A sleep stage estimation algorithm based on cardiorespiratory signals derived from a suprasternal pressure sensor

Automatic estimation of sleep structure is an important aspect in moving sleep monitoring from clinical laboratories to people's homes. However, the transition to more portable systems should not happen at the expense of important physiological signals, such as respiration. Here, we propose the use of cardiorespiratory signals obtained by a suprasternal pressure (SSP) sensor to estimate sleep stages. The sensor is already used for diagnosis of sleep‐disordered breathing (SDB) conditions, but besides respiratory effort it can detect cardiac vibrations transmitted through the trachea. We collected the SSP sensor signal in 100 adults (57 male) undergoing clinical polysomnography for suspected sleep disorders, including sleep apnea syndrome, insomnia, and movement disorders. Here, we separate respiratory effort and cardiac activity related signals, then input these into a neural network trained to estimate sleep stages. Using the original mixed signal the results show a moderate agreement with manual scoring, with a Cohen's kappa of 0.53 in Wake/N1–N2/N3/rapid eye movement sleep discrimination and 0.62 in Wake/Sleep. We demonstrate that decoupling the two signals and using the cardiac signal to estimate the instantaneous heart rate improves the process considerably, reaching an agreement of 0.63 and 0.71. Our proposed method achieves high accuracy, specificity, and sensitivity across different sleep staging tasks. We also compare the total sleep time calculated with our method against manual scoring, with an average error of −1.83 min but a relatively large confidence interval of ±55 min. Compact systems that employ the SSP sensor information‐rich signal may enable new ways of clinical assessments, such as night‐to‐night variability in obstructive sleep apnea and other sleep disorders.


| INTRODUCTION
Sleep clinics currently employ one or more nights of polysomnography (PSG) to determine sleep structure and assess the presence of potential sleep disorders.The recommended approach (Berry et al., 2012) in the measurement of sleep architecture relies on the manual or semiautomatic inspection of 30-s segments of physiological signals, discriminating several sleep stages, tracing back to the pioneering work done in the 20th century on electroencephalography (EEG) signals (Blake et al., 1939;Dement & Kleitman, 1957).However, there is growing evidence that many sleep disorders, like sleep-disordered breathing (SDB), circadian dysfunction, and insomnia, cannot be fully assessed during a single night of sleep (Khoo et al., 2011;Fossion et al., 2017;Chouraki et al., 2022).In these cases, standard PSG would be too burdensome in terms of human and technical effort, and novel means of measuring sleep become necessary.
Many current research efforts aim to automate sleep staging with a smaller subset of sensors.A recent review of sensing technologies outlines three major research domains outside clinical settings: brain activity measured with EEG (usually with as a single sensor), cardiac activity derived from electrocardiography (ECG) or photoplethysmography (PPG), and body movements, frequently combined with other sensors (Imtiaz, 2021).Respiration sensors, such as respiratory inductance plethysmography (RIP) belts and nasal airflow detectors, are the main component in only a fraction of research endeavours (Ebrahimi & Alizadeh, 2021).Two issues may motivate their absence.Respiration sensors are often bulkier and more expensive (e.g., mattress sensors and radar-based sensors) than other highly miniaturised and cheaper devices, to the point of characterising respiration through its effect on other physiological signals, like the ECG or wrist-PPG, rather than measuring the phenomenon itself (Papini et al., 2020).Additionally, the development of respiratory sensors frequently focuses on narrow applications or specific physiological mechanisms, such as detecting respiratory disturbances or estimating respiratory rate, ignoring the sleep architecture.
If we consider these applications of respiratory sensors, there is another advantage given by accurate sleep staging.There are different portable systems on the market to detect the presence of obstructive sleep apnea (OSA), known as home sleep apnea tests (HSATs), which allow the quantification of its severity through the apnea-hypopnea index (AHI), or the number of obstructive and central apnea and hypopnea events/h of sleep.However, the American Association of Sleep Medicine (AASM) allows the total recording time (TRT) as the denominator instead of total sleep time (TST), potentially underestimating OSA severity when long bouts of wakefulness are present (Berry et al., 2012).This underestimation is relevant in OSA populations due to the high prevalence of comorbid insomnia (30%-38%, Sweetman et al., 2019).
