market response to the need for pulse oximeters

over the years 2013 to 2025 as predicted by (Grand View

Research, 2019). ... 7 Figure 3. The classification of Wearable Health Devices (WHD) indicates the

monitoring tasks characteristic to specific device application. From (Dias et al. 2018). ... 9 Figure 4. A commercial Nellcor® monitor. This monitor is one of the few

current devices for the measurement of respiration unobtrusively.

heart rate by OHR sensors. From (SEIKO EPSON Corp. 2019). ... 13 Figure 6. A) The pulsatile arterial blood and deep, non-pulsatile tissue put

together the PPG Signal waveform. B) The photoplethysmography is an illumination method applied either in reflectance or

transmission mode, former typical for WHDs, while latter is

commonly characterized to pulse oximeters. From (Bilgaiyan et al.

2018). ... 22 Figure 7. The absorption bands of red and near-IR light during distinction of

oxygenated hemovariants. From (Chan et al. 2013) ... 23 Figure 8. Microcirculation and large blood vessels in dermis. The

anastamosis (the bifurcation of microvasculature) thicken at the illumination site following change in transmural pressure due to a

cardiac pulse. From (Kamshilin 2015). ... 23 Figure 9. The formation of PPG waveform in the optical path of illuminated

light. The PPG signal, given below, is typically inverted due to convention with photodetector operation. From (Tamura et al.

2014). ... 26 Figure 10. Differences in the PPG signal morphology with age and monitoring

site. The PPG waveform exhibits change in morphology with

arterial aging. From (Allen and Murray 2003). ... 27 Figure 11. The PPG waveform exhibits cyclic changes akin to that

encountered in HRV analysis from ECG. Prospects for HRV measurements have been studied from local waveform maxima.

From (Vandenberk et al. 2017). ... 29 Figure 12. Common respiratory-induced variabilities (RIV), the RIFV, RIAV,

and RIIV. Furthermore, two other variability signals have been indicated, and will be studied concurrently with three

acknowledged modulations. From (Pirhonen and Vehkaoja 2019). ... 31 Figure 13. Prospective measurement sites on mapped by camera images in

two subjects, a) and b). c) and d) illustrates the sites of maximal

amplitude in PPG waveform. From (Kamshilin et al. 2015). ... 33 Figure 14. Changes in photoplethysmographic waveform morphology

conditional to measurement location. Note the phase difference.

From (Hartmann et al. 2019). ... 34 Figure 15. The spectral power content of respiratory component in PPG

signal according to disparate measurement sites (Nilsson et al.

2007) ... 35 Figure 16. A) The conventional setups on studying illumination in

photoplethysmography. B) The illuminated light is transmitted or

(10)

reflected on a photodetector or a camera for further analysis. From (Wang 2017). ... 36 Figure 17. The physics governing the light-tissue interactions. From

(Anderson and Parrish 1982). ... 38 Figure 18. The fundamental blocks of electronics in the operation of the pulse

oximeter circuitry. From (Allen 2007). ... 39 Figure 19. The fiducial points as distinguished by our analysis. Specifically,

we will track the most distinguishable morphologies: peaks, dicrotic notches and peaks, and troughs. ... 41 Figure 20. Top: the conventional PPG waveform, exhibiting a weak dicrotic

notch. Bottom: The second derivate of the PPG signal (APG), which classifies important fiducial points on PPG signal. (From

Elgendi 2012). ... 42 Figure 21. The common modulative properties of the PPG signal, which are

likely generated to some degree by surrogated, physiological

events such as respiration. From (Charlton et al. 2016). ... 44 Figure 22. PPG signal features of pulse area and their ratio. Top: The PPG

offers some measures of waveform that may modulate from the structural properties of amplitude/width relation or quadrature (=area integral). Bottom: one of proposed measures differentiates areas to be cut-off at dicrotic notch. From (Valsalan 2017)... 47 Figure 23. The effect of choosing the sample rate on PPG in time-domain.

SDNN refers to PRV measure for the standard deviation of normal peaks interval. Choosing arbitrarily high sampling rates may implicate difficulties in data acquisition, so a compromise is

necessary. From (Choi and Shin 2017). ... 48 Figure 24. Common artefacts coupling to movement: a) gross movement, b)

tremor, c) bout of coughing, and d) gasp or yawn. From (Allen

2007). ... 50 Figure 25. The absorption spectra of common medical dyes that may emerge

as prospective artefacts in PPG following administration. From

(Chan et al. 2013). ... 51 Figure 26. In the most fortunate situations, such as one displayed here, the

AM-FM components are clearly distinct and specify the component of the ROI. Red dots implicate the largest spectral component at

each time instant of the representation. ... 55 Figure 27. Top: The PPG waveform has a distinct morphology, observable

during rest. Bottom: During movement of measurement site or device, the PPG is readily distorted, and as such more difficult to evaluate accurately for physiological information. From (Pimentel

et al. 2017). ... 56 Figure 28. The physiological structure of human cutaneous tissue. From

(Kadam et al. 2014). ... 58 Figure 29. The illuminated light frequencies penetrate varying depths of skin

structure, with green (reflective) mode absorbing to mere dermis

level. From (Moço et al. 2016). ... 59 Figure 30. The structural partitioning of the human respiratory system.

