Cerinaetal 2023 Physiol. Meas. 10.1088 1361-6579 Acb12b

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/366925214
Extraction of cardiac-related signals from a suprasternal pressure sensor

during sleep
Article in Physiological Measurement · January 2023

DOI: 10.1088/1361-6579/acb12b
CITATIONS READS
0 24
6 authors, including:
Luca Cerina Gabriele Papini

Eindhoven University of Technology Philips
34 PUBLICATIONS 233 CITATIONS 28 PUBLICATIONS 241 CITATIONS
SEE PROFILE SEE PROFILE
Pedro Fonseca Sebastiaan Overeem

Philips Eindhoven University of Technology
87 PUBLICATIONS 1,252 CITATIONS 312 PUBLICATIONS 16,964 CITATIONS
SEE PROFILE SEE PROFILE
Some of the authors of this publication are also working on these related projects:
the role of orexin in goal-directed behaviour View project
reconfigurable architectures for fog computing View project
All content following this page was uploaded by Luca Cerina on 22 February 2023.
The user has requested enhancement of the downloaded file.

Physiological Measurement
ACCEPTED MANUSCRIPT • OPEN ACCESS
Extraction of cardiac-related signals from a suprasternal pressure sensor

during sleep
To cite this article before publication: Luca Cerina et al 2023 Physiol. Meas. in press https://doi.org/10.1088/1361-6579/acb12b
Manuscript version: Accepted Manuscript

Accepted Manuscript is “the version of the article accepted for publication including all changes made as a result of the peer review process,
and which may also include the addition to the article by IOP Publishing of a header, an article ID, a cover sheet and/or an ‘Accepted
Manuscript’ watermark, but excluding any other editing, typesetting or other changes made by IOP Publishing and/or its licensors”
This Accepted Manuscript is © 2023The Author(s). Published on behalf of Institute of Physics and Engineering in Medicine by IOP
Publishing Ltd.
As the Version of Record of this article is going to be / has been published on a gold open access basis under a CC BY 3.0 licence, this Accepted
Manuscript is available for reuse under a CC BY 3.0 licence immediately.
Everyone is permitted to use all or part of the original content in this article, provided that they adhere to all the terms of the licence
https://creativecommons.org/licences/by/3.0
Although reasonable endeavours have been taken to obtain all necessary permissions from third parties to include their copyrighted content
within this article, their full citation and copyright line may not be present in this Accepted Manuscript version. Before using any content from this
article, please refer to the Version of Record on IOPscience once published for full citation and copyright details, as permissions may be required.
All third party content is fully copyright protected and is not published on a gold open access basis under a CC BY licence, unless that is
specifically stated in the figure caption in the Version of Record.
View the article online for updates and enhancements.
This content was downloaded from IP address 131.155.142.187 on 22/02/2023 at 09:30

Page 1 of 35 AUTHOR SUBMITTED MANUSCRIPT - PMEA-104906.R1
1
2
3
4
pt
5
6
7
8
9
Extraction of cardiac-related signals from a
cri
10
11 Suprasternal Pressure Sensor during sleep
12
13
14 Luca Cerina1 , Gabriele B. Papini1,2 , Pedro Fonseca1,2 ,
15
16
Sebastiaan Overeem1,3 , Johannes P.van Dijk1,3 , Rik Vullings1
us
1
17 Electrical Engineering, Technische Universiteit Eindhoven, Eindhoven, Noord
18 Brabant, The Netherlands
19 2
Philips Research, Eindhoven, Noord Brabant, The Netherlands
20 3
Center for Sleep Medicine, Kempenhaeghe Foundation, Heeze, Noord Brabant, The
21 Netherlands
22
23
24
25
26
27
28
an
E-mail: {l.cerina*, r.vullings}@tue.nl,
{pedro.fonseca, gabriele.papini}@philips.com,
{overeems, dijkh}@kempenhaeghe.nl
Abstract. Objective The accurate detection of respiratory effort during polysomnog-