We strongly believe that a point of convergence for the two domains, the estimation of sleep stages and the detection of respiratory events, exists.There are known interactions between cardiorespiratory activity and sleep (Imtiaz, 2021;Penzel et al., 2007), with some works showing promising results with respiratory signals alone (Yang et al., 2016).Additionally, new respiratory sensors may close the gap with other wearable devices present on the market, e.g., suprasternal pressure (SSP) sensors (Meslier et al., 2002).In its simplest form, the sensor consists of a pressure transducer placed on the notch above the sternum inside an airtight capsule, which detects variations in intrathoracic pressure caused by respiration.Previous literature demonstrated its usefulness in analysing SDB, as the SSP sensor signal is a faithful representation of respiratory effort.During airway obstruction, inspiratory effort increases with each breath, with the SSP sensor signal showing larger negative pressure waves (Glos et al., 2018).
Other than respiration, another physiological mechanism is visible in the SSP sensor signal as an artefact caused by the mechanical activity of the heart, therefore named cardiogenic oscillations (Adams et al., 1993;Glos et al., 2018).These oscillations are sometimes used as an indication of central apneas when no respiratory activity is present but otherwise ignored as a source of noise (e.g., when they are masking respiratory fluctuations, as in Figure 2).Yet, cardiogenic oscillations can be a direct measure of cardiac activity.We hypothesise it is possible to use the SSP sensor to perform sleep staging, employing its high-quality respiratory effort signal and cardiogenic oscillations.
We present a method to estimate sleep stages using cardiorespiratory signals extracted using only the SSP sensor.Our proposal takes advantage of the presence of cardiogenic artefacts as a source of information instead of noise.We previously demonstrated (Cerina et al., 2023)  The SSP sensor is not present in all the recordings of the SOMNIA database, and these 100 recordings were collected during the years 2020-2021 in a sub-project regarding new methods of respiratory effort measurement.This subset comprises a single recording of each individual participant.Table 1 shows the demographics of participants together with the TRT.Participants with known diagnoses of cardiac arrhythmia or with known usage of beta-blockers at the date of PSG were excluded a priori from this subset.Each recording includes 30-s epoch sleep stages manually scored by trained sleep technicians at the Center for Sleep Medicine Kempenhaeghe (Heeze, the Netherlands) according to AASM rules (Berry et al., 2012), which we employed as the ground truth for sleep staging and TST estimates.
We combined sleep stages N1 and N2 and relabelled them as N1-N2.
As participants underwent PSG for suspected breathing disorders, we also highlighted their whole night AHI.Using the AASM thresholds that define OSA severity according to AHI, our subset included 18 patients with no OSA (AHI <5 events/h), 19 with mild OSA (AHI 5-15 events/h), 23 moderate (AHI 15-30 events/h), and 40 severe (AHI >30 events/h).The respiratory events are scored using standard PSG sensors and AASM guidelines.
The SOMNIA dataset covers a wide variety of sleep disorder diagnoses, frequently overlapping with SDB, and our subset is representative of that, as visible in Table 2. Diagnoses are performed and annotated in the database according to International Classification of Sleep Disorders, third edition.

| Suprasternal pressure sensor
We employed a custom-designed SSP sensor, consisting of a pressure transducer inside a three-dimensionally printed capsule secured to the skin notch above the sternum with an adhesive patch, which also guaranteed airtight separation from the external environment.
Figure 1 shows the placement of the sensor and a visual impression of its physical dimensions.The variations of air pressure in the trachea move the skin, which then compresses or expands the air in the capsule, exciting the sensing element.The sensor measures only relative pressure changes, as it is not zeroed during the PSG setup.The SSP sensor records its signal with a sampling frequency (f s ) of 1024 Hz and filters it at the front-end with a low-pass filter at 285 Hz.The signal's pre-processing involved a high-pass filter at 0.15 Hz to remove unwanted baseline drifts and a low-pass filter at 32 Hz that removes powerline interference and potential noise sources in the audio range (e.g., snoring sounds).The signal is then re-sampled at 256 Hz using a finite impulse response (FIR) polyphase filter.We will refer to this signal as 'S SSP ' and assume this to be composed by three components: with S Resp representing the respiratory effort, S Cardio containing the pressure fluctuations caused by cardiac activity, and ε representing residual noise not suppressed by pre-processing.