Retrieved from (Commons 2007). ... 62 Figure 31. The physiological structure of respiratory muscles in the thorax.

From (Peracchia et al. 2014). ... 63 Figure 32. Activation of respiratory muscles. The movement of the diaphragm

upwards during expiration relaxes external muscles and activates the internal muscles. Vice versa is observed during inspiration.

From (Peracchia et al. 2014). ... 64

(11)

Figure 33. Neural pathways in respiration control. The neural control of respiration originates through medulla oblongata and pons, from which the signaling transmits to intercostal muscles. From (Betts et al. 2014). ... 65 Figure 34. The physiological partitioning of pneumotaxic and ventral

respiratory groups, and the mapping of related neurons. From

(Zaidi et al. 2018). ... 67 Figure 35. The connections of chemoreceptor afferents to heart as

conductors of respiratory information. From (Pearson Ed. 2011)... 69 Figure 36. The RSA (frequency-component of RR-variability) culminates

distinctively on this polygraph following the holding of breath. From (Hirsch and Bishop 1981). ... 70 Figure 37. Prospective measurement sites for photoplethysmography. From

(Castaneda et al. 2018). ... 72 Figure 38. The ventilatory relations to types of hypopnea and apnea for the

air-flow and movement of thorax. From (Brenner et al. 2008). ... 74 Figure 39. Examples of time-frequency representations of a renal test signal

(SNR of 8 dB with added white noise). a) Short-time Fourier Transform (STFT), b) Continuous wavelet transform (CWT), c) Wigner-Ville, d) complex demodulation, e) TFR-based

autoregressive (AR) method. From (Scully et al. 2013). ... 79 Figure 40. The physiological signal is studied in fixed window lengths and

subsequent FFT provides the means for a crude TFR resolution.

From (Chiu et al. 2015). ... 81 Figure 41. Formulation of wavelet transform. CWT translates a signal in the

time domain through convolutions with wavelets formulated from a mother window. Adapted from (Kirilina et al. 2013). ... 83 Figure 42. Different wavelet function families applied in conventional signal

analysis applications. From (Faust et al. 2015). ... 85 Figure 43. The TFR response of S transform. a) Original signal, b) S

transform, c) and d) STFT with differing window lengths. The performance of S transform is clearly enhanced at low frequencies.

From (Stockwell et al. 1996). ... 88 Figure 44. Wigner-Ville TFR response to an example signal. Interferences

emerge from cross-correlation components. From (Majkowski et al.

2014). ... 89 Figure 45. The partitioning of intrinsic mode function in EMD for a PPG signal.

In practice, one of the extracted lower-frequency IMFs could

correspond with respiration. From (Wang et al. 2010b). ... 91 Figure 46. The synchrosqueezing operator reassigns ('squeezes') the

spectral ridges into their energy centroids. The operator works optimally when spectral ridges are not localized near in frequency.

Here, an example from the wavelet synchrosqueezing of a single,

quadratic chirp signal. From (MathWorks 2019). ... 93 Figure 47. In this figure, different TFRs, including three synchrosqueezing

methods, were compared for the extraction of the respiratory PPG signal. Slight differences in spectral energy distributions are visible for each proposed method. From (Pirhonen et al. 2018). ... 95 Figure 48. Spectral fusion analysis on the signal coherence of separate PPG

variability modulations. This use of Wigner-Ville spectrum appears to improve the detection of respiratory rate (Orini et al. 2011). ... 99 Figure 49. The conventional steps governing sequential Monte Carlo, or

particle filtering. a) The state vector comprises particles localized conditional to prior step. b) The next state vector is acquired, c) a likelihood of particle localization evaluated, d) the redundant

(12)

particles removed, and e) new particle distribution approximated.

From (Berg et al. 2019). ... 103 Figure 50. The development workflow of a particle filter. From (MathWorks

2019). ... 105 Figure 51. Resampling constitutes a process to remove degeneracy in low-

weight particles and redundancy in high-weight particles. From

(Wang and Gao 2015). ... 117 Figure 52. Particle filtering scheme applied in this work. The reallocation

resampling provides means to diminish particle degeneration with

a simple, yet intuitive approach. From (Pirhonen et al. 2018). ... 120 Figure 53. Distributions over subject parameters that were acquired prior to

the measurements. ... 125 Figure 54. The monitoring protocol for the study that was conducted for the

generation of a dataset ‘MARSH’. The protocol was envisioned to include a variety of respiratory actions. From (Pirhonen and

Vehkaoja 2019). ... 127 Figure 55. Amplitude responses to typical filter designs in preprocessing

algorithms. From (Podder et al. 2014). ... 130 Figure 56. The effect of filter choice on the morphology of the PPG waveform.