dM
raphy is a critical element in the diagnosis of sleep-disordered breathing conditions such
29 as sleep apnea. Unfortunately, the sensors currently used to estimate respiratory effort
30
are either indirect and ignore upper airway dynamics or are too obtrusive for patients.
31
One promising alternative is the suprasternal notch pressure (SSP) sensor: a small ele-
32
ment placed on the skin in the notch above the sternum within an airtight capsule that
33
34 detects pressure swings in the trachea. Besides providing information on respiratory
35 effort, the sensor is sensitive to small cardiac oscillations caused by pressure perturba-
36 tions in the carotid arteries or the trachea. While current clinical research considers
37 these as redundant noise, they may contain physiologically relevant information. Ap-
38 proach We propose a method to separate the signal generated by cardiac activity from
pte
39 the one caused by breathing activity. Using only information available from the SSP
40 sensor, we estimate the heart rate and track its variations, then use a set of tuned fil-
41 ters to process the original signal in the frequency domain and reconstruct the cardiac
42 signal. We also include an overview of the technical and physiological factors that may
43 affect the quality of heart rate estimation. The output of our method is then used as
44
a reference to remove the cardiac signal from the original SSP pressure signal, to also
45
optimize the assessment of respiratory activity. We provide a qualitative comparison
46
against methods based on filters with fixed frequency cutoffs. Main results In compar-
ce
47
48 ison with ECG-derived heart rate, we achieve an agreement error of 0.06±5.09bpm,
49 with minimal bias drift across the measurement range, and only 6.36% of the estimates
50 larger than 10bpm. Significance Together with qualitative improvements in the char-
51 acterization of respiratory effort, this opens the development of novel portable clinical
52 devices for the detection and assessment of sleep disordered breathing.
Ac
53
54
55
56 Keywords: Sleep Disordered Breathing, Polysomnography, Suprasternal Notch Pressure,
57 Cardiogenic Oscillations, Heart Rate, Respiratory Effort
58
59
60
Submitted to: Physiol. Meas.
AUTHOR SUBMITTED MANUSCRIPT - PMEA-104906.R1 Page 2 of 35
1
2
3 Cardiac signals extraction from a Suprasternal Pressure Sensor during sleep 2
4
pt
5 1. Introduction
6
7 Several disorders are affecting breathing during sleep. Obstructive sleep apnea (OSA) is
8 the most common manifestation, with a complex pathophysiology and diverse outcomes
9
cri
10 that range from increased cardiovascular risk to excessive daytime sleepiness (Lévy
11 et al. 2015). There is a growing interest in portable diagnostic devices, or home sleep
12 apnea tests (HSAT) (Abrahamyan et al. 2017)), to reduce costs of OSA screening,
13
14 but despite recent technical advancements they still provide an incomplete view of OSA
15 mechanisms. To obtain a complete perspective of brain, cardiac, and respiratory activity,
16
us
and body movements altered by sleep disorders, an overnight polysomnography (PSG)
17
18 recording conducted in specialized sleep clinics remains the gold standard.
19 Concerning respiratory mechanics, PSG provides two types of signals: oronasal
20 airflow, measured with a thermistor or pressure sensor, and respiratory effort,
21
22 representing the work of breathing during inspiration and expiration. PSG can rely
23
24
25
26
27
28
an
on two methods to measure respiratory effort: respiratory inductance plethysmography
(RIP) belts and, albeit less common, esophageal pressure (Pes) sensors. RIP belts
measure thoracoabdominal displacement during breathing and are easy to use but
account only for thoracic and abdominal movements. These belts measure the
dM
respiratory effort indirectly through continuous phase relationships. Conversely, Pes
29
30 provides a more direct measure of respiratory effort by assessing intrathoracic pressure
31 swings. Despite being the recommended gold standard (Berry et al. 2012), Pes sensors
32 are invasive, difficult to place accurately, and poorly tolerated by patients during
33
34 sleep (Brochard 2014, Glos et al. 2018).
35 The suprasternal notch pressure (SSP) sensor represents a potential non-invasive
36 alternative to the esophageal sensor, with comparable reliability in both adults (Glos
37
38 et al. 2018) and children (Amaddeo et al. 2016). The sensor consists of a small pressure-
pte
39 sensitive element placed inside an airtight capsule on the skin notch above the trachea.
40 Due to their specific position, researchers observed that both the Pes and SSP sensors
41
42 measure more than one physiological signal (Ayappa et al. 1999, Glos et al. 2018).
43 Specifically, the SSP signal comprises mainly information about respiratory effort and
44 high-frequency content associated with swallowing and snoring. Upon careful inspection,
45
46 a component with a smaller amplitude is also visible, and quite evident during central
ce
47 apneas when the respiratory effort is completely absent. This component has been
48 associated with cardiac activity (Suarez-Sipmann et al. 2012), hence dubbed cardiogenic
49
50 or cardiac oscillations. Some examples of the presence of cardiac oscillations during
51 normal respiration and apneic events are illustrated in Appendix A
52 Typically, the SSP sensor is used to analyse only respiratory activity, and cardiac
Ac
53
54
oscillations are often filtered out (Glos et al. 2018, Mukhopadhyay et al. 2020). However,
55 we hypothesize that the content of these oscillations may provide additional insights into
56 cardio-respiratory physiology and is therefore worthy of further analysis.
57
58
We propose a method to extract the cardiac signal using only the pressure signal of
59 the SSP sensor. Our contribution is an algorithm capable of estimating heart rate (HR)
60
1
2
4
pt
5 as the fundamental frequency of cardiac oscillations in the SSP signal and a method
6 to separate these oscillations from the respiratory signal with a tuned filterbank. We
7
tested the accuracy of our estimates against the HR measured by the ECG signal.
8
9 As a secondary outcome of our research, we perform a qualitative comparison of the
cri
10 respiratory effort signal filtered with our method. Since the extraction of cardiac
11 oscillations from SSP or Pes is not treated extensively in literature, we discuss the
12
13 shortcomings of the proposed method and challenges encountered, including technical
14 or physiological factors that may influence the quality of the signals.
15
16
us
17 2. Methods
18
19 2.1. Dataset
20
21 To test and develop our method we employed full single-night PSG recordings which
22
23
24
25
26
27
28
an
are part of SOMNIA (van Gilst et al. 2019), a clinical database designed to facilitate
research on sleep disorders and unobtrusive monitoring of sleep. We used a subset of
100 recordings where the PSG included a synchronized recording of the SSP sensor. In
this subset each recording is unique participant-wise. Demographics of participants are
shown in Table 1 together with total sleep time (TST) and apnea-hypopnea index (AHI).
dM
29 The values are expressed as average ± standard deviation and range. AHI is expressed
30
as median and inter-quartile range (IQR)(range) due to its skewed distribution.
31
32
33 Table 1: Demographics of the participants
34
35 mean±std or
36 Variable (range)
37 median and IQR
38
Sex [#] 43 (57 male)
pte
39
40 Age [years] 47±16 (18 : 80)
41 BMI 1,*
[kg/m ]2
25.8±3.6 (19 : 36)
42 TST 2
[minutes] 418±69 (241 : 622)
43 3
44 average HR [bpm] 63±10 (43 : 106)
45 AHI 4
[events/hr] 10.9, 18.8 (0 : 109.9)
46
1
Body-Mass Index; 2 Total Sleep Time; 3 Heart Rate;
ce
47
4
48 Apnea-Hypopnea Index; * Available only for 34 participants
49
50 In our dataset, 71 participants out of 100 presented an AHI above 5 events/h
51
52 (median 16.4), which is the current clinical threshold for OSA diagnosis. Concerning the
Ac
53 manifestation of OSA, 94 participants had more obstructive apneas and hypopneas with
54 respect to central and mixed apneas, with a median ratio of 99.3%. Among the remaining
55
56
six, four presented more central and mixed apneas than obstructive events but had very
57 low AHI below 1 events/h. The last two participants had more central or mixed events
58 than obstructive ones, a severe AHI above 40 events/h, but also respiratory instability in
59
60
the form of Cheyne-Stokes respiration patterns. The electrocardiography (ECG) Lead
1
2
4
pt
5 II data available in the PSG is used as ground truth to test the estimated heart rate
6 and for the quantification of the average HR in Table 1.
7
8
9 2.2. SSP signal characterization
cri
10
11 We can represent the raw signal SSSP as the following:
12
13 SSSP = SResp + SCardio + SAudio + ε. (1)
14
15
SResp represents the pressure swings caused by respiratory activity, while SCardio contains
16
us
17 the pressure waves coming from carotid and pulmonary arteries. SAudio represents noise
18 in the audio frequency range, for example, vocalization or speaking during wakefulness,
19
or snoring during sleep. The ε component is an umbrella term for other noise sources
20
21 present in the signal: coughs, deglutition, vibrations caused by body movements during
22 sleep, or other unknown transient artifacts.
23
24
25
26
27
28
an
The SSP signal was recorded with a sampling frequency fs of 1024 Hz with a low-
pass filter DC-285 Hz. The sensor’s front-end automatic gain control and filters were
disabled to avoid unwanted alterations in the signal’s dynamic range. The SSP sensor
can suffer from technical limitations. For example, a faulty sealing changes the pressure
dM
29 in the cavity, affecting the amplitude range of the whole signal (Figure 1).
30
31 0.1
32
33 0.05
SSP raw [a.u.]
34 0
35
36 -0.05
37
38 -0.1
00:30 01:00 01:30
pte
39
40
3
41 20
Power/Frequency [dB/Hz]
42 2.5
Frequency [Hz]
15
43 2
44 1.5 10
45 1
46 0.5
5
ce
47 0
48 00:30 01:00 01:30
Time [hh:mm]
49
50
51 Figure 1: Time and frequency domain representation of a signal with a loss of dynamics
52 due to sealing failure. Respiration rate (0.25 Hz), fundamental frequency (1.25 Hz) and
Ac
53
54 2nd harmonic of heart rate (2.5 Hz) are visible before the breaking point.
55
56
57
When this happens, the vibrations in the trachea are not strong enough and get
58 dispersed before exciting the sensor. In 9 cases, the sensor failed to record any valuable
59 data, and we decided to introduce an exclusion criterion to detect faulty SSP recordings.
60
1
2
4
pt
5 We opted for the percentage of samples with amplitude smaller than 2% of the sensor
6 range of 0.5App (amplitude peak-to-peak) to measure the magnitude of the problem.
7
The median percentage of low amplitude samples in the whole dataset is 50.8% with an
8
9 IQR of 25.76%. We opted for an exclusion criteria set at 90%.
cri
10 The SSP sensor can also suffer from saturation and clipping artifacts, likely an
11 effect of sensor misplacement, body movements, or other unknown factors. Figure 1
12
13 illustrates an example of spikes observed before the sealing failure. This class of artifacts
14 happened sparsely in the dataset (< 0.01% of the samples), thus we did not exclude
15 any SSP recording due to excessive saturation. Instead, we tried to detect local spikes
16
us
17 as artifacts and excluded only those segments of the signal from our analysis.
18
19
2.3. Methodology overview
20
21 Figure 2 provides an overview of the proposed method and the motivation behind the
22
23
24
25
26
27
28
ˆ
an
computational steps performed. The overarching goal is to get an accurate estimation
of the heart rate at each time step n, HR(n), so that we can tune a filter to separate
the respiratory and cardiac signals SResp and SCardio . For all the following processing
steps depicted in Figure 2 we opted for a time window of 10 s with a 2.5 s step.
We must note that our algorithm includes multiple parameters that concur to its
dM
29 robustness, and that we selected according to technical factors and known physiological
30
31 ranges during sleep, or that we optimized to improve the performance in our specific
32 population. Nevertheless, none of the parameters was tuned upon known characteristics
33 of the participant, making them entirely participant-agnostic. For reproducibility
34
35 purposes, all the parameters are listed in Appendix C.
36
37
38
pte
39
40
41
42
43
44
45
46
ce
47
48
49
50 Figure 2: High level representation of the proposed methodology. Each color represent a
51 sub-section of Section 2, except for the detection of signal artifacts, treated extensively
52
Ac
53 in Appendix B. The symbol n represents the n-th time step in the computation.
54
55
56 Initially, we preprocess the raw SSP signal to remove the SAudio component and
57 powerline interferences that are not related to respiratory and cardiac information using
58 a 4th order Butterworth low-pass filter at 32 Hz. Then the signal is re-sampled from
59
60
1
2
4
pt
5 fs = 1024 Hz to fs = 256 Hz using a FIR polyphase filter. This preprocessing optimally
6 removes the SAudio noise component, but ε will still be present in different forms.
7
After preprocessing, we would ideally identify a precise frequency separating
8
9 respiration and heart rate and design a high-pass filter to remove the respiratory signal.
cri
10 However, the process is not trivial because the heart rate may overlap with the higher
11 harmonics of the respiratory rate (RespR), or because broad noise reduces the spectral
12
13 contrast necessary to distinguish harmonics. In these situations, a filter with the wrong
14 cutoff frequency will either have SResp leaks or lead to the removal of the main component
15 of SCardio . Therefore we opted for a different solution that tries to boost the signal in
16
us
17 the heart rate frequency range and attenuate the respiration as much as possible.
18 We now have a signal where Scardio is present along with minimal residuals from
19 SResp , that we will now consider part of noise ε. We expect that the fundamental
20
21 frequency F 0 of this signal will coincide with the heart rate, such that F 0 ≈
22 HR. However, residual noise may have a magnitude comparable with the cardiac
23
24
25
26
27
28
an
oscillations, and maximum likelihood methods that use discrete fourier transform
(DFT) representations will not work reliably. We chose a time-domain method
that identifies F 0 from its autocorrelation function (ACF), which has already been
demonstrated to be more robust to noise in challenging situations, such as speech
dM
29 analysis (Strömbergsson 2016).
30 The downside of the ACF representation compared to DFT is that there is not
31 a single spectral peak describing a signal with a certain frequency F 0, but multiple
32
33 peaks at lags (indicative of the signal period) that correspond to fs /F 0, its multiples
34 fs /(k ∗ F 0), and dividers fs /(F 0/k) with k = N > 1, even if the signal is a single
35 sinusoid. This ambiguous representation in the ACF implies that we need a plausible
36
37 guess of the range of lags around the expected heart rate and then update the search
38 space over time as the heart rate fluctuates. Starting with a broad range covering
pte
39 observed physiological frequencies ([0.6 : 1.8] Hz) we refine the boundaries according to
40
41 heuristics and assumptions about heart rate dynamics.
42 We further divide each time window to get multiple sub-estimates of HR and
43 to minimize the impact of short transient artifacts, assuming that sub-estimates
44
45 will distribute around its true value. Then we select sub-estimates considering the
46 distribution’s skewness as a potential indicator of outliers. We also try to detect evident
ce
47 artifacts in the SSSP signal to exclude them a priori. The artifact detection method is
48
49 not our main contribution, so we will only describe it briefly. For a detailed description
50 please refer to Appendix B.
51 We then design a filter tuned on the estimated heart rate that separates SCardio
52
Ac
53 and SResp from the preprocessed SSSP signal. For each estimate, we create a band-
54 pass filter from the HR up to its 3rd harmonic and apply it as a convolution in the
55 frequency domain to extract SCardio . We derive SResp directly in the frequency domain
56
57
by subtracting the now known cardiac components from the SSSP spectrum.
58
59
60
1
2
4
pt
5 2.4. Respiration signal attenuation
6
7 The first step of our method attenuates most of the SResp frequencies so that SCardio
8 becomes the predominant signal in which we can identify the heart rate.
9 Let us consider the signal in Figure 3 as a common example for our separation
cri
10
11 process. Often, SResp and SCardio are not perfectly separable, so we need to design a
12 filter that considers the respiration rate RespR, but also other spectral characteristics
13 of the SSSP signal. The identification of RespR is relatively straightforward using DFT
14
15 coefficients since SResp is predominant in the power spectrum and with a high signal-
16 to-noise ratio (SNR). Nevertheless, we must account for the edge case represented by
us
17 breathing disturbances, in which the estimated RespR is absent or much lower than the
18
19 real respiration rate, and how that influences our filter design.
20
21 0.04 1
22 SSP
an
SSP Pressure [a.u.]
0.02 0.8 ECG

23 Heart rate
Density [n.u.]
24 0 0.6
25 -0.02 0.4
26
27 -0.04 0.2
28
dM
-0.06 0
29 0 1 2 3 4 5 6 7 8 9 10 0 0.5 1 1.5 2 2.5 3
Time [s] Frequency [Hz]
30
31 Figure 3: (left) Example of one 10s window of the SSP signal. (right) Normalized power
32
33 spectrum and heart rate marked with dot-dash line at 0.65 Hz, highlighting its overlap
34 with respiration harmonics.
35
36
37 In our example, we extracted the heart rate from the ECG signal, observing a
38 frequency of 0.65 Hz. This frequency is not clearly detectable in the SSSP power
pte
39
40 spectrum, with the closest peaks around 0.53 Hz and 0.875 Hz. The first peak is likely
41 the average between the heart rate and the respiration rate (0.21 Hz) 2nd harmonic. In
42 this and many other cases, the SSSP signal will not guarantee reliable spectral features
43
44
for a high-pass filter at the optimal edge of respiration and cardiac oscillations.
45 We then opted for a different approach. Instead of a finely tuned high-pass filter,
46 we use a low-pass filter and subtract its output from the original signal. In this way,
ce
47
48
lower frequencies (where low-pass gain GainLP = 1) disappear entirely while frequencies
49 around the low-pass cutoff are boosted (Gain(1−LP ) > 1). While a boosting gain
50 would be undesirable in filtering operations, it helps us detect frequencies that resonate
51
52
together. This method does not guarantee the absence of noise under 1Hz or respiration
Ac
53 harmonics, but it gives us confidence that Scardio frequencies will stand out for the heart
54 rate detection step and it has the advantage of working well also with noisy spectra.
55
56
To determine the low-pass filter’s cutoff frequency, we search for the first valley point
57 fmin in the range [0.40 : 0.95] Hz. If none exists, we use 0.66 Hz as a fallback candidate
58 (or the second harmonic of a maximum RespR around 20 breaths-per-minute). This
59
60
frequency is assigned to a new range [fmin : 2.4] Hz, where we detect all local maxima of
1
2
4
pt
5 the power spectrum. A pairwise comparison of these maxima flags those with potential
6 higher harmonics. If a harmonic exists, we use its frequency as the low-pass cutoff, as it
7
may be related to the heart rate, and they will resonate together. Other noise sources
8
9 spread across a wide band of frequencies and should not have definite harmonics. If
cri
10 we cannot find harmonics, we set a fallback value of 1.6 Hz, boosting frequencies from
11 0.62 Hz to 2.8 Hz (considering a 4th order Butterworth filter).
12
13
14 0.04 3.5
high-pass filter
15 0.02 booster
high-pass [a.u.]
16
us
3
0
17
18 -0.02
19 original 2.5
-0.04 high-pass filter
20 booster
21 -0.06
Gain [linear]
2
0 1 2 3 4 5 6 7 8 9 10
22
23
24
25
26
0.04
0.02 an 1.5
booster [a.u.]
1
27 0
28
dM
-0.02
0.5
29
-0.04
30
31 -0.06 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5
32 Time [s] Frequency [Hz]
33
34
35 Figure 4: (left) Signal filtered using a 4th order Butterworth high-pass (0.375 Hz) or by
36 proposed low-pass subtraction (1.6 Hz). (right) Linear gain of the two filters.
37
38
pte
39 In Figure 4, we see our example signal filtered using a 4th order Butterworth high-
40 pass (0.375 Hz) or by low-pass subtraction (1.6 Hz). The first has some of the frequency
41
42 content of respiration leaking into the signal. Conversely, our filter obtains unit gain near
43 the real heart rate and a boost peak at its 2nd harmonic. We have chosen a 4th order
44 Butterworth filter taking in consideration the desired boosting effect, its bandwidth and
45
46
how much it attenuated frequencies in the respiratory band in our specific population
ce
47 and with this specific SSP sensor, and tested the performance of the filter in the whole
48 estimation pipeline, in comparison with other filter typologies, both FIR and IIR.
49
50
51 2.5. Heart Rate estimation
52
Ac
53 We proceed to estimate HR from the SCardio approximating signal. Assuming its