| Separation of SSP sensor signal components
We processed the signal using the methods described in our previous work to separate the cardiac component in the SSP sensor signal (Cerina et al., 2023).In short, the system attenuates respiratory com- obtained through SSP sensor-derived HR estimates, and : jopt for optimal HR estimates.It is important to note that this approach does not imply the usage of an ECG in the sleep staging process or in the derivation of the IHR signal; we still detected the IBIs on the SSP S Cardio peaks, which may differ substantially from R-peaks intervals observable on the ECG.

| Signal quality selection
During the night, it may happen that the SSP sensor sealing loosens or the participant touches it, dramatically decreasing the quality of the signal.In those cases, the method still outputs a sleep stage estimate, F I G U R E 3 Example of a recording scored with our proposed method (S Respjopt signal with IHR opt ) and the equivalent manual scoring.The selected recording is representative of the average demographic in our data.REM, rapid eye movement sleep; W, Wake.
but its accuracy may be compromised.We included a parameter that describes how reliable the signal is for each epoch using an experimentally determined threshold dependent on the full range of the sensor.An epoch is valid if at least 95% of the samples are >1% of the sensor range (À0.25, 0.25 mV in our specific sensor).Henceforth, we will designate as 'coverage' the percentage of valid epochs over the total observed in the PSG scoring.
We opted to exclude from the results those obtained in recordings with a very low coverage.The exclusion threshold has been set only for extreme cases with a coverage of <20%, as the performance degradation may be dependent on other mechanisms that are still unexplored and on which we can only speculate at the present time.
With this criterion, we excluded nine of 100 recordings.A representation of the phenomenon is shown in Figure 4. Supplemental data S2 shows an example of a recording with high coverage and one excluded due to low coverage.

| Statistical analysis
In our primary analysis, we evaluated the epoch-by-epoch estimation accuracy and Cohen's kappa agreement for different levels of estimation complexity and clinical relevance (Djonlagic et al., 2020): (1) complete resolution (Wake/N1-N2/N3/rapid eye movement sleep [REM]); (2) Wake, non-REM and REM discrimination (Wake/ NREM/REM); and (3) binary comparison of Wakefulness versus Sleep.Each stage was also evaluated separately in binary comparisons (e.g., N1-N2 versus N3, REM, and Wake together) and results are presented in detail in Supplemental data S1, Tables S1.1 and S1.2.In the case of binary comparisons, we also calculated the sensitivity, specificity, and positive predictive value (PPV).The recordings were scored by different technicians, but all from within a single centre, and moreover, an internal institutional assessment yielded an high average agreement of 85.6% (range 83%-88%), so scorer-specific differences are unlikely to have relevant influence on the present findings.
For each level of estimation complexity we conducted four experiments (combinations from here on) with: the original S SSP signal without filtering cardiac oscillations and no IHR, the filtered S RespjSSP signal without IHR, the two signals combined S RespjSSP and IHR SSP , and finally S Respjopt signal with IHR opt .Ideally, the quality and informative content of the signals improves with this order, and so should the performance of the sleep staging.Under the hypothesis that additional and better information would improve the performance, we evaluated each combination against the previous one using one-sided Wilcoxon's signedrank test at 5% significance.As the sleep staging architecture can also operate using only the IHR signal, we included an overview of its performance in Supplemental data S1.Here, we highlighted only the contribution of the respiratory effort and its combination with the novel SSP sensor-derived IHR.
To measure the agreement between the manual TST and that estimated using the SSP sensor, we summed all the epochs predicted as sleep after sleep onset and before the final wake-up and measured bias and 95% confidence intervals (CIs) using Bland-Altman agreement analysis.
While the CReSS algorithm always outputs a stage estimate, there are segments in the manual scoring that are not classified (and labelled as 'U').In the SOMNIA dataset these segments occur only at the beginning of the recording, when the participant receives instructions and goes to bed, and after waking up.We considered these segments irrelevant to the problem at hand and excluded them from the calculation of Cohen's kappa, TST, and coverage.