It has been argued that for short segments, Chebyshev filter may

prove beneficial. From (Elgendi 2019). ... 132 Figure 57. Savitzky-Golay smoothing slightly relieves the morphological

challenges brought by slight mounds in PPG waveform. Such morphological aberrations would complicate the distinction of

fiducial points in the pulse contour. ... 133 Figure 58. The permutations of five RIV family signals are compared for

distance-correlation of the highest magnitude component. The occurrences of sufficiently close ridges are recorded, and finally summed onto a baseline RIV spectra of choice. From (Pirhonen

and Vehkaoja 2019). ... 142 Figure 59. In this figure, the occurrence matrix of sufficiently closely located

spectral ridges are displayed. Each color represents a distinctive permutation pair. This information is conveyed to a baseline

spectra in the proposed fusion method. ... 143 Figure 60. Fusion methodology (top) improves the continuity of a signal ridge,

enhancing the performance of particle filtering, and thus the accuracy of respiratory mode tracking. Bottom: particle filter

response without fusion. From (Pirhonen and Vehkaoja 2019). ... 146 Figure 61. The main differences governing the statistical inference in

Bayesian and Frequentist logic. From (Van de Schoot et al. 2014). ... 154 Figure 62. Visual representation of a Bayesian estimation by 'BEST' library.

This example, on the difference of pooled group means of RILAV with fusion intervention indicates highly credible improvement not evaluable by frequentist means. ‘v’ here indicates the degrees of

freedom in the t-distribution sense. ... 156 Figure 63. An example on estimators and their credibility intervals (HDI). We

observe 99.814% credibility that the fusion intervention improved

the sample mean value among other information. ... 157 Figure 64. In this violin plot, one may differentiate the change in distribution of

MAE for five different modulations, with or without the fusion method (F, NF, respectively). The bars indicate 90%, 75%, 50%

percentiles with red, yellow, and green colors, and 1.96σ limits of agreement (LOA) of mean (diamond marker) as the magenta interval. These results refer to computation of 100 particles, one

ridge, and use of Savitzky-Golay smoothing filter. ... 165

(13)

Figure 65. The violin plots on RMSE for five different modulations, with or without the fusion method. These results refer to the computation of 100 particles, one ridge, and use of Savitzky-Golay smoothing

filter. ... 166 Figure 66. The percentile and CP2 indicators for MAE as a cumulative

distribution. Each permutation of fusion and use of smoothing has been presented, along with instructions on reading CP2 value for

RIFV (fusion, Savitzky-Golay). ... 179 Figure 67. The violin plots for the distribution of CP2 for different RIVs with or

without fusion. These results refer to the computation of 100

particles, one ridge, and use of Savitzky-Golay smoothing filter. ... 179 Figure 68. The improvement of MAE at different frequency ranges brought

along the introduction of the fusion method, with 95% HDI

indicated by shaded regions. ... 181 Figure 69. The effect of particle size to observed MAE. We observe negligible

changes in MAE after the increase in particle set size. These

results do not justify the inevitable increase in computation time. ... 182 Figure 70. The increase in computational time following the increase in the

particle set size. 95% line of agreements are given as a blue-

shaded area. ... 183 Figure 71. The change brought by fusion on each RIV following fusion, with

bars denoting 100, 500, 1000, and 2000 particles, respectively.

The plus and minus signs refer to the same designation as given on the result tables VII to XVII. ‘o’ indicates not available for

application. ... 184 Figure 72. The change in computational time following the change in number

of spectral ridges, with 95% limits of agreement in shaded areas. ... 185 Figure 73. Change in statistical measures by Savitzky-Golay. We observe

slight indication towards a benefit, but it is not statistically conclusive, except for RIAV. MAE and SD presented are the differences in respected errors with or without the smoothing filter.

Negative value indicates improvement after Savitzky-Golay. This tabularised figure represents method with 100 particles, one ridge, and fusion. ... 186 Figure 74. MAE of each separate measurement is shown here with the

median, 25/75% percentiles, and whiskers of extreme data. We observe one of the subjects to skew the results for MAE and other measures. ... 189 Figure 75. Arrow-plot on common statistical estimators. From each

modulation source: RIV-colored arrowhead points to change (mean of the interval) imposed by intervention. Colorbars indicate 95% credibility interval. Effect sizes for each measure are given below in the arrow stem (unitless). From (Pirhonen and Vehkaoja 2019). MAE-FB indicates MAE during the phases of unlabored,

free breathing (i.e., eupnea). ... 191 Figure 76. Mean-difference plot on the results acquired in RIFV with

additional information visually incorporated. Towards the inner circle: mean-difference plot, histogram of points at the frequency

interval, SD, and MAE. From (Pirhonen and Vehkaoja 2019). ... 192 Figure 77. A particularly well-behaving EDR signal, extracted by a penalized

greedy algorithm, closely matches the with the trend of the IP

reference. ... 193 Figure 78. Conclusions over the statistical implications and each RIV as we

infer from the statistical discussion prior. ... 195

(14)

ABBREVIATIONS

AM Amplitude Modulation

ANS Autonomic Nervous System APG Accelerated Plethysmography BA(S)N Body Area (Sensor) Network BPM Breaths per Minute

CAGR Compound Annual Growth Rate CLT Central Limit Theorem

CNS Central Nervous System

COPD Chronic Obstructive Pulmonary Disorder CP2 Coverage Probability (=% below 2 BPM MAE) CWT Continuous Wavelet Transform

ECG Electrocardiography

EWS Early Warning Score

F/NF Fusion/No fusion

FM Frequency Modulation

FIR Finite Impulse Response

FB Free Breathing (= eupnea, unlabored respiration) FSST Fourier Synchrosqueezing Transform

FUS Fusion intervention GCP Good Clinical Practice HDI High Density Interval

HR Heart Rate

HRV Heart Rate Variability ICU Intensive Care Unit

IID Independent and Identically Distributed (random variable), i.i.d.