54 fundamental frequency F 0 ≈ HR, we can reduce the heart rate estimation problem
55
56
to the identification of the signal’s main oscillatory component. In a noisy scenario,
57 DFT-based techniques are influenced by residual noise, so time-domain methods such
58 as the ACF are preferred. They were demonstrated in earlier works on speech analysis
59
60
(Strömbergsson 2016) or digital transmissions (Chaudhari et al. 2009) to be more
1
2
4
pt
5 robust to noise and have looser constraints on the signal’s stationarity. In its general
6 formulation the ACF is:
7
M
8 1 X
9 rx′ (τ ) = x(k) ∗ x(k + τ ), (2)
rx (0) k=1
cri
10
11 M
12
X
rx (0) = x(k) ∗ x(k). (3)
13 k=1
14
15 where x is our signal of interest, and τ represents the time lag between samples over
16
a window of M samples. The rx′ formulation is already normalized compared to the
us
17
18 original rx (τ ) ACF coefficients, dividing them by the global maxima rx (0). In a noise-free
19 situation, the autocorrelation function (ACF) decreases to zero as τ increases and has
20
local maxima at τ = fs /F 0, its multiples τ = fs /(k ∗ F 0), and dividers τ = fs /(F 0/k)
21
22 with k = N > 1. This representation is the principal distinction between DFT where the
23
24
25
26
27
28
an
fundamental frequency and its harmonics are distinct (given a frequency resolution high
enough), and ACF where they conflate. Consequently, we can only estimate the correct
frequency by searching the local maximum in a known range of interest [fbottom : ftop ]
in Hz, to be determined. The number of samples M is dependent on the largest period
dM
of oscillation (and consequently minimum frequency) that we want to observe correctly
29
30 as a local maximum. In our case the minimum samples will be ideally M > 2 ∗ fs /F 0.
31 Since the frequency to lag conversion lag(f ) = ⌊fs /f ⌉ introduces a discretization error,
32 we refine the lags’ search range by selecting those that would be closest to the original
33
34 boundaries when the inverse function f (lag) = lag(f )−1 is applied.
35 Additional steps are needed to improve the likelihood that the F 0 detected in the
36 ACF is truly representative of HR because residual noise may influence the shape of
37
38 ACF at certain lags, masking the peak of F 0. Our solution is to subdivide our 10s
pte
39 analysis window into smaller frames and shift the ACF function, to obtain multiple
40 sub-estimates Fˆ0(i). If the heart rate remains relatively stable and noise is limited, we
41
42 would expect that Fˆ0 distributes around the real F 0 as:
43 N
44 ˆ ∼ N( 1
X
45 HR Fˆ0(i), σ 2 ) + γ1 . (4)
N
46 i=0
ce
47
48 We already account for the fact that the distribution will have a variance σ 2 and a
49 non-zero skewness γ1 caused by noisier frames or short heart rate spikes. We will not
50 treat the problem of skewed distributions analytically, but each step of our algorithm
51 ˆ close to a non-skewed normal distribution.
52 applies some heuristics trying to keep HR
Ac
53 We identified N = 20 as the optimal number of frames using a parameter grid-

54 search (range N = [10 : 30], step size 2) on the entire dataset (the same applies to other
55
56
hyperparameters introduced here). The size M of each frame and subsequently the step
57 size between frames (step = (L − M )/(N − 1) with L = fs ∗ 10) is dependent on the
58 lower boundary of the search range as M = 2.0 ∗ lag(fbottom ). If the HR is low, the
59
60
ACF needs more samples to characterize it, but the corresponding peak in the ACF will
1
2
4
pt
5 change slowly over time. Instead, higher frequencies need fewer samples, but frames will
6 have a a higher overlap to better observe rapid variations.
7
Sometimes the search range may be misplaced, with local maxima falling on sub-
8
9 harmonics 1/2 ∗ F 0. Then, for each candidate Fˆ0(i), we extend the search range up
cri
10 to 2 ∗ Fˆ0(i) (or the first zero-crossing of ACF) to verify the presence of a larger local
11 maximum. If a new maximum exists and it exceeds 2 ∗ ACF (Fˆ0(i)), it becomes the
12
13 new sub-estimate. Both the M multiplier and the amplitude threshold for harmonics
14 were selected after a parametric grid-search (range [1.5 : 2.5] step 0.1 for both the
15 parameters). Nevertheless, there is a small but non-negligible chance that the extended
16
us
17 range and undetected noise in the specific time frame introduces some outlier estimates.
18 If these outliers cause a shift in the Fˆ0 skewness we remove them using the following
19 criteria based on the quantiles of Fˆ0:
20
21 (
22 Fˆ0(i) < Q1 − 1.5 ∗ IQR, if γ1 < 0
23
24
25
26
27
28
outliers =
an
Fˆ0(i) > Q3 + 1.5 ∗ IQR, if γ1 > 0
With Q1 and Q3 being the specific quartiles and IQR the interquartile range. We
(5)
identified a threshold in the skewness value of ±1.5 as the best performing in our dataset,
dM
although it does not remove all unbalanced estimates.
29
30
31 2.6. Heart rate estimate rejection
32
33 In some circumstances, the method presented fails to estimate the heart rate, and some
34 corrective techniques are needed to manage unreliable segments of the signal.
35
36 In segments where the sensor disconnects from the PSG recorder the value of ACF
37 is rx (0) ≃ 0, the autocorrelation function does not cross zero or it does not exhibit any
38 local maxima. In all three cases, we skip the estimation process.
pte
39
40 In other situations, short, large transient artifact negatively affects the methods
41 used to estimate the heart rate, so we opted for two levels of information rejection when
42 they are present. We detected artifacts using symbolic aggregate approximation (SAX)
43
44 technique (Keogh et al. 2005, Senin et al. 2014). SAX transforms a time series into a
45 sequence of characters using quantized levels, then artifacts (here called discords) are
46 selected as sub-sequences that are maximally different from all the others. A detailed
ce
47
48
description is available in Appendix B. For each analysis window, if artifacts account for
49 more than 12.5% of its samples, the entire window is rejected. The exclusion threshold
50 is a compromise between performance loss caused by artifacts, and coverage. If the
51
52
artifact samples are few but scattered across frames, we reject only the sub-estimates
Ac
53 that contain those samples and calculate the HR averaging the others. If these frames
54 are more than a threshold arbitrarily set at 3/4 ∗ N , we reject the entire window. For
55
56
both artifacts and disconnected sensors, we store the estimate as invalid (NaN, not-a-
57 number) and reset the search range for the next window at its default value.
58
59
60
1
2
4
pt
5 2.7. Search range tracking
6
7 The selection of the proper search range is crucial for our method. If the search range is
8 too wide, ACF noise will lead to bi-modal distributions of Fˆ0(i) and erroneous estimates.
9 Conversely, a too narrow range may not respond well to sudden variations in the heart
cri
10
11 rate caused by several phenomena during sleep, such as apneas and arousals (Azarbarzin
12 et al. 2014, Azarbarzin et al. 2021). We then distinguish two operational phases: an
13 initial bootstrap phase, where we do not know the heart rate nor can make proper
14
15 assumptions, and a tracking phase, in which we aim to keep the search range in a soft
16 spot centered around the average estimates.
us
17 The bootstrap phase comprises the first ten estimates (HR ˆ booth ) with a default
18
19 range of [0.6 : 1.8] Hz, identified by looking at the heart rate distribution extracted
20 from the ECG in the same initial interval. After the bootstrap phase we define an
21 initial floor estimate fbottom equal to median(HRˆ booth ) − 0.5 ∗ IQR(HRˆ booth ). This early
22
23
24
25
26
27
28
an
approximation is a reasonable guess of the bottom range of heart rate, which we will
update with subsequent estimates. In the tracking phase, we reduce the fbottom by a
small margin of 0.05 Hz, clipping at a minimum value of 0.6 Hz since frequencies lower
than that are unlikely in our sample population. If the bootstrap estimates are too
high, the margin allows for corrections of fbottom . The upper search limit ftop is instead
dM
29 dependent on a smoothed average of the last 10 HR estimates, as:
30
31 n−1
1 X ˆ
32 ˆ ˆ
HRsmooth (n) = αHR(n) + (1 − α) HR(i), (6)
33 10 i=n−10
34 ˆ smooth (n) + 0.25 Hz.
35 ftop (n + 1) = HR (7)
36
37
With n the nth estimate of the heart rate, α a smoothing factor equal to 0.1. We chose
38 both α and the 0.25Hz margin assuming that generally HR remains stable over time,
pte
39 but it can also increase suddenly and then return to the average value. This two factors
40
41
offer a compromise between having enough margin to observe spikes while minimizing
42 ˆ
the risk of the tracker deadlocks in higher search ranges. If the estimate HR(n) is
43 invalid we keep the last smoothed value.We opted to focus the HR tracking algorithm
44
45
on the control fbottom as it links to how many samples we use to calculate the ACF in
46 the sub-estimates. The value of fbottom is moved up or down to track slower changes in
ce
47 the average heart rate according to these rules:

48
49 ˆ smooth > fbottom + 0.64 ∗ (ftop − fbottom ) → fbottom + 0.01 Hz,
HR (8)
50
51 ˆ smooth < fbottom + 0.36 ∗ (ftop − fbottom ) → fbottom − 0.01 Hz.
HR (9)
52
Ac
53 We chose the two thresholds assuming that HR ˆ smooth should stay close to the center of
54 the search range. A shift towards fbottom or ftop invalidates this assumption and updates
55
56
the search range. We selected both thresholds empirically, optimizing for the minimum
57 variance of the estimation error.
58 Furthermore, participants are most likely awake in the initial phase of the night
59
60
(before sleep onset), and in this phase, heart rate remains relatively constant. After the
1
2
4
pt
5 person falls asleep, the heart rate slightly decelerates around 0.05 Hz∼3 bpm (Shinar
6 et al. 2006), then decreases further during the first half of the night. We introduced
7
a coarse correction mechanism trying to follow this deceleration. After the bootstrap
8
9 phase, for each window, we decrease fbottom by 1e − 5 Hz, roughly equivalent to a
cri
10 reduction of 1 bpm every 10 min of recording.
11
12
13 2.8. Signal separation
14
15 After heart rate estimation, we separate respiratory and cardiac signals to retain their
16 informative content. The separation happens on the original preprocessed signal before
us
17
attenuating the respiration.
18
19 We opted for a filtering method that employs the short-time fourier transform
20 (STFT) of the SSSP signal. Each time segment of the STFT corresponds to each window
21 in which we estimated the heart rate. The HR estimation tunes a filter applied to every
22
23
24
25
26
27
28
an
segment to obtain two distinct representations, one for SCardio and one for SResp . The
reconstruction of neighboring windows in the STFT domain will have a smooth change
in the cardiac frequency, similarly to time-variant adaptive filters.
At first, we calculate a frequency representation of the signal using the STFT as
XSSP = ST F T {Sssp }. Each time segment of the STFT is defined with the same length
dM
29 ˆ
and overlap (10s, 75%) used so far for the analysis. Then, for each estimated HR(n),
30
31 with n the iterator of all estimates in time, we design a FIR bandpass filter with 2048
32 ˆ
taps tuned from HR(n) to the 3rd harmonic and estimate its response Hf ilter (n) in the
33 ˆ
frequency domain. If HR(n) ˆ
was discarded we use the last valid value of HR(0 : n). A
34
35 single large bandpass filter guarantees gain = 1 in the Scardio bandwidth with a better
36 separation result. If other non-gaussian noise sources (e.g., very high harmonics of Sresp )
37 are present in-between the harmonics, they will leak into the final signal.
38
The signal Scardio is then filtered using the next formulation:
pte
39
40
41 Xcardio (n) = XSSP (n) ∗ |Hf ilter (n)|, (10)
42
43
Xcardio (n) = |Xcardio (n)| ∗ e(i∗∠Xcardio (n)+i∗θ(n−1)) . (11)
44
45 Here we opted for a FIR filter with a high number of taps to obtain the highest resolution
46 possible in Hf ilter (n) and to multiply it directly with XSSP (n). With θ(n) in Eq. 11
ce
47
48
accounting for small phase shifts that exist between the phase of Hf ilter (n − 1) and
49 ˆ
Hf ilter (n) at the frequency of HR(n). Now we consider Xcardio (n) as additive noise in
50 XSSP and therefore obtain Sresp as:
51
52
Xresp (n) = XSSP (n) ∗ (1 − |Xcardio (n)|/|XSSP (n)|). (12)
Ac
53
54
55 Additionally, Xresp is low-pass filtered (FIR 1024 taps) using the 3rd harmonic of the
56 respiratory rate RR(n) as the cutoff point to remove high-frequency noise higher than
57
58
that. We finally reconstruct the signals from the frequency domain to the time domain
59 using the inverse STFT and store them.
60
1
2
4
pt
5 2.9. Performance evaluation
6
7 To evaluate how well our method can estimate the heart rate from the SSP sensor, we
8 extracted the heart rate from the synchronous ECG as a ground truth. QRS complexes
9 are detected and localized with the ecg peaks method available in the NeuroKit2 toolbox
cri
10
11 (Makowski et al. 2021), checked for detection quality using the ecg quality method (based
12 on (Zhao & Zhang 2018)), and then visually inspected to correct misclassified ECG peaks
13 with the R-DECO software (Moeyersons et al. 2019). The HRecg is the average of the
14
15 interbeat intervals on the same 10s time windows used by our proposed method. If the
16 HR is outside a plausible physiological range (set broadly at [24 : 220] bpm) or the signal
us
17 quality is deemed too low, we discarded the window.
18
19 We pooled the estimations on the entire population to give a complete comparison
20 between HRecg and HRssp using Bland-Altman agreement analysis. Besides the average
21 estimation error (bias) and its deviation in beats-per-minute (bpm), we also considered
22
23
24
25
26
27
28
an
the percentage of estimates above 10 bpm and below 5 bpm, and the coverage, or
the amount of valid HRssp estimates over the valid HRecg measurements. Since some
estimation errors may be large but otherwise sparse in the signal, we also considered
their median and median absolute deviation (MAD) since these metrics are more
robust to outliers. The same metrics were calculated separately per recording to detect
dM
29 problematic SSP signals.
30
31 We calculated the skewness and kurtosis of the sub-estimates for each estimation
32 window and used a Shapiro-Wilk test to verify if they are normally distributed. We
33 compared the effect of having absolute skewness greater than one, excess absolute
34
35 kurtosis greater than 0.5, or non-normal distributions using the one-sided Wilcoxon
36 Rank-Sum test on the absolute estimation error.
37 To quantify how respiratory events dynamics impact our performance, we employed
38
the scored PSG annotations to separate the estimates obtained during normal breathing
pte
39
40 and during sleep disordered breathing events for each recording. Then, we subdivided
41 breathing events along with their onset and termination, where the SSP signal may
42
43 be turbulent, and during the event where it is relatively stable. We compared the
44 results using the Wilcoxon Rank-Sum test against normal breathing (null hypothesis:
45 the median difference is 0) and between event classes to evaluate if they have a larger
46
impact on error. To compensate for inter-recording variability, the effect is calculated
ce
47
48 as the difference between the median estimate error during events and during normal
49 respiration. Lastly, we assessed linear effects of age, body-mass index (BMI), and AHI
50
51
on our performance indexes using Kendall τ correlation coefficient to reduct the risk of
52 bias caused by model residuals. Potential effects of sex were assessed with two-sided
Ac
53 Wilcoxon Rank-Sum test.

54
55
56
57
58
59
60
1
2
4
pt
5 3. Results
6
7 3.1. Heart Rate estimation
8
9 Figure 5 shows the Bland-Altman agreement analysis on the entire population. The
cri
10 error bias is -0.06 bpm with a dispersion (standard deviation SD) of 5.09 bpm. Most
11
12 of the measured bias is near 0 bpm, with 85.49% of errors less than 5 bpm and only
13 6.36% over 10bpm. Specifically, the median error is 0.11 bpm with a median absolute
14 deviation of 0.43 bpm. With respect to ECG coverage, the average per-recording is
15
16 94.4%, with the first and third quartiles at 92.5% and 97.1%, respectively. 86 out of
us
17 100 recordings have a coverage higher than 90%. The average error of all the under-
18 estimated segments is −2.84 bpm, slightly larger than the over-estimated segments, with
19
20 an average of 2.03 bpm. In the same way, the maximum under-estimation error exceeds
21 over-estimation with a value of -64.65 bpm and 42.70 bpm, respectively.
22
23
24
25
26
27
28
an
dM
29
30
31
32
33
34
35
36
37
38
pte
39
40
41
42
43
44
45 Figure 5: Bland-Altman analysis plot for all the per-segment HR estimations. The
46 colorbar indicates the count of samples to better illustrate how they concentrate near
ce
47
48 the 0 bpm line. Average bias and dispersion of -0.06±5.09 bpm, median error, and
49 deviation 0.11±0.43 bpm.
50
51
52 Figure 6 illustrates the Bland-Altman analysis for each participant, with a median
Ac
53 bias of -0.12 bpm and a median deviation of 3.42 bpm. 94% of the recordings have an
54 average bias smaller than 2 bpm, while few recordings have a greater contribution to the
55
56
error seen in Figure 5, with high bias, large error dispersion, or both. Our method only
57 severely under-performs in four recordings, with more than 50% of estimates having an
58 error larger than 5 bpm. In one of these recordings the participants presented Cheyne-
59
60
Stokes breathing patterns for most of the night.
1
2
4
pt
5
6
7
8
9
cri
10
11
12
13
14
15
16
us
17
18
19
20
21
22
23
24
25
26
27
28
an
dM
29
30 Figure 6: Bland-Altman analysis of each recording separately with kernel density
31 estimate. Median bias between recordings is -0.12 bpm, median deviation (dash-dot
32 line) 3.42 bpm.
33
34
35
36 Of the 100 recordings analyzed, 8 fit the exclusion criteria presented in Section 2.1
37 and 2 are excluded because of persistent Cheyne-Stokes breathing patterns. After
38 excluding these recordings, the bias becomes -0.26 bpm with a standard deviation of
pte
39
40 3.88 bpm, with the median error reduced to 0.09 ± 0.36 bpm. The percentage of errors
41 under 5 bpm increases to 90.07% with only 3.24% above 10 bpm. All of the following
42 results do not exclude any recording to give a perspective on the whole experimental
43
44 population.
45
46 3.2. Assumptions on normality of estimations
ce
47
48 We assumed the sub-estimates of each window to be minimally skewed and normally
49
50 distributed, particularly after our correction. Concerning sub-estimates’ skewness,
51 84.93% of the windows considered have an absolute skewness lower than 1, with a
52 median value and standard deviation of 0.08 ± 0.67. We contained distributions with
Ac
53
54 absolute skewness above 1.5 using the method presented in Section 2.5, but we observed
55 that a minimum amount of estimates (1.14%) remains unaffected by the correction.
56 When the absolute skewness is above 1, the absolute estimation error is significantly
57
58 higher (p < 1e-04) with a median of 1.30 bpm versus 0.39 bpm.
59 The sub-estimates’ kurtosis (Pearson’s notation) is distributed around a median
60
1
2
4
pt
5 of 2.31 with an interquartile range from 1.88 to 2.99. It implies a slightly platykurtic
6 tendency of estimates, which may be an outcome of the outliers rejection presented in
7
Section 2.5. The sub-estimates with excess kurtosis are 74.2% of the total, but we did
8
9 not observe significant differences in the error.
cri
10 Using the Shapiro-Wilk test with a rejection α = 0.05, 42.35% of the valid windows
11 reject the null hypothesis, and they cannot be assumed to be normally distributed.
12
13 When the sub-estimates distribution is non-normal, the absolute error is significantly
14 higher (p < 1e-04) with a median of 0.87 bpm versus 0.34 bpm.
15
16
us
17 3.3. Effect of respiratory events
18
19 We can observe in table 2 that respiratory events introduce an estimation error smaller
20 1 bpm on average. However, some recordings report non-negligible differences compared
21 to normal respiration. The onset of an event contained the HR estimates from 5s before
22
23
24
25
26
27
28
an
to 2.5s after its start. Conversely, the end of an event is considered from 2.5s before
its termination to 5s after. Out of 100 participants, 2 did not have respiratory events
during the night, hence they are not considered here.
Table 2: Effect of respiratory events on method performance as difference with average

dM
29 normal respiration error. Number of observations: 98
30
31
32 Type median ∆ error [bpm] range IQR1 p
33
34
Event end -0.19 -11.97 : 4.04 0.83 <1e-05
35 Event onset 0.03 -8.18 : 10.66 0.20 0.054
36 During the event -0.04 -8.96 : 5.51 0.25 <5e-02
37
1
38 Inter-Quartile Range
pte
39
40
41
Although small, the median difference is significantly different from zero for event
42 transitions and during the events. A paired comparison between the different portions
43 of events suggests that our method performs significantly worse at the end and during
44
45
the respiratory events than at their onset (table 2). Upon closer inspection, we found
46 that the largest positive and largest negative outliers come from subjects that can be
ce
47 excluded according to the criteria in Section 2.1.