Lastly, we quantified the potential effects of age and the clinical severity of breathing sleep disorders (employing manually scored AHI), considering a linear relationship and quantifying the correlation index R 2 .We tested the effects of biological sex with the Mann-Whitney U test.

| Sleep stage estimation
Table 3 shows the performance of all the different combinations with different levels of estimation complexity (number of sleep stages discriminated) on the selected set of 91 recordings.We denote as combination the type of filtering applied on the original signal and the source of the HR signal.Starting from a comparison between the original unfiltered respiratory effort signal (S SSP ) and the same signal with filtered cardiac oscillations removed (S RespjSSP ), we can observe that both kappa and accuracy improved in the four and three classes estimates.In the wake/sleep experiment, the median kappa slightly decreases, although not significantly, while accuracy increases.
Once we include the IHR, both with sub-optimal (IHR SSP ) and optimal estimates (IHR opt ), we can observe significant improvements in the highest level of discrimination (Wake/N1-N2/N3/REM), with the median kappa going from 0.533 to 0.636 (+19.3%), and the F I G U R E 4 Cohen's kappa as function of coverage in four class discrimination with S RespjSSP + IHRS .Separation cut-off at 76% of coverage.median accuracy going from 71.9% to 77.5%.Similar results were found in the three class discrimination task (Wake/NREM/REM), with the median kappa going from 0.591 to 0.699 (+18.3%) and the median accuracy going from 80.8% to 86.0%.
In the wakefulness versus sleep task we can notice that the performance is high even with the non-filtered respiratory signal.However, the filtering process and the addition of the cardiac signal remains beneficial, with significant improvements in all performance indices.Kappa improves from 0.622 to 0.713 (+14.6%),accuracy from 88.9% to 92.2%, sensitivity from 80.4% to 84.2%, specificity from 92.9% to 95.6%, and PPV from 68.2% to 77.1%.The distinction of wakefulness and sleep is also relevant for certain indices of sleep quality, namely the sleep onset latency, the duration of wake after sleep onset, and the sleep efficiency.We included relevant results in the supplemental data S1, Table S1.4.
The frequency of respiratory events showed only a slight influence on Cohen's kappa performance, with the R 2 correlation coefficient between À0.19 and À0.21 ( p < 0.05), independently from the signals' combination or the sleep staging task.The tasks in Table 3 were selected according to their clinical relevance, such as discriminating wakefulness from sleep, or the time onset of specific sleep stages.We did not observe any significant effect of age and biological sex on any performance index.
We compared the distributions of Cohen's kappa of the recordings below a decreasing coverage cut-off against the recordings above that using the Mann-Whitney U test.We then observed that under 76% coverage, all the performance indices were significantly worse (α = 0.01) and decreased almost linearly, as seen in Figure 4.The complete results, including the whole population of 100 participants and all stage-specific comparisons, are available in Supplemental data S1, Tables S1.1 and S1.2.While the exclusion of nine subjects lead to small improvements in performance, none of the indices showed a statistically significant improvement when tested with oneside Mann-Whitney U test (α = 0.01).

| Total sleep time estimation
In our population, the TST scored by sleep technicians on the selected 91 recordings had a median (interquartile range [IQR]) of 432 (381-472) min (note that this value differs from those in Table 1 due to the nine excluded recordings).The distribution of estimated median (IQR) TST with the current working combination S Respjssp , IHR ssp was 432 (387-468) min.The discrepancy with manual TST, calculated as TST manual À TST SSP =TST manual ð Þ Â 100 with Wilcoxon test was not significant (p = 0.15) with a median (IQR) of À0.58% (À4.35% to 2.01%).
Despite the large error spread, the results were better framed in comparison with the TRT.In this case the discrepancy with the manual TST was significant ( p < 1e-5) with a median (IQR) error of 16.3% (7.9%-29.6%).The Bland-Altman agreement (Figure 6) shows a bias of 81.2 (95% CI 0-192.4)min.The TRT was always larger than TST by construction, therefore it could only overestimate the TST.