IIR Infinite Impulse Response IMF Intrinsic Mode Function IMT Intrinsic Mode Type -function

IP Impedance Pneumography

iPPG Imaging (Noncontact) Photoplethysmography LCD Liquid-Crystal Display

LED Light Emitting Diode LOA Limits of Agreement

MAE Mean Absolute Error

MAP Maximum a posteri

MARSH Monitoring Associable to Respiration with Simulated and Hab. rates MAX1/5 Maximum 1/5 peak(s) at specified time

MC Monte Carlo

MCMC Markov Chain Monte Carlo

MDR Medical Device Regulations OHR Optical Heart Rate

OSST Oblique Synchrosqueezing Transform PDF Probability Density Function

PEP Pre-Ejection Period

PF Particle filtering

PNS Peripheral Nervous System PPG Photoplethysmography PRV Pulse Rate Variability PTT Pulse Transit Time

RAR Rapidly-Adapting Receptor RIV Respiratory Induced Variability

RIAV Respiratory Induced Amplitude Variability

(15)

RIFV Respiratory Induced Frequency Variability RIIV Respiratory Induced Intensity Variability

RIPV/RIDAV Respiratory Induced Peaked/Dicrotic Amplitude Variability RILAV Respiratory Induced Lower-Amplitude Variability

RM Reassignment Method

RMSE Root-mean square error RR Respiration/Respiratory rate

SA Sinoatrial

SaO2 Arterial blood oxygenation

SD Standard Deviation

SIR Sequential Importance Resampling SIS Sequential Importance Sampler SNA Sympathetic Nervous Activity SNR Signal-to-Noise Ratio

SpO2 Partial Oxygen saturation (in blood) SST Synchrosqueezing Transform SQI Signal Quality Indices

TSR Time-Scale Representation

VSST Vertical Synchrosqueezing Transform WNN Weighted Nearest Neighbor

WHD Wearable Health Device WPO Wearable Pulse Oximeter

WSST Wavelet Synchrosqueezing Transform WVD Wigner-Ville Distribution

Mathematical notations

A Amplitude

a Scale

arg Complex argument function b Translational value

ℂ Field of complex numbers Cψ Admissibility constant δ Dirac delta function

E Expected value

η Temporal frequency

ε Smoothness (intrinsic-mode type) f Ordinary frequency

fr Rayleigh frequency fs Sampling rate

Ƒ DFT operator

ḟ DFT transform

gt Dynamic mapping of state and noise L^p Lebesgue function space

L^∞ Maximum norm

Lv Largest available scale max Maximal element med Median element

μx Mean (of x)

N Sample size

ℕ Set of natural numbers Neff Effective sample size O(∙) Computational complexity

(16)

𝜔 Angular frequency p Probability distribution ϕ (Instantaneous) phase

𝛱 Smoothing kernel (Cohen class) Pr Empirical probability

Ṗr Empirical probability estimate (MC)

𝜓 Wavelet function

q Importance (or ‘instrumental’) function R Number of random numbers drawn

ℝ Real field

rect Rectangular function σ Standard deviation

T Window length

𝜏 Lag

U Uniform distribution

u Random number draw

v Voices per Octave

vt Process noise sequence w Particle weight

Wf Wigner function xt State sequence x[n] Discrete signal

ẊMAP Maximum a posteri estimate ẊMEDIAN Median error estimate

ẊMMSE Minimum mean-square error estimate w[n] Window function

zt Observation mapping

(17)

1. INTRODUCTION

Contemporary advances in engineering facilitate the development of clinical solutions.

Recent advances in health informatics and computational sciences have set forth substantial prospects for the implementation of elegant, yet structurally intricate engineering solutions to clinical applications. Amongst these visions emerge trends in the manufacturing of small-scale, low-power instrumentation designs, and matura- tion of approaches in the detection and classification of physiological curiosities in homeostatic feedback control. In the engineering perspective, an emphasis is grad- ually directed towards closing the information gap between the acquisition of vital parameters in computational, built-in instrumentation designs, and informing their respective implications over to a medical practitioner or laymen. Appropriately, the recent advances ensue through facilitation of novel methodologies that reflect the cross-interaction of various fields, such as follows from the allocation of recent knowledge on the scaling of electronics and signal processing methodologies. For the clinical environment, the prospective applications include improvements in therapeutic applications or designs of continuous monitoring for vital physiological parameters, the latter indicating tasks such as the classification of respiratory behavior or estimation of hemodynamic parameters. To complement on such tasks, the surg- ing comprehension on the particulars of human physiology paves the conceptualiza- tion and diagnostic tools for the preventative diagnosis of physiological aberrations at disparate time scales. As an increasing trend, the instrumentation designs offer prospects that specify the needs of an end-user: varying the downscaling and operating principles offer the possibilities of automatization or wireless support, which may allow for the reallocation of time and human resources in clinical settings. Facil- itating the ease of access to corresponding devices may also prove as an insurance to the patient and clinicians, so that effective and safe treatment practices are advo- cated.