48
If we distinguish the absolute effect of apneic events, both obstructive and central,
49
50 and obstructive hypopnea events (table 3) the effect of apneas is significantly larger at
51 event onset and during the event, with an increased error of 0.24 bpm on average, but
52
not at event end. It must be noted that out of 98 participants considered, 33 had only
Ac
53
54 hypopnea events and no apneas, hence they are excluded from the comparison.
55 Although respiratory events lead to local errors in the estimation accuracy, we did
56
not observe strong correlations between the AHI of each participant and our performance
57
58 indexes. We converted the average and median bias to their absolute value, while other
59 variables are unchanged. The correlation coefficients are: average bias (τ of 0.16, p
60
1
2
4
pt
5 Table 3: Effect of apneic events (obstructive or not) and hypopneas on absolute error
6 as difference with average normal respiration error. Number of observations: 65
7
8
9
Apneas |∆ error| [bpm] Hypopneas |∆ error| [bpm]
Type p
cri
10 median IQR1 median IQR
11
12 Event end 0.27 1.21 0.46 1.14 0.662
13 Event onset 0.29 0.93 0.10 0.33 <1e-03
14 During the event 0.41 0.93 0.11 0.26 <1e-05
15
16 1
Inter-Quartile Range
us
17
18
19 < 0.05), error variance (0.17, p < 0.05), median bias (0.15, p < 0.05), median absolute
20
21
deviation (0.25, p < 0.001), percentage of errors above 10 bpm (0.19, p < 0.01), and
22 percentage of errors below 5 bpm (-0.22, p < 0.01).
23
24
25
26
27
28
3.4. Effects of demographic factors an
The manifestation of OSA in patients may differ significantly with age, sex or gender,
and BMI and consequently influence our method’s performance. After the exclusion of
dM
29 faulty recordings, our dataset includes 40 females and 50 males. None of our performance
30
31
indexes exhibited significant differences (p < 0.05) with respect to gender. We observed
32 minor correlations between performance and the age of the participant: average bias
33 (τ of 0.18, p < 0.05), error variance (0.16, p < 0.05), median bias (0.18, p < 0.05),
34
35
median absolute deviation (0.27, p < 0.001), percentage of errors above 10 bpm (0.21,
36 p < 0.01), and percentage of errors below 5 bpm (-0.27, p < 0.001). We did not observe
37 significant relationships between the BMI of participants and our performance indexes.
38
Other details of the participants may influence the quality of the signal, but are currently
pte
39
40 not accounted, such as skin elasticity, and other anatomical (e.g. fat depositions in that
41 area) or non-anatomical factors (e.g. mucus accumulations caused by chronic obstructive
42
43
pulmonary disease (COPD), physiological status of the lungs, body position etc.)”
44
45 3.5. Qualitative improvement of respiratory effort signal
46
ce
47 Figures 7a, 7b and 8 illustrate examples chosen to highlight how the signal changes after
48
49
separating Sresp and Scardio . The aim of these examples is to show how a well tuned
50 filter can lead to better visual evaluations of the signal.
51 Figure 7a shows that our method smoothens out the small cardiac oscillations in
52
Sresp , but the quality of Scardio is low despite the low estimation error. The spectrogram
Ac
53
54 shows that all three harmonics of the HR (0.84 Hz) are present, but another strong
55 signal at 1.12 Hz is the most likely cause of noise here.
56
57
In the obstructive apnea example in Figure 7b our method filters smaller cardiac
58 oscillations at the beginning and during the apnea, so that we can better understand
59 if there is a pattern of increased respiratory effort caused by the obstruction. Also,
60
1
2
4
pt
Proposed method
5 Normal breathing filtered SSP Sresp Preprocessed signal
6 0.04
7 0.02
[a.u.]
8 0
-0.02
9
cri
-0.04
10 0 5 10 15 20 25 30 35 40
11 Filtered SSP Scardio HR 50.6 bpm, error -0.351 bpm
12
0.01
13
[a.u.]
14 0
15 -0.01
16
us
0 5 10 15 20 25 30 35 40
17
SSP Scardio spectrogram [dB], minimum threshold -12dB
18 2
Frequency [Hz]
19 1.5 0
20 1 -5
21 0.5
22 -10
23
24
25
26
27
28
5 10 15 20
an
Time [s]
(a)
25 30 35
Proposed method
40
dM
Obstructive Apnea filtered SSP Sresp Preprocessed signal
29 0.04
30 0.02
31
[a.u.]
0
32 -0.02
33 -0.04
0 5 10 15 20 25 30 35 40
34
Filtered SSP Scardio HR 50.4 bpm, error 0.0463 bpm
35
36 0.01
37
[a.u.]
0
38
pte
39 -0.01
40 0 5 10 15 20 25 30 35 40
41 SSP Scardio spectrogram [dB], minimum threshold -12dB

2
42 5
Frequency [Hz]
43 1.5 0
44 1
-5
45 0.5
-10
46 5 10 15 20 25 30 35 40
ce
47 Time [s]
48
49 (b)
50
51 Figure 7: Examples with normal (a) and obstructed breathing (b) with separated signals.
52
Start and end of apneic event are marked with dotted lines.
Ac
53
54
55
56
57
58
59
60
1
2
4
pt
5 the cardiac signal shows a better resemblance with cardiac pressure waves (Dillon &
6 Hertzman 1941), qualitatively speaking. We can also see that respiratory overshooting
7
at the end of the event may introduce noise in the signal spread across cardiac harmonics.
8
9 The central apnea example in Figure 8 shows very interesting results. We can
cri
10 remove the cardiac oscillations from the central apnea, which results in a more accurate
11 representation of the full cessation of breathing effort. Although the cardiac information
12
13 during the event is fully retained in Scardio , the first half of this example (from 0 to 20s)
14 presents the same issues observed in the first example, with a higher spectral noise
15 between the fundamental frequency and the first harmonic, which disappears in the
16
us
17 second half.
18
19 Central Apnea filtered SSP Sresp
Proposed method
Preprocessed signal
20 0.04
21 0.02
[a.u.]
22
23
24
25
26
27
28
-0.02
-0.04
0.01
0
0 5 10 15
an 20 25
Filtered SSP Scardio HR 56.5 bpm, error -2.09 bpm

30 35 40
[a.u.]
dM
0
29
-0.01
30
31 0 5 10 15 20 25 30 35 40
32 2
SSP Scardio spectrogram [dB], minimum threshold -12dB
33 5
Frequency [Hz]
1.5
34 0
1
35 -5
36 0.5
-10
37 5 10 15 20 25 30 35 40
38 Time [s]
pte
39
40 Figure 8: Example with central apnea. Start and end of apneic event are marked with
41
42 dotted lines.
43
44
45
46 4. Discussion
ce
47
48
49
We describe a method to extract the cardiac signal present in SSP recordings and
50 separate it from the respiratory signal, retaining the information contained in both.
51 The goal was twofold: first, we aimed to improve the filtering of cardiac oscillations
52
so that respiratory events are better characterized in the resulting signal; second, we
Ac
53
54 aimed to retain the potentially informative content of the cardiac signal. To achieve
55 these goals we designed a processing system that estimates the heart rate fluctuations
56
in the signal, and exploits this information to tune a bandpass filter. This filter is
57
58 employed to separate the two signals of interest in the frequency domain. Heart rate
59 was estimated using the combination of time-domain representation of the signal in the
60
1
2
4
pt
5 form of autocorrelation function (ACF), and a set of domain-driven heuristics to track
6 and improve estimates during sleep.
7
We measured the quality of the method on the agreement between the SSP-
8
9 estimated heart rate and the equivalent measured from a synchronized ECG signal.
cri
10 Generally, the heart rate estimation error with the proposed method is low, with an
11 average bias of -0.06 bpm and a standard deviation of the error of 5.09 bpm over 10 s,
12
13 relatively constant over the entire range of measurements, and well contained, with more
14 than 90% of the recordings having an average bias smaller than 2bpm. The combination
15 of artifacts detection and outlier removals allowed us to obtain a very high coverage of
16
us
17 94.4% with respect to ECG-derived heart rate.
18 The quality of the estimation, or at which point the error can be considered large or
19 small, depends on the application of the heart rate signal, the sampling rate of the source,
20
21 and the intrinsic characteristic of the source signal itself. The authors in (Peláez-Coca
22 et al. 2022) have shown that signals like ours or photoplethysmography (PPG) signals do
23
24
25
26
27
28
an
not have well defined fiducial points like the ECG does, and consequently these signals
have some discrepancies with the ECG even at high sampling rates. Furthermore, for
certain applications the difference between pulse rate variability (e.g. from the PPG)
and heart rate variability (from ECG) may be non-negligible (Mejı́a-Mejı́a et al. 2020).
dM
29 Since this is the first time, to the best of our knowledge, that cardiogenic oscillations
30 are considered as a separate signal, there are no studies in literature examining their
31 physiological significance and their usage in beat-by-beat analysis applications, such as
32
33 pulse rate variability or the detection of abnormal beats. On the other hand, we believe
34 that our level of accuracy may be sufficient in applications where the heart rate dynamics
35 are smoothed over time, such as monitoring trends in heart rate or sleep staging (Bakker
36
37 et al. 2021).
38 In a broader multi-modal scenario, we believe that the information carried by the
pte
39 SSP signal could be used to improve other signals. For example, the SResp respiratory
40
41 signal may be used to reduce fluctuations visible in ECG and PPG signals to improve
42 the detection of heart-beats. Additionally, the peaks in the SCardio signal could be used
43 as candidates of pulse locations in majority voting fusion systems (after accounting for
44
45 pulse transit time effects) (Rankawat & Dubey 2017).
46 We also believe that our proposal opens new research developments in respiratory
ce
47 analysis and that a better comprehension of respiratory effort measurements can lead
48
49 to their widespread usage in clinical settings. Other than enriching the information
50 available in PSG recordings, we foresee the usage of the SSP sensor in novel integrated
51 home sleep apnea tests devices, with an improved characterization of respiratory effort
52
Ac
53 and upper airways dynamics.

54
55 4.1. Factors influencing the accuracy of heart rate estimation
56
57 In 10 out of 100 recordings the results were unsatisfactory, due to a combination of
58
59 high signal noise and sensor failures, large transients in the signal caused by turbulence
60
1
2
4
pt
5 associated with respiratory events, and technical shortcomings of our method.
6 The physical properties of the SSP sensor are one of the factors that influences
7
the quality of the raw signal. Requiring an airtight capsule makes it more prone to
8
9 disruptive effects caused by detachments or faulty sealing. While in some cases the
cri
10 signal loss is evident (Figure 1), we also observed drift, baseline wander, or baseline
11 jumps effects potentially caused by smaller leaks, and that are not correctly filtered
12
13 by our preprocessing. In both cases the system would require additional sensors that
14 measure the capsule status, body movements, or environmental effects, paired with
15 corrective algorithms, such as adaptive filters, source separation techniques or gain
16
us
17 correction methods at the sensor frontend. If a correction is not possible, it would
18 be necessary to quantify the reliability of each heart rate estimate and flag or remove
19 them if the quality is too low. Furthermore, analysing the residuals between the original
20
21 SSP signal and the separated SResp and Scardio signals after the estimates may help us
22 understanding how well the information content of the signal was captured
23
24
25
26
27
28
an
In addition to currently uncontrolled artifacts, multiple phenomena can drastically
and rapidly alter the temporal evolution of the SSSP signal. For example, the signal’s
dynamics can change, both with natural physiological mechanisms of sleep (e.g., heart
rate changes with sleep stages and arousals) and pathological respiratory events (e.g.,
dM
29 changes in breathing rate during apneas), or anatomical changes in the lungs, trachea
30 and heart that may depend on the age of the participant (see Section 3.4). This
31 situation imposes a constraint on the signal processing techniques we can apply, as
32
33 signal stationarity is quite limited in time. During the analysis, long time windows
34 provide more reliable estimates of low frequencies, but average heart rate may change
35 too rapidly. On the contrary, shorter windows would guarantee limited heart rate
36
37 fluctuations, but the characterization of slow-varying SResp before its subtraction would
38 be more challenging. Our selection of a 10s time window is a compromise between these
pte
39 two contrasting situations. As a future development of our algorithm, we foresee it