| DISCUSSION
We describe a method to estimate sleep stages using cardiorespiratory signals extracted from SSP sensor recordings.We believe that automating sleep staging using integrated cardiorespiratory sensors,  F I G U R E 6 Bland-Altman agreement of total sleep time TST (min) based on total recording time versus manual scoring.Bias 81.26 (95% CI 0-192.4)min.The underestimation confidence interval is not representative due to construction of TST, always lower than total recording time.TST, total sleep time.
the SSP sensor significantly improved the sleep staging performance, with an effect proportional to the quality of the HR estimation.To achieve these goals, we combined the output of an algorithm designed to separate respiration and cardiac activity the SSP sensor signal with a neural network previously trained with comparable sensor modalities, namely the respiratory effort derived from RIP belts and HR derived from ECG or PPG.
We measured the quality of our method as the agreement between estimated sleep stages and the scoring of a trained sleep technician.We also included accuracy, sensitivity, and specificity metrics for binary class estimation tasks.Considering 91 recordings with enough data coverage, we observed a median Cohen's kappa agreement close to 0.6 for four classes estimation and slightly higher (0.661 and 0.679) for three classes and for wakefulness versus sleep estimation.An agreement above 0.6 is generally considered satisfactory, with a recent meta-analysis of inter-rater agreement among human scorers that found an average Cohen's kappa of 0.76 (Lee et al., 2022).
We know from our previous work that respiratory events may influence (among other factors) the HR estimation, with a cascading effect on the quality of the filtering and the performance of the sleep staging.
Nevertheless, our method performed well independently of the presence or the frequency of respiratory events, with Cohen's kappa only slightly decreasing as this frequency increases (À0.001 per unit of AHI, p < 0.02).We observed that SSP sensor cardiac signal filtering improves sleep staging performance, even with imperfect HR estimates during the separation phase.The optimal HR estimation and subsequent detection of IBIs in the separated SSP Cardio signal led to a further boost in performance.We want to remark that IHR opt is not equivalent to IHR ecg as we still detect heartbeats and IBIs on the SSP sensor signal, but an exemplification of the results obtainable with a perfect HR estimator.This result supports the hypothesis that cardiogenic oscillations are a valid and interesting signal and not an artefact to be eliminated.
It is important to remember that the outcome of the sleep staging process, also known as hypnogram, is almost always an intermediate step in the quantification of other clinically relevant parameters, such as sleep onset or REM phase latencies, time awake during the night, number of sleep cycles, etc.A high Cohen's kappa is a desirable outcome, but only if it is framed correctly for the problem at hand.In this work, we considered the TST estimation as an important element in the evaluation of SDB.We observed a low bias of À1.38 min, but a relatively large CI close to 55 min, both in over-and underestimation.Further improvements are needed, but we can frame these results under the potential applications of the estimated sleep stages, e.g., using our estimated TST in a hypothetical HSAT device.Employing the TRT would lead to a bias of À6.1 events/h, with an underestimation up to À26.5 events/h in our population, with notable detrimental effects on the quality of the diagnostic process.Conversely, using the estimated TST as the AHI denominator leads to a bias of only À0.61 (95% CI À10.6 to 9.42) events/h.These results reinforce the need for updated requirements in Type 3 diagnostic devices, particularly for people with suspected insomnia (Hermans et al., 2020).
In nine recordings, we observed a very low coverage of the signal during the night, potentially caused by sensor issues rather than the effect of a physiological mechanism.The specific experimental conditions (a real-world clinical PSG routine versus a controlled laboratory setting) prevent us from obtaining a perfect signal, but several explanations are possible regarding how a low signal's coverage may influence the performance of our algorithm.For example, a high percentage of low amplitude samples will skew the normalisation of signals before the neural networks.Additionally, if the sensor does not measure information reliably that would affect the HR estimation (Cerina et al., 2023), the detection of IBIs, and the quality of the respiratory effort signal itself.Future versions of the CReSS neural network could also learn an undefined class that mimics epochs deemed unscorable by sleep technicians and then discard them.In real-world scenarios, we foresee the presence of multiple quality assessments (e.g., coverage, signal-to-noise ratio of respiration and cardiac signal, excessive movements) prior to the proposed method to guarantee that the recording is good enough for a reliable analysis.