An important portion of the administered resources in clinical units tie either directly or indirectly to continuous monitoring, which facilitates the workflow on numerous medical interventions. The development of appropriate sensor networks, including the interest set forth by Internet of Things, points out the allure in uniting multiple sources of instrumentation data to facilitate computer-aided classification, even at

(18)

distance to the measured individual. At the same time, amplified competition over clinical accuracy, as well as introduction and recent transition to modernized medical device regulations (MDR) thereof instigate further competition on the improvement of monitoring instrumentation in the performance, safety, and ease of application.

Incidentally, many novel monitoring designs have been incorporated, if not specifically tailored, outside clinical settings. Respective products raise the health awareness in the consumer market and, when such information is jointly conveyed to clinicians, may complement with the early warning score (EWS), triage, or follow-up period.

Correspondingly, the design of instrumentation for physiological monitoring entails perspectives over various fields of engineering and product development expertise.

Traditionally, such development further bifurcates to company management, and research and development, R&D. Particularly, unobtrusive device designs have been extensively pushed forward due to their clear implications for the ease and/or comfort of application. Such unobtrusive, physiological measurands of contemporary interest range from well-reviewed instances of heart rate estimation to such emerging tech- nologies as respiratory, glucose, or stressor monitoring.

1.1 Respiratory monitoring

Respiratory parameters provide an intriguing case example on monitoring workflow.

One could argue that the development of novel engineering products, such as algorithms or operation designs for instrumentation, indicates a collaborative enterprise of diverging considerations; in this thesis we will provide a case example of such a workflow during the planning and deployment of a respiration measurement protocol, including the perspectives over data collection, algorithm construction, and parameter estimation of a vital physiological sub-parameter, respiratory rate (RR). Respira- tion constitutes an indispensable, homeostatic response to needs of tissue oxygenation; correspondingly, instances of nervous signaling blockage, a primary forewarn- ing encountered in clinical care, readily translates to cessation of the muscle response in the lungs and further to imbalance in oxygen exchange for tissues. More- over, the respiratory acidosis or alkalosis that follows from the loss of respiratory regulation may be life-threatening. Particularly, respiratory rate is a sub-parameter in a larger context of ventilatory measurands and covers general information over patient status in long-term monitoring settings as well as on intermittent periods. The physiological aspects, in which we base the foundation for monitoring the respiration, connect to circulatory dynamics in peripheral vasculature,

(19)

Table I. The progression of the discussion, with differentiation of core concepts to be accentuated.

particularly arteries, through a technique known as transmissive photoplethysmography, which we will discuss in further detail along this thesis. The dedicated segments of our discussion relate to how mathematics of respiratory coupling in signal is col- lectable by a pulse oximeter device. Our emphasis and goal include the understand- ing about concepts in physiology and mathematics corresponding to the conventional, developmental timeline of novel engineering features and getting familiarized with the conventional data analysis tools for novel algorithm verification. We omit here, for brevity, the discussion over specifics of electronics design and medical device regulations, save the discourse on good clinical practice (GCP) as a part of de- signing a scientific, measurement protocol. While our treatment accentuates the theory of respiratory physiology and respective challenges imposed by signal processing, similar discussion and insight may be translated to a variety of applications in monitoring of vital, clinical parameters, or otherwise.

The core aims and impacts on the following chapters are as follows: the reader should understand the meaning and extent of physiological monitoring as a collaborative enterprise of disparate engineering fields. We revise respiratory physiology to provide the groundwork for details governing measurement protocols for data acquisition. Following this examination, the reader is introduced to the mathematical foundation of performing spectral signal analysis and component tracking by Bayesian inference. In addition, we emphasize a novel fusion method, distance-correlation of instantaneous-component spectral ridges, united to the original tracking principle of synchrosqueezing and sequential Monte Carlo (MC) method (i.e., particle filtering).

In data analysis, we initiate a proposal for the use of Bayesian logics as part of statistical inference for this field. The core objective of our results discussion is to convey the efficiency of advanced signal analysis tools to the reader through literature-es-

Core Aims

Coupling of respiratory events in circulation. Intro- duction to photoplethysmogra- phy. Extraction of

respiratory rate as a clinical pa-

rameter.

Core Concepts Insight to respiratory physiology.

Novel application of mathematical concepts to physiological monitor-

ing. Suggested solution A novel algorithm construct on localization of respiratory signal

components.

Synchrosqueezing and particle filtering with a fusion

methodology. Performance Evaluation

Bayesian statistical inference on

standard measures of dis-

persion. Evalu- ation of plausible

defects.

(20)

tablished statistical measures and visual presentation tools. In conclusion, we exam- ine the future directions of respiratory monitoring in general perspective. We indicate the general differentiation over topics to be accentuated in this thesis on Table I.