40
41 would be possible to detect apneic events on the SResp signal, and manipulate the time
42 window as necessary. For example, windows with higher overlap after apneic events, or
43 longer and less overlapped windows when the respiration and heart rate are changing
44
45 slowly.
46 Secondarily, we observed how respiratory events and abnormal respiratory patterns
ce
47 such as Cheyne-Stokes breathing increased the estimation error. The error measured at
48
49 the end of respiratory events may be caused by typical post-event phenomena, such as
50 hypercapnia, arousals, coughs, or gasps that introduce large artifacts in the SSP signal
51 and hide the effect of cardiac oscillations. Figure 9 shows an example of such events. An
52
Ac
53 obstructive hypopnea event accompanies two arousals and snoring events that impact
54 the signal quality, to the point that the resulting artifacts lead to the rejection of the
55 estimates. Although both hypopneas and apneic events had a significant effect on the
56
57
estimation error, we observed that hypopneas are less disruptive. We hypothesize that
58 the reduction in the airflow without complete obstruction does not introduce the same
59 extreme pressure swings visible during apneas.
60
1
2
4
pt
5
6
7
8
9
cri
10
11
12
13
14
15
16
us
17
18
19
20
21
22
23
24
25
26
27
28
an
Figure 9: SSP signal during a respiratory event with arousals (13s, 46s) and snoring
(65s,72s). The empty section in the HR error plot (60 to 70s) is due to an artifact in
the signal that caused an estimate rejection
dM
29 Some of the assumptions we made in the design phase of the system are invalidated
30 by the transient non-linearities of respiratory events and corrective measures are
31
32 necessary. For example the time window should update dynamically if respiratory
33 turbulence is detected, potentially leading to a better attenuation of the respiratory
34 signal. Furthermore, the boosting filter may be the culprit of these errors because of
35
36 its design. The boosting effect is relatively broad in frequency. On the one hand, a
37 broadband boost helps us as we do not need to identify the cutoff frequency perfectly to
38 catch the cardiac oscillations. On the other hand, the filter may also boost some noise,
pte
39
40 leading to uncertainties in heart rate detection later. The boosting filter needs also a
41 phase-aware design, as the current implementation introduces undesired phase shifts in
42 the range of boosted frequencies. Phase flips between harmonics are detrimental for
43
44 ACF representation, leading to uncertainties in the heart rate estimation.
45 There is also space to improve the heart rate estimation algorithm to address some
46 weak points of ACF representation. For example, using multiple representations of the
ce
47
48 signal and ensembling the results (Brüser et al. 2013), employing estimates of sleep
49 stages or respiratory events visible in the signal itself (Fonseca et al. 2015).
50 While we did not observe significant differences in the performance with respect
51
52 to sex and BMI, we observed small correlations with age and AHI. Given our specific
Ac
53 population and the known complex interplay of age, gender and BMI in the severity
54 of OSA, further tests on healthy elderly participants are required to understand the
55
56 nature of these relationships. Currently we can only assume that the strength or nature
57 of cardiogenic oscillations in the SSP signal may be partially influenced by the effects
58 of aging heart and vasculature.
59
60
1
2
4
pt
5 Another important consideration is that comorbid cardiac conditions (quite
6 common in OSA) may produce a discrepancy between the electric activity of the heart
7
and the manifestation of heart pulsations visible in arteries (Sološenko et al. 2017, Gil
8
9 et al. 2013). Further research is necessary to discriminate how irregular pulses of the
cri
10 heart affect the SSP signal. The average night-time heart rate of our participants is in
11 line with other OSA and disturbed sleep cohorts (Huang et al. 2018, Choi & Kim 2011).
12
13 Nevertheless, one participant maintained a HR> 100 bpm and remained awake for most
14 of the night. In this specific case, the correction for sleep-induced heart rate deceleration
15 likely introduced an unwanted drift. As we already know that the heart rate does not
16
us
17 decrease linearly for the entire night, a better mechanism could employ a more precise
18 prior of the deceleration curve, sleep staging information, or data from other sensors
19 (e.g., body movements) to know if the person is asleep or not.
20
21 Lastly, we observed how skewed distributions of Fˆ0 lead to larger errors. Other
22 than spurious elements in the signal, for example undetected artifacts, it may happen
23
24
25
26
27
28
an
that our algorithm introduces artificial errors at upper harmonics. When they are not
corrected, wrong sub-estimates introduce error creeps, both in the single estimate and
by drifting the heart rate tracking. Better selection algorithms are required to properly
compensate for skewed or bimodal distributions.
dM
29
30 4.2. Considerations on harmonic reconstruction error
31
32 The fact that the proposed method is resilient to small estimation errors (as seen in
33 Figure 8) opens the way to some considerations on what is the largest HR ˆ error tolerable
34
35 if we want to separate respiratory and cardiac signals. If we slightly overestimate HR,
36 the current transition bandwidth of the Scardio filter guarantees that, up to 5 bpm (or
37 0.1 Hz), the signal will stay over the -3dB threshold. The amplitude of the signal peaks
38
will not be usable as a feature, but it should be good enough to improve the estimation of
pte
39
40 HR through the detection of inter-beat intervals. On the contrary, HR ˆ underestimations
41 have a larger operative margin, thanks to full bandpass bandwidth from HR ˆ to 3 ∗ HR.
ˆ
42 ˆ the filtered Scardio will contain at least the fundamental
43 If the real HR < 2 ∗ HR,
44 frequency and 2nd harmonic, as mistaken for the 2nd and 3rd harmonic. However, two
45 limitations remain: every signal or noise with frequency HR ˆ < f < HR will leak into
46
Scardio . Similarly, if higher harmonics of Sresp are in present in that bandwidth, they
ce
47
48 will be filtered out. We will need more sophisticated filters to guarantee that signals’
49 separation does not disrupt valuable information when we reconstruct SCardio and SResp .
50
51
52 4.3. Physiological interpretation of the phenomenon
Ac
53
54 We can make some hypotheses on the physiological mechanism driving the cardiac
55 oscillations.While earlier studies connected cardiogenic oscillations to the vibration of
56
57 the heart through the lungs (West & Hugh-Jones 1961), researchers later demonstrated
58 that they derive from the pressure wave of the pulmonary artery near the trachea’s
59 bifurcation (Suarez-Sipmann et al. 2012). If this is the case, they would be more visible
60
1
2
4
pt
5 when the airflow is low, such as at the end of the expiration or in the presence of central
6 apneas, as illustrated in our examples. It must be noted that the experiments conducted
7
by Suarez-Sipmann and colleagues were performed on pigs, which may present slight
8
9 anatomical differences compared with human subjects. A second potential source in
cri
10 humans may be the aortic arch that runs left of the trachea, but we are unaware of
11 experimental studies directly supporting this. At the same time, since oscillations seem
12
13 dampened during normal breathing and amplified during apneas, they could also be
14 related to blood pressure variations. Different studies (Alex et al. 2017) showed that
15 blood flow velocity and pressure both increase during apneas. In this case, the vibration
16
us
17 of the carotid artery may be transmitted transversely through the still air in the trachea.
18 It is also possible that the temporal shift between the two is so small that they both
19 concur in the signal. Unfortunately, we cannot verify these hypotheses with the data
20
21 available at the moment. Future experiments with the same pressure sensor used to
22 detect the SSP signal or with a PPG sensor placed on the carotid artery could provide
23
24
25
26
27
28
an
the means to determine how the two signals are related.
Concerning the physiological role of cardiogenic oscillations in sleep-disordered
breathing conditions, some studies on rodents (Sullivan & Szewczak 1998, Dubsky
et al. 2018) showed a positive oxygen intake and lung air mixing as the result of increased
dM
29 cardiogenic activity during induced apneas or hibernation. We cannot be sure that
30 indirect air mixing represents an active protective mechanism against lowering oxygen
31 levels. Yet, we can hypothesize that strong cardiogenic activity may be a phenotypical
32
33 trait of OSA. For example, cardiogenic activity may correlate to higher pressure in
34 the pulmonary arteries. If this is the case, it may be a part of the bi-directional link
35 between OSA and pulmonary hypertension, as proposed by different authors (Sharma
36
37 et al. 2021, Mesarwi & Malhotra 2020). Similarly, the interaction between intrathoracic
38 pressure and cardiogenic activity may represent another indicator of the physiological
pte
39 mechanisms that govern cardioventilatory coupling (Sin et al. 2012). That would add
40
41 insights to the techniques (such as the loop gain) currently employed to characterize
42 patients with elevated cardiovascular risk (Edwards et al. 2019).
43 From a clinical perspective, multiple controlled experiments are necessary to unravel
44
45 the exact physiological mechanisms that manifest as cardiac oscillations in the SSP
46 signal. A better comprehension of the phenomenon could then lead to improved
ce
47 algorithms to analyze this information-rich signal.

48
49
50 5. Conclusions
51
52 This paper presents a method to estimate heart rate and extract the correlated cardiac
Ac
53
54 signal from a Suprasternal Notch pressure sensor during polysomnographic recordings of
55 people with suspected sleep disorders. Performance of heart rate estimation is promising,
56 and we observed definite qualitative visual improvements in the characterization of
57
58 respiratory events with correct filtering of cardiac oscillations. Future developments
59 will cover both technical improvements necessary to get more reliable estimates and a
60
1
2
4
pt
5 cleaner cardiac signal, but also to better understand the potential clinical applications
6 of the system. If the information carried by the SSP signal correlates with other cardiac
7
sensors like the PPG, it means that we could design new home sleep apnea tests systems
8
9 that need fewer sensors but provide a level of detail comparable with obtrusive and
cri
10 expensive PSG systems at a fraction of the cost. Potential coupling mechanisms between
11 the respiratory effort signal and cardiac oscillations may be complementary to all other
12
13 signals as a descriptor of the complex OSA dynamics. These can also represent an almost
14 direct probe of the activity of the pulmonary artery due to their cardiogenic nature, and
15 could be used, for example, to possibly quantify the risk of pulmonary hypertension in
16
us
17 OSA patients (Sharma et al. 2021).
18
19
20
Acknowledgements
21
22 This work was performed within the IMPULS framework of the Eindhoven MedTech
23
24
25
26
27
28
an
Innovation Center (e/MTIC, incorporating Eindhoven University of Technology, Philips
Research, and Sleep Medicine Center Kempenhaeghe), including a PPS-supplement
from Dutch Ministry of Economic Affairs and Climate Policy. Additional support by
STW/IWT in the context of the OSA+ project (No. 14619).
dM
29
30 References
31
32 Abrahamyan L, Sahakyan Y, Chung S, Pechlivanoglou P, Bielecki J M, Carcone S M, Rac V E,
33 Fitzpatrick M & Krahn M 2017 Diagnostic accuracy of level iv portable sleep monitors versus
34 polysomnography for obstructive sleep apnea: a systematic review and meta-analysis Sleep and
35 Breathing 22, 593–611.
36 Alex R M, Mousavi N D, Zhang R, Gatchel R J & Behbehani K 2017 Obstructive sleep apnea: brain
37
hemodynamics, structure, and function Journal of Applied Biobehavioral Research 22(4), e12101.
38
Amaddeo A, Fernandez-Bolanos M, Olmo Arroyo J, Khirani S, Baffet G & Fauroux B 2016 Validation
pte
39
40 of a suprasternal pressure sensor for sleep apnea classification in children Journal of Clinical
41 Sleep Medicine 12(12), 1641–1647.
42 Ayappa I A, Norman R G & Rapoport D M 1999 Cardiogenic oscillations on the airflow signal during
43 continuous positive airway pressure as a marker of central apnea. Chest 116 3, 660–6.
44 Azarbarzin A, Ostrowski M, Hanly P & Younes M 2014 Relationship between arousal intensity and
45 heart rate response to arousal Sleep 37(4), 645–653.
46 Azarbarzin A, Sands S A, Younes M, Taranto-Montemurro L, Sofer T, Vena D, Alex R M, Kim
ce
47 S W, Gottlieb D J, White D P et al. 2021 The sleep apnea–specific pulse-rate response