Other than coverage, different factors may explain the limitations of our method, both from a technical and from a physiological or conceptual point of view.The recordings in the SOMNIA dataset are collected in a third-line diagnostic centre for sleep medicine, and therefore only a low percentage of the participants had no diagnosis of a sleep disorder.Further tests are warranted on a more general population to ensure the consistency of our results.Our method relies on the HR estimation performed directly on the SSP sensor signal to drive the separation filter.We know that this estimate is not perfect due to sensor issues, partial detachments, low signal-to-noise ratio, transient motion artefacts, or respiratory disturbances (Cerina et al., 2023).All these issues harm the quality of the separated signals and degrade the performance of the whole method.The effect of imperfect estimates is visible and quantifiable when we separate the signals with an optimal estimate, with significant improvements across all performance indices.Future developments will focus on improving the mechanical case of the sensor to minimise external noise and on methodological updates necessary to achieve better HR estimates.
For example, employing all the harmonics of the cardiac signals may improve the detection of IBIs and subsequently of the IHR, but it may also introduce beat detection errors if the SNR of S Cardio is not high enough.A second option is to use a multi-modal system, including known cardiac sensors such as the ECG to measure the IHR or inertial sensors to tackle motion artefacts.This option comes at the expense of increased system complexity, cost, and comfort for the patient.
Similar devices on the market followed the path of integration in polygraphy (Cidelec PneaVoX) or standalone systems (Acurable Acupeb- how to separate respiratory and cardiac signals, achieving reliable heart rate (HR) estimates and better interpretability of the respiratory effort signal.Once separated, we extract the instantaneous HR (IHR) from the cardiac signal.Then, together with the clean respiratory effort signal, we use the IHR as the input of a previously developed cardiorespiratory sleep staging algorithm (Bakker et al., 2021), with the notable advantage of employing a single sensor instead of three (further details in Section 2.4).The present study presents three contributions towards the development of SSP sensor-based systems for sleep staging.(1) To show that the SSP sensor respiratory effort signal can provide reliable estimates of sleep staging in agreement with manual scoring, and how removing cardiogenic oscillations leads to a direct performance improvement; (2) to show how the novel extracted cardiac signal could further improve the performance, proportionally to the quality of the estimated IHR; and (3) to quantify the reliability of our method in the estimation of the TST and how that could affect potential applications.This work also aimed to inspire the development of novel home sleep polygraphs that require less sensors, are more comfortable for the user, and ideally cost less, without compromising on the quality of the physiological signals recorded. 2 | METHODS 2.1 | Dataset We employed data from a set of full single-night PSG recordings which are part of Sleep and Obstructive Sleep Apnoea Monitoring with Non-Invasive Applications (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 synchronised recording of the SSP sensor.
ponents using a spectral estimate of the respiration rate then it estimates the HR as the fundamental frequency in the autocorrelation domain of overlapping 10-s segments of the S SSP signal.Two complementary filterbanks tuned on these estimates are then used to separate the S SSP signal into S Resp and S Cardio .The filtering of S Resp includes both the subtraction of HR harmonics and a low-pass filter set at three times the estimated respiratory rate (around 0.7-1.2Hz) to remove high-frequency noise or residual components in the audio range related to snoring or tracheal sounds.Figure2shows an example of an SSP sensor signal with strong cardiogenic oscillations and the two output signals.The improvement in the S Resp respiratory effort signal is notable, with a distinct breathing waveform that would be difficult to detect reliably in the original SSP sensor signal.While our original algorithm included upper harmonics of the HR, here we employed only the fundamental frequency to tune the filters, resulting in a simpler S cardio signal comprising a single frequencyvarying sinusoid.We detected potential heartbeats as local maxima in S Cardio to obtain a better temporal resolution of the HR compared to initial estimates.From the heartbeats, we calculated the inter-beat intervals (IBIs), removing the ones outside a physiological range of 0.3 s ($200 beats/min [bpm]) and 2 s ($30 bpm).We determined the IHR SSP from S Cardio IBIs and sampled it at 10 Hz with previous value interpolation.An IHR value of 0 is set for excluded IBIs.In the illustrative example in Figure2, we can see how the IHR obtained from the S Cardio signal is close to that obtained with the simultaneously recorded ECG.To further explore how our method would perform with perfect HR estimates, we also employed the HR measured on the ECG signal to tune our filter banks.Following the same process as above, we obtain two other S Resp and S Cardio signals, detect peaks and a new IHR signal.From here on, we will use the suffix : jSSP to specify signals Instantaneous heart rate @ 10Hz SSP-derived IHR ECG-derived IHR F I G U R E 2 Example of suprasternal pressure (SSP) sensor signal with large amplitude cardiogenic oscillations.After the separation of S Resp and S Cardio , heartbeats are detected on the S Cardio signal and converted into instantaneous heart rate.a.u., arbitrary unit; b.p.m., beats/min; ECG, electrocardiography.