Respiratory rate is an indispensable indicator for recognizing primary signs of decline in patient health.

The vital physiological parameters comprise the blood pressure, heart rate (HR), respiratory rate (RR), oxygen saturation (SpO2), and core temperature. Additionally, (El- liott et al. 2012) argues over the parameters for pain, consciousness, and urine output as prospective factors towards the onset of homeostatic deterioration. In this thesis, our main consideration emphasizes accurate estimation of respiratory rate, which, as an indispensable variable towards the control of myriad, vital physiological processes, may provide a range of primary forewarnings towards decline in health (Pirhonen and Vehkaoja 2019, Fleming et al. 2011). Albeit an essential baseline measure, the respiration rate monitoring is not commonly facilitated by medical equipment save the exception of interventions during anaesthesia, critical care, or related, uncompromising environments. These instances justify the use of obtrusive instrumentation to support the prospects of patient recovery. Unfortunately, contemporary, clinical devices are bulky and relatively expensive to operate. In this thesis, we base the measurement variables and much of the discussion about signal analysis on an emerging, unobtrusive technique of photoplethysmography (PPG) to measure respiration through peripheral blood volume variations. The corresponding monitoring device is called the pulse oximeter, a peg-shaped probe device and its respective computational unit (which may be integrated), which proves accessible and applicable to most clinical environments due to its ease and familiarity of use. Partic- ularly, we generalize the signal analysis to the consideration of the physiological, digitalized signal, photoplethysmogram, which is predominantly regarded as a signal source for hemodynamics and the oxygen saturation of hemoprotein variants in blood tissue. We partition the discussion in sections of relevant physiology of monitoring this signal waveform and its respective properties through novel insights in the emergence of respiratory rate components. Importantly, we consider, in a comprehensive manner, the algorithmic construction based on novel mathematical tools, and finally propose the practicality of the algorithm on respiratory rate estimation. Moreover, we study the results conditional to the properties and nature of the dataset (protocol) that is designed for this study. Additionally, data from other clinical monitoring parameters complement the available, statistical inference as well as future studies: a short de-

(21)

scription over these physiological signals of electrocardiography, impedance pneumography, and oral-nasal signals are covered in subsequent discussion, where we discuss their respiration-induced features. These signals act here as known refer- ences in the verification process of novel respiratory rate algorithm formulations and providing additional insight for the proposed measurement protocol.

1.2 Objectives of the thesis

The main objectives presented in this thesis revolve around the discussion of considerations for both the theory and realization of a PPG-RR algorithm and the statistical inference over the characterization of the algorithm performance. We make a hypothesis that a novel spectral method and tracking based on Bayesian methodology is viable in characterization of the RR component. Furthermore, we propose a novel, spectral fusion method and evaluate its effect on the proposed algorithm. In addition, we aim to provide a new dataset of transmissive PPG signals to the development of novel RR -algorithms. This dataset is constructed on the basis that the measurement protocol includes various respiratory patterns, and may prove chal- lenging, transient changes between high and low respiratory rates. Accordingly, one of the objectives of this thesis is also to study the PPG-RR algorithm conditional to the qualities of this dataset.

1.3 The economy of clinical monitoring

Monitoring of vital, physiological parameters proves as a major financial partaking.

In the field of physiological monitoring, algorithms act on the sensory output of a medical device, decomposing analogous, physiological signals to meaningful, digital interpretations for the end-users. A comprehensive review on vital physiological monitoring by (Khan et al. 2016) discusses the design for monitoring devices for each vital parameter. In the forefront, selected physiological variables operate in the framework of continuous recording, i.e., in a monitoring setup that supposedly re- flects the ever-fluctuating system of homeostasis in real-time. Ultimately, the end purpose is to facilitate diagnosis following the potential onset of any medical abnor- mality by examining multiple, different signal properties simultaneously. Accordingly, research on instrumentation design and algorithms form the fundamental fields of study in developing physiological monitoring solutions. Particularly, the general interest in industry is illustrated through the prospects of monitoring by integrated systems and patient monitors, and the investment of medical industry on these devices.

(22)

As of this date, techniques in physiological monitoring, and thus in the development of medical advances, demonstrate a major financial undertaking; contemporary esti- mates for medical device market, between the years 2018 to 2023, propose a compound annual growth rate (CAGR) of 4.5%, this translating to estimated 409.5 billion U.S. dollars market volume worldwide by the year 2023 (SMT 2018). Accordingly, the competitive industry seeks further increase in instrumentation performance, with dependence on the engineering research and development of algorithm and sensor networks. Moreover, the socio-economic implications to as far as the degrees of so- cial security and the public health care in detection of chronic diseases should not be understated: the performance of an algorithm may decrease response time in emer- gency and/or accelerate patient rotation (Malasinghe et al. 2019). Additionally, in the consumer market, the respective methodologies may be integrated into devices to raise awareness in consumer health, and as such provide passive monitoring solutions. It is likely that increased prevalence of chronic diseases, which increasingly load the healthcare system, necessitates self-monitoring applications to the consumer market, further inflating and diversifying the instrumentation range. In the consumer electronics, the wearable device market is projected to exhibit major growth:

the worldwide sales revenue of 33.78 billion U.S. dollars in 2019 is expected to cross over 73 billion U.S. dollars in 2022 (Tractica 2019). Computational resources to operate such remote, home devices may have benefited from recent development in small-scale consumer electronics and/or mobile device applications.