48 predicts cardiovascular morbidity and mortality American journal of respiratory and critical
49 care medicine 203(12), 1546–1555.
50
Bakker J P, Ross M, Vasko R, Cerny A, Fonseca P, Jasko J, Shaw E, White D P & Anderer P 2021
51
Estimating sleep stages using cardiorespiratory signals: validation of a novel algorithm across a
52
Ac
wide range of sleep-disordered breathing severity Journal of Clinical Sleep Medicine 17(7), 1343–
53
54 1354.
55 Berry R B, Budhiraja R, Gottlieb D J, Gozal D, Iber C, Kapur V K, Marcus C L, Mehra R,
56 Parthasarathy S, Quan S F et al. 2012 Rules for scoring respiratory events in sleep: update
57 of the 2007 AASM manual for the scoring of sleep and associated events: deliberations of the
58 sleep apnea definitions task force of the american Academy of Sleep Medicine Journal of clinical
59 sleep medicine 8(5), 597–619.
60
1
2
4
pt
5 Brochard L J 2014 Measurement of esophageal pressure at bedside: pros and cons Current Opinion in
6 Critical Care 20, 39–46.
7 Brüser C, Winter S & Leonhardt S 2013 Robust inter-beat interval estimation in cardiac vibration
8 signals. Physiological measurement 34 2, 123–38.
9 Chaudhari S, Koivunen V & Poor H V 2009 Autocorrelation-based decentralized sequential detection
cri
10 of ofdm signals in cognitive radios IEEE Transactions on Signal Processing 57, 2690–2700.
11 Choi S J & Kim J S 2011 Clinically different phenotypes of obstructive sleep apnea according to 24
12 hours heart rate tachogram pattern Sleep Medicine Research 2(1), 21–26.
13 Dillon J B & Hertzman A B 1941 The form of the volume pulse in the finger pad in health,
14
arteriosclerosis, and hypertension American Heart Journal 21, 172–190.
15
Dubsky S, Thurgood J, Fouras A, R Thompson B & Sheard G J 2018 Cardiogenic airflow in the lung
16
us
17 revealed using synchrotron-based dynamic lung imaging Scientific reports 8(1), 1–9.
18 Edwards B A, Redline S, Sands S A & Owens R L 2019 More than the sum of the respiratory events:
19 personalized medicine approaches for obstructive sleep apnea American journal of respiratory
20 and critical care medicine 200(6), 691–703.
21 Fonseca P, Long X, Radha M, Haakma R, Aarts R M & Rolink J 2015 Sleep stage classification with
22
23
24
25
26
27
28
Engineering 60, 3149–3155. an
ecg and respiratory effort Physiological measurement 36(10), 2027.
Gil E, Laguna P, Martı́nez J P, Barquero-Pérez Ó, Garcı́a-Alberola A & Sörnmo L 2013 Heart
rate turbulence analysis based on photoplethysmography IEEE Transactions on Biomedical
Glos M, Sabil A, Jelavic K S, Schöbel C, Fietze I & Penzel T 2018 Characterization of respiratory
events in obstructive sleep apnea using suprasternal pressure monitoring Journal of Clinical
dM
Sleep Medicine 14(3), 359–369.
29
30 Huang Z, Goparaju B, Chen H & Bianchi M T 2018 Heart rate phenotypes and clinical correlates in a
31 large cohort of adults without sleep apnea Nature and Science of Sleep 10, 111 – 125.
32 Keogh E, Lin J & Fu A 2005 Hot sax: Efficiently finding the most unusual time series subsequence in
33 ‘Fifth IEEE International Conference on Data Mining (ICDM’05)’ Ieee pp. 8–pp.
34 Lévy P, Kohler M, McNicholas W T, Barbé F, McEvoy R D, Somers V K, Lavie L & Pépin J L 2015
35 Obstructive sleep apnoea syndrome Nature Reviews Disease Primers .
36 Makowski D, Pham T, Lau Z J, Brammer J C, Lespinasse F, Pham H, Schölzel C & Chen S H A
37 2021 NeuroKit2: A python toolbox for neurophysiological signal processing Behavior Research
38 Methods 53(4), 1689–1696.
pte
39 Mejı́a-Mejı́a E, May J M, Torres R & Kyriacou P A 2020 Pulse rate variability in cardiovascular
40 health: A review on its applications and relationship with heart rate variability Physiological
41
Measurement 41(7), 07TR01.
42
Mesarwi O & Malhotra A 2020 Obstructive sleep apnea and pulmonary hypertension: A bidirectional
43
44 relationship Journal of Clinical Sleep Medicine 16(8), 1223–1224.
45 Moeyersons J, Amoni M, Huffel S V, Willems R & Varon C 2019 R-deco: an open-source matlab based
46 graphical user interface for the detection and correction of r-peaks PeerJ Computer Science 5.
ce
47 Mukhopadhyay S K, Zara M, Telias I, Chen L, Coudroy R, Yoshida T, Brochard L & Krishnan S

48 2020 A singular spectrum analysis-based data-driven technique for the removal of cardiogenic
49 oscillations in esophageal pressure signals IEEE Journal of Translational Engineering in Health
50 and Medicine 8, 1–11.
51 Peláez-Coca M D, Hernando A, Lázaro J & Gil E 2022 Impact of the ppg sampling rate in the pulse rate
52 variability indices evaluating several fiducial points in different pulse waveforms IEEE Journal
Ac
53 of Biomedical and Health Informatics 26, 539–549.

54 Rankawat S A & Dubey R 2017 Robust heart rate estimation from multimodal physiological
55
signals using beat signal quality index based majority voting fusion method Biomedical Signal
56
Processing and Control 33, 201–212.
57
58 Senin P, Lin J, Wang X, Oates T, Gandhi S, Boedihardjo A P, Chen C, Frankenstein S & Lerner M 2014
59 Grammarviz 2.0: a tool for grammar-based pattern discovery in time series in ‘Joint European
60
1
2
4
pt
5 conference on machine learning and knowledge discovery in databases’ Springer pp. 468–472.
6 Sharma S, Stansbury R, Hackett B & Fox H 2021 Sleep apnea and pulmonary hypertension: A riddle
7 waiting to be solved Pharmacology & Therapeutics 227, 107935.
8 Shinar Z, Akselrod S, Dagan Y & Baharav A 2006 Autonomic changes during wake–sleep transition:
9 A heart rate variability based approach Autonomic Neuroscience 130, 17–27.
cri
10 Sin P Y W, Webber M R, Galletly D C & Tzeng Y C 2012 Relationship between cardioventilatory
11 coupling and pulmonary gas exchange Clinical Physiology and Functional Imaging 32.
12 Sološenko A, Petrėnas A, Marozas V & Sörnmo L 2017 Modeling of the photoplethysmogram during
13 atrial fibrillation Computers in biology and medicine 81, 130–138.
14
Strömbergsson S 2016 Today’s most frequently used F0 estimation methods, and their accuracy in
15
estimating male and female pitch in clean speech. in ‘INTERSPEECH’ Dresden pp. 525–529.
16
us
17 Suarez-Sipmann F, Santos A, Peces-Barba G, Bohm S H, Gracia J L, Calderón P & Tusman G 2012
18 Pulmonary artery pulsatility is the main cause of cardiogenic oscillations Journal of Clinical
19 Monitoring and Computing 27, 47–53.
20 Sullivan S G & Szewczak J M 1998 Apneic oxygen uptake in the torpid pocket mouse Perognathus
21 parvus Physiological Zoology 71, 624 – 632.
22
23
24
25
26
27
28
9(11), e030996. an
van Gilst M M, van Dijk J P, Krijn R, Hoondert B, Fonseca P, van Sloun R J, Arsenali B, Vandenbussche
N, Pillen S, Maass H et al. 2019 Protocol of the SOMNIA project: an observational study
to create a neurophysiological database for advanced clinical sleep monitoring BMJ open
West J B & Hugh-Jones P 1961 Pulsatile gas flow in bronchi caused by the heart beat. Journal of
applied physiology 16, 697–702.
dM
Zhao Z & Zhang Y 2018 Sqi quality evaluation mechanism of single-lead ECG signal based on simple
29
30 heuristic fusion and fuzzy comprehensive evaluation Frontiers in physiology 9, 727.
31
32
33
34
35
36
37
38
pte
39
40
41
42
43
44
45
46
ce
47
48
49
50
51
52
Ac
53
54
55
56
57
58
59
60
1
2
4
pt
5 Appendix A. Cardiac oscillations under various respiratory conditions
6
7
8
40
40
40
40
9
cri
10
11
12
13
35
35
35
35
14
15
16
us
17
18
30
30
30
30
19
20
21
22
23
an
25
25
25
25
24
25
Obstructive Hypopnea
Obstructive Apnea
26
Central Apnea
27
Eupnea
Time [s]
28
dM 20
20
20
20
29
30
31
32
33
15
15
15
15
34
35
36
37
38
pte
39
10
10
10
10
40
41
42
43
ECG R-peaks
44
5
45
46
SSSP
ce
47
48
49
0
50
0
0
0.04
0.02
0.04
0.02
0.04
0.02
0.04
0.02
-0.02
-0.04
-0.02
-0.04
-0.02
-0.04
-0.02
-0.04
51 [a.u.] [a.u.] [a.u.] [a.u.]