Our method extends a previously developed algorithm for CardioRespiratory Sleep Staging (CReSS;Bakker et al., 2021).The algorithm was originally trained on different datasets and validated on publicly available Multi-Ethnic Study of Atherosclerosis (MESA) and Sleep Heart Health Study (SHHS) datasets (Chen et al., 2015; Quan et al., 1997), using respiratory information from airflow (nasal cannula or thermistor), respiratory effort (thoracic or abdominal RIP belt), and cardiac information, concretely IHR derived from ECG or PPG.The algorithm was developed as such that any arbitrary combination of a subset of these input signals can be used, with the inputs that are not used simply set to a constant value of zero.In the present study, we used S Resp as the respiratory effort input and the SSP sensor-derived IHR, ignoring the airflow signal.The IHR and the respiratory effort signals are resampled at 10 Hz and normalised between the 0.1% and 99.9% percentiles.The two signals feed into a neural network model, which estimates a sleep stage for every 30-s non-overlapping segment, corresponding to the epochs used for manual sleep scoring.The CReSS algorithm consists of two parts: feature extraction and sleep stage classification.The feature extractor is a neural network architecture comprehending three convolutional neural network (CNN) layers and three bidirectional long-short-term memory (LSTM) layers.The neural network extracts 64 cardiac features from the IHR signal, plus 44 generic respiratory features and 20 effort-specific features from the respiratory effort signal.The original architecture includes 64 flow-derived features from another set of CNN and LSTM layers that we ignored in the present study.The set of 128 features is the input of a sleep staging architecture, composed of fully connected layers that reduce their dimension to 64 features, a LSTM layer, and a Softmax layer that estimates sleep stage probabilities for each epoch.The final output is the label with the highest probability.An example of the sleep stages estimated with our method and the equivalent manual scoring is shown in Figure 3.The selected recording is representative of the average demographic in our data in terms of age, TST and AHI and with a good overview of sleep stages and sleep cycles.
such as the SSP, represents an uncharted area of research compared to the usage of signals derived from non-respiratory sensors.While sensors that detect respiratory activity by proxy are valuable and T A B L E 3 Performance indices for sleep stage discrimination experiments, N = 91.Results presented as median (interquartile range).Metric Original signal S SSP Filtered, effort only S RespjSSP Filtered effort and IHR, filters tuned with SSP-derived HR S RespjSSP ,IHR SSP Filtered effort and IHR, filters tuned with ECG-derived HR S Respjopt ,IHR opt 4 classes: Wake/N1-N2/ ble), which to the best of our knowledge do not use the signal for sleep staging, so the field of research remains open.For example, movement and HR data could be extracted from wearable devices (e.g., smart wristbands) to complement or extend the SSP sensor signal or employ ensemble learning techniques to merge the results of multiple sleep staging models.Furthermore, it could be valuable to explore the usage of sleep staging models and respiratory event detectors synergistically, with sleep stage data informing the probability of a respiratory event, or vice versa the presence of a respiratory event informing the probability of certain sleep stages.An example would be the addition of nasal airflow sensors, already present in most HSATs and also employed in the original manifestation of the CReSS algorithm.Even with additional sensors, we must note that there is a hypothetical upper bound, yet to