To date, the home health monitoring devices have included a variety of tests such as biochemical sensing, activity, and sleep monitoring. These devices do not necessarily fulfill the requirements of monitoring accuracy in clinical environments, save

Figure 1. The differing aspects governing the early assessment of a monitoring de- vice.

(23)

the regulation of blood glucose monitors, yet their design and operation may neces- sitate further evaluation by local authorities over their safe use, as the risks of infec- tion or injury may persist. The monitoring for markers of non-communicable diseases, such as blood glucose levels, a prospective parameter for pulse oximetry (Yama- koshi et al. 2017), may alleviate the workload of healthcare systems globally. The vast range of considerations towards the marketing of medical devices has been illustrated in Figure 1. Such considerations include rigorous market research, yet further illustrate the clinical context of physiological variable acquisition, latter the key- note which we will mostly accentuate on in this thesis.

In the following discussion, we emphasize the arising sub-field of such physiological monitoring, photoplethysmography-derived signal analysis, with prospects on eco- nomical and conventional integration to contemporary medical environment. Of no- table importance in this thesis, we emphasize the importance of respiratory rate monitoring, which encompasses fundamental implications not only towards respiratory tract health, but also to the plummeting of neuronal control during the maintain of homeostasis. As it occurs, respiration presents an intriguing parameter to measure;

the respiratory rate is among the primary indicators determining the decline of patient condition, while currently restricted in use owing to low access and obtrusiveness of monitoring. Accordingly, the issue is multi-faceted: while we emphasize the mathematical analysis of respiration, considerations over instrumentation design and physiology prove essential. Thus, our discussion will concisely include the introductory

Figure 2. A prospective market response to the need for pulse oximeters over the years 2013 to 2025 as predicted by (Grand View Research, 2019).

(24)

framework of related concepts, including respiratory physiology, research design, and tools in signal processing.

Monitoring of wellbeing instigates engineering designs outside the clinical environ- ment.

The market share evaluation of photoplethysmography in the global perspective has been discussed in a Market Research Report by (Grand View Research, 2019). Ac- cording to the report, the ever-increasing readmissions of patients, specifically the cost that incurs from these to clinical environments, has accelerated the interest in monitoring networks, and particularly in the use of both clinical and consumer -grade pulse oximeter devices. An estimate of 1.8 billion U.S. dollars market share over the year of 2018 and the CAGR of 6.3%, one of the largest growing rates in the sector, propose that the physiological monitoring has become increasingly prevalent measure to combat diseases such as Chronic Obstructive Pulmonary Disorder (COPD), pulmonary fibrosis, or cardiac arrhythmias. The prospective market trajectory towards the demand for pulse oximeters has been illustrated in Figure 2. The use of pulse oximeters improves the quality of life when integrated with other devices, for example during patient triage (Shah et al. 2015). However, the reader should note that pulse oximeters, or monitoring devices for respiratory rate, are not therapeutic devices for accessing human respiration: a different, substantially larger market ex- ists for such devices, particularly through positive airway pressure machines and ventilators.

The wireless pulse oximeter devices further provision the health care in telemedicine as a part of the body area sensor network (BASN) discussed by (Poon et al. 2006).

An important, supporting notion in this thesis regarding this topic is the following:

currently, the firmware of many medical devices, particularly multiparameter units in monitoring settings, may be updated with new capabilities (effectively, algorithm soft- ware, but also data management), since these additions do not necessarily require changes in hardware structure. Accordingly, the prospects of integrating novel, enhanced features into pulse oximeter devices, which are evidently stocked in clinical and consumer environments, may become definitive selling points for the affiliated industry. As pulse oximeter units are relatively cheap, the introduction of novel features, such as respiration rate extraction discussed in this thesis, may promote eco- nomically sustainable policies on clinics. In the consumer market, these prospects are readily integrable to activity and home care devices, such as in mobile activity

(25)

watches or wireless devices, with recording provided by mobile applications. The advantages and critical issues have been studied by (Baig et al. 2015). Incidentally, few devices, such as Oura® ring, offer the option for respiratory trend measurements, primarily applicable during sedentary, resting states. When the measurements are conducted concurrently, wirelessly with many prospective devices, the monitoring is said to involve body-area networks (BAN), surveyed by a few authors, such as (Patel and Wang 2010a, Chen et al. 2011).

A core portion of medical devices are operated during adapted clinical interventions.

On the other side of the spectrum, Wearable Health Devices (WHD) include a growing sector of customer-based, ambulatory monitoring in use of everyday life, providing information over vital signs, such as recovery statistics, to the individual or a cli- nician, as discussed by Dias et al. (Dias et al. 2018). Technically, the content of this thesis is readily applicable to the devices under the umbrella term ‘WHD’, bounded by online or offline applications and remote or clinical environments, as most customer and clinical devices comply with the same signal analysis tools proposed here.

Figure 3 demonstrates the partitioning of the WHDs, assisting the classification of application. Some of these applications are regulated due their use in clinical applications, while the portion of consumer-oriented ‘wellness’ devices are not. Im- portantly, whether it is the heart rate, respiration rate, or any of the other fundamen- tally monitoring parameters, there is a distinction in engineering designs for alarming

Figure 3. The classification of Wearable Health Devices (WHD) indicates the monitoring tasks characteristic to specific device application. From (Dias et

al. 2018).

(26)

over health complications and informing over activity. Much of this thesis concen- trates on the prediction segment; we will mostly consider the results in the perspective of algorithm performance of respiration measurement accuracy, and not construe a warning system for any medical complication or formulate a system for diagnostics.

1.4 Development of monitoring solutions

Monitoring devices support on the partitioning of measurement design, physiology, mathematics, and the application and intervention with the information.

The foremost separation in physiological monitoring occurs at the techniques incorporated on the instrumentation side, particularly in the mathematics that govern rapid feature deduction of non-stochastic, cyclic features from stochastic, deteriorating components of ambient (extracorporeal) environment. Moreover, approximation and sampling variance may be introduced following the digitalization of an analogous input signal, although this may prove mostly negligible in comparison with analogous signal noise level. Furthermore, algorithms that address physiological parameters comprise only a minor segment of the monitoring system, operating under the theo- retical premise of a known signal waveform or inherent property thereof, originating from intricate, chemical or mechanical signaling of physiological conductors. Yet, the algorithm design may be conditional to the measurement protocol or even actively guide it, specifically whenever the signal is perturbated or enhanced through some intervention. Particularly, active sensors may be calibrated according to the algorithm response though most sensors remain passive in this sense during use. With respect to timing, the emergence of physiological processes precedes the response of a sensing element constructed according to the theory surrounding the relevant physiology, such as synchronous firing of neurons in electroencephalography of the brain or circulatory blood perfusion on the periphery in photoplethysmography. Conditional to the measurement site, the sensor and its connection to data transmission are pro- tected by both measurement environment and coupling of ambient error sources.

Thus, engineering applications are obliged to shield and otherwise facilitate the at- tenuation of ambient noise and provide means to sample the signal from analogous content to digital representation. Typically, a signal is the presentation of measured variable with respect to time and/or image space, which induces different dimensions for the input data. As an example, the respiratory rate is a one-dimensional variable that evolves as a rate information with respect to time, while physiological imaging, such as x-ray imaging, is based on spatial, multi-dimensional data management. In this thesis, we emphasize techniques that involve a single-dimensional variable that

(27)

evolves in non-stochastic fashion with respect to time. The operational flow of algorithms, the primary interest of this thesis, presents a complex integration of one or several mathematical relations that map an unprocessed, digital signal to a compre- hensible output, such as the heart rate or respiratory rate, often through the extraction of a specific morphological or frequency-domain feature. Finally, the end-user or a classification system formulates a decision over procedures according to algorithm output. The concluding phase in the physiological monitoring setup includes consideration in the representation, classification, and possible storage of data. The classification may involve intricate advanced, applied topics in mathematical logics, such as fuzzy inference networks as examined by (Seera and Lim 2014).

Assessing respiratory behavior proves important yet complex task in clinical monitor- ing.

Respiration monitoring forms an integral part of clinical care in neonatal and intensive care units (ICU), analgaesia, and in the supervision of respiratory complications. The patient triage and early warning scores may attain further accuracy with the evaluation of the baseline respiratory response. Additionally, contemporary awareness towards medical maladies in the airways, such as observed during apnea, asthma, or lower respiratory tract infections has turned practitioners to diagnosing levels of respiratory effort and ventilation. Respiratory information complements well in any clinical monitoring system as complications such as chronic obstructive pulmonary dis- ease (COPD), hypertension, or cystic fibrosis may bring about pathophysiological forewarnings developing to abnormalities in breathing behavior or patterns. Particu- larly, COPD and apnea, overlapping in pathophysiology, require monitoring during sleep stages (McNicholas 2009). Moreover, it is widely known in medical community that deterioration of baseline respiratory rate is among the first indicators of onset in decline of overall patient health. For example, in neuromuscular diseases the respiratory control is lost due to muscular inactivity, leading to hypoxia (the lack of tissue oxygenation), elevated carbon dioxide levels, and, when left untreated, death. An- other example instance would call for the deterioration of neuronal pathways in me- dulla oblongata and pons, the location for respiratory rhythm control by the pre- Bötzinger complex. Currently, the respiratory monitoring instrumentations are expensive, bulky instruments that provide little comfort for the patient. Such methods may incorporate oral-nasal thermistor or pressure masks (Norman et al. 1997) or elec- trode setups (Ernst et al. 1999) to distinguish respiratory events. The practitioner may also hesitate on the use of instrumentation that may interfere with other medical procedures. Some clinical environments may have inadequate instrumentation supply

Considerations for Spectral Tracking of Respiration in Photoplethysmography

Mikko Pirhonen