52
Ac
53
54
Figure A1: suprasternal notch pressure (SSP) signal during normal respiration and
55 different respiratory events of a single person with visible cardiac oscillations. Start and
56 end of events are marked with vertical lines. Dots indicate the instant of the R-peaks
57
from synchronized ECG
58
59
60
1
2
4
pt
5 Figure A1 shows the SSP signal under different conditions: normal respiration
6 (ideally Eupnea), partial or fully obstructed breathing, and central apneas, where the
7
airway is still partially open, but the activity of respiratory muscles is absent. In all
8
9 cases, cardiac oscillations are visible at different levels. Normal respiration masks them
cri
10 entirely, except for the expiration phase with small visible inflections. They become more
11 evident during hypopneas, with distinctive inflections in the signal, and follow the same
12
13 pattern during obstructive apneas, although with a dampened amplitude profile. Lastly,
14 cardiac oscillations are dominant during central apneas, with a temporal profile quite
15 similar to the known photoplethysmography (PPG) shape (Dillon & Hertzman 1941).
16
us
17 The respiratory profile seems to shift towards cardiac oscillations in this example, even
18 before the onset of the apnea.
19
20
21 Appendix B. Artifact Rejection with HOT-SAX method
22
23
24
25
26
27
28
Approximation) method (Keogh et al. 2005, Senin et al. 2014).
an
We describe how we can estimate the presence of an artifact (here called discords)
using the HOT-SAX (Heuristically Ordered Time series using Symbolic Aggregate
We start with the sample visible in Figure B1 with a small artifact present. The
dM
29
method may be prone to false positives (given our choice in the parameters), but the
30 presence of some noise also in the electrocardiography (ECG) signal increases the chance
31 it is a real artifact.
32
33
S SSP with highlighted artifact
34 0.05
35
36
37
[a.u.]
38
pte
39 SSSP
40 -0.05 detected artifact
41 0 1 2 3 4 5 6 7 8 9 10
42 10-3 ECG
43
2
44
45 1
[a.u.]
46
ce
47 0
48
49 -1
0 1 2 3 4 5 6 7 8 9 10
50 Time [s]
51
52 Figure B1: SSP signal sample with synchronized ECG. The identified artifact is
Ac
53
54 highlighted
55
56
57 The algorithm starts by dividing the signal into N overlapping sub-sequences (with
58 1 sample step) of size winsize . This value determines the dimension of the discords
59 we will find. The optimal value depends on the sampling frequency, the length of the
60
1
2
4
pt
5 analysis window, the expected length of discords, etc. In our case winsize = 96 is
6 a reasonable compromise between granularity and coverage, as a value too large may
7
force us to reject many estimates and at the same time miss smaller artifacts. Each sub-
8
9 sequence is normalized subtracting the mean and dividing it by the standard deviation if
cri
10 larger than a threshold equal to 0.01. An example with winsize = 96 and winsize = 512
11 is included in Figure B2 and it hints already where large unexpected variations are.
12
13
14
15
16
us
17
18
19
20
21
22
23
24
25
26
27
28
an
dM
29
30
31
32
33 Figure B2: Representation of overlapping normalized sub-sequences with different sizes
34
35
36 Then each sub-sequence is quantized using Piece-wise Aggregate Approximation
37 and thresholds based on the normal probability distribution. The first step is to divide
38 the sub-sequence into equally spaced frames (in our case paasize = 5) and calculate
pte
39
40
the mean of each frame. Then the dimensionality is reduced by assigning a letter to
41 each frame according to the desired resolution. We opted for an alphabet with size
42 asize = 5. The choice of paasize and asize is not as critical as winsize . The original
43
44
authors observed that asize between 3 and 5 is good for most applications, while paasize
45 should be high if the signal has fast-changing dynamics. Figure B3 shows an example
46 on a single sub-sequence.
ce
47
48
Once all sub-sequences are converted into words, the heuristic of HOT-SAX starts
49 to look for discords sorting words from the least frequent, which is more likely to be a
50 unique discord. To identify if the sub-sequence represents a discord or not, it calculates
51
52
the euclidean distance between the original sub-sequence, normalized but not quantized,
Ac
53 and all other sub-sequences in random order. The sub-sequence which has the largest
54 distance with the others is the discord. If the parameters are tuned correctly, the
55
56
ordering heuristic provides a higher chance of finding the sub-sequence with the largest
57 distance early on.
58 Even with this heuristic, the research process can be time-consuming if we are
59
60
looking for small artifacts in a signal that is hours long. Therefore we sub-divided the
1
2
4
pt
5 2
6
7 1.5
e
8
9 1
cri
10
11 [n.u.] 0.5 d
12 c
13 0
14
15 -0.5
16 a
us
-1 a
17
18
-1.5
19 0 20 40 60 80 100
20 Samples
21
22
23
24
25
26
27
28
thresholds and alphabetic equivalence
an
Figure B3: Quantized representation of a single sub-sequence with highlighted
signal again in 10s windows. The shift is 10s if no discord is found, or the final instant
dM
of the discord otherwise. We apply a distance threshold to separate large discords
29
30
from false positives. We are still exploring the optimal combination of parameters,
31 as it sparsely happens that normal breathing waves are classified as discord, or that
32 windows with a lot of noise may be self-similar internally and therefore missed by this
33
34
method. Figure B4 shows two examples of the problem. Future optimizations will look
35 for parameters that balance heart rate (HR) estimation performance and coverage.
36
37
S SSP False positives
38 0.02
pte
39 0.01
40
0
41
[a.u.]
42 -0.01
43 -0.02
44 -0.03
45 5180 5190 5200 5210 5220 5230 5240
46 S SSP Missed artifacts

ce
47 0.02
48 0
49
[a.u.]
50 -0.02
51 -0.04 SSSP
52 detected artifacts
Ac
53 -0.06
12170 12180 12190 12200 12210 12220 12230
54 Time [s]
55
56 Figure B4: Two samples: (above) false positives during stable respiration, (below)
57
58 potential artifacts missed by the algorithm.
59
60
1
2
4
pt
5 Appendix C. Algorithm parameters
6
7 We describe here how different parameters in the algorithm operate and the motivation
8 behind their value. Variations of these parameters may be necessary if the method is
9
cri
10 applied on other populations, or during different physiological states that are not related
11 to sleep. All the parameters are fixed and do not change during the signal processing,
12 while tuned parameters do in accordance to certain rules and are noted in the table with
13
14 the symbol †. None of the parameters is tuned a-priori upon known characteristics of
15 the participant.
16
us
17
18
19
Table C1: Parameters employed in our algorithm.
20
21
22
23
24
25
26
27
28
Parameter
an
Initial value
Section 2.2
Motivation
dM
29
30
Small signal selection Elbow cutoff point in performance of our
90%
31 threshold dataset
32
33
34 Section 2.3
35
36
37 Compromise between respiratory and heart
38 Analysis window (L) 10s
rate stationarity (see also Section 4.1)
pte
39
Compromise between potential maximum
40
41 rate of change in heart rate during sleep,
Window step 2.5s
42 time resolution of estimates, and algorithm
43
execution speed
44
45 Reasonable margin for lower boundary of
46 SAudio lowpass cutoff 32Hz
audible audio frequencies (18 Hz : 20 kHz)
ce
47
48
49 Signal subsampling 256Hz Algorithm execution speed
50
51
52
Ac
53 Section 2.4
54
55
56
57 Continued on next page
58
59
60
1
2
4
pt
5
6 Table C1 – continued from previous page
7
8
9
cri
10 Parameter Initial value Motivation
11
12
13
Hypothesis of respiration rate harmonics
14 Respiration fmin range [0.4, 0.95] Hz during sleep between approximately be-
15 tween 12 and 30 breaths-per-minute
16
us
17 Hypothesis of respiration rate close to 20
18 fmin fallback candidate 0.66Hz
breaths-per-minute
19
20
Hypothesis on average heart rate observed.
21 Harmonics search range† [fmin , 2.4] Hz
22 See Section 2.1.
23
24
25
26
27
28
Harmonic
candidate
fallback
Respiration filter cutoff †

1.6 Hz
none
an Hypothesis on average heart rate observed.
See Section 2.4, function of the detected

dM
harmonics.
29
30 Optimization of booster bandwidth and
31 Order of respiration filter Butterworth 4th
32 signal gain. See also Figure C1
33
34
35 Section 2.5
36
37
38 Empirical optimization of algorithm perfor-
Number of frames N 20
pte
39 mance on ours dataset

40
41 Compromise between respiratory and heart
42 Size of window (L) f s ∗ 10
43 rate stationarity (see also Section 4.1)
44
45 Sampling criterion for ACF fbottom fre-
Size of frame (M)† 2.0 ∗ lag(fbottom )
46 quency resolution
ce
47 Initial value based on heart rate during

48 †
49 fbottom , ftop HR range [0.6, 1.8] Hz sleep 36 to 108bpm, then updated according
50 to Section 2.5 rules.
51
52 Outlier threshold Q3,1 ± 1.5 ∗ IQR Common choice for outliers removal
Ac
53
54
55 Empirical optimization of algorithm perfor-
56 Skewness threshold 1.5
mance on our dataset
57
58
59 Continued on next page
60
1
2
4
pt
5
6 Table C1 – continued from previous page
7
8
9
cri
10 Parameter Initial value Motivation
11
12
13
14 Section 2.6
15
16
us
17 Correspondent to half window step, arbi-
18 Rejection threshold 12.5% or 1.25s
trary choice
19
20 Arbitrary choice on good performance in
Sub-estimates rejec-
21 3/4*N out dataset in detecting sensor disconnec-
22 tion threshold
23
24
25
26
27
28
an tions.
Section 2.7
dM
median(HRˆ booth ) − Assumption on reasonable lower boundary
29 Bootstrap fbottom †
ˆ booth )
0.5 ∗ IQR(HR ˆ booth
close to 1st quartile of HR
30
31
32 † Arbitrary margin for potential corrections
fbottom margin fbottom − 0.05 Hz
33 of erroneous Bootstrap fbottom
34
35 See relevant text. Based on assumptions of
ftop margin † ˆ smooth (n) + 0.25 Hz
HR
36 heart rate dynamics during sleep and OSA
37
38
fbottom , ftop update Empirical optimization of algorithm perfor-
pte
39 [0.36, 0.64]
40 thresholds † mance on our dataset
41 Assumption on heart rate deceleration
42 fbottom deceleration †
1e-5 Hz/window during sleep. See Section 4 for relevant
43
44 issues.
45
46
Section 2.8
ce
47
48
49 Balance between optimal nfft for a 10 s
50
FIR and STFT window 2048 window (2560 samples) and resolution of
51
52 the filter response.
Ac
53
54
55
56
57
58
59
60
1
2
4
pt
5 Gain effect of Butterworth multiple orders and Chebyshev II @ 1.6Hz
6 4
7 BW1
BW2
8 3.5
BW3
9
cri
BW4
10 3 BW5
11 BW6
12 ChebyII
2.5
13
14
gain
15 2
16
us
17 1.5
18
19 1
20
21
0.5
22
23
24
25
26
27
28
0
0 0.5 1 1.5 2
an
2.5 3
Frequency [hz]
3.5 4 4.5
Figure C1: Comparison of boosting effect of Butterworth filter at different orders and
5
dM
29 Chebyshev II. See also Section 2.4 in Table C1
30
31
32
33
34
35
36
37
38
pte
39
40
41
42
43
44
45
46
ce
47
48
49
50
51
52
Ac
53
54
55
56
57
58
59
60
View publication stats

Cerinaetal 2023 Physiol. Meas. 10.1088 1361-6579 Acb12b

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cerinaetal 2023 Physiol. Meas. 10.1088 1361-6579 Acb12b

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Extraction of cardiac-related signals from a suprasternal pressure sensor

Article in Physiological Measurement · January 2023

Luca Cerina Gabriele Papini

SEE PROFILE SEE PROFILE

Pedro Fonseca Sebastiaan Overeem

SEE PROFILE SEE PROFILE

the role of orexin in goal-directed behaviour View project

reconfigurable architectures for fog computing View project

The user has requested enhancement of the downloaded file.

ACCEPTED MANUSCRIPT • OPEN ACCESS

Extraction of cardiac-related signals from a suprasternal pressure sensor

Manuscript version: Accepted Manuscript

View the article online for updates and enhancements.

This content was downloaded from IP address 131.155.142.187 on 22/02/2023 at 09:30

Abstract. Objective The accurate detection of respiratory effort during polysomnog-

0.02 0.8 ECG

53 We proceed to estimate HR from the SCardio approximating signal. Assuming its

53 We identified N = 20 as the optimal number of frames using a parameter grid-

47 the average heart rate according to these rules:

53 Wilcoxon Rank-Sum test.

Table 2: Effect of respiratory events on method performance as difference with average

47 excluded according to the criteria in Section 2.1.

41 SSP Scardio spectrogram [dB], minimum threshold -12dB

Filtered SSP Scardio HR 56.5 bpm, error -2.09 bpm

53 and upper airways dynamics.

39 two contrasting situations. As a future development of our algorithm, we foresee it

47 algorithms to analyze this information-rich signal.

47 S W, Gottlieb D J, White D P et al. 2021 The sleep apnea–specific pulse-rate response

47 Mukhopadhyay S K, Zara M, Telias I, Chen L, Coudroy R, Yoshida T, Brochard L & Krishnan S

53 of Biomedical and Health Informatics 26, 539–549.

51 [a.u.] [a.u.] [a.u.] [a.u.]

40 -0.05 detected artifact

46 S SSP Missed artifacts

Respiration filter cutoff †

See Section 2.4, function of the detected

39 mance on ours dataset

47 Initial value based on heart rate during

View publication stats

You might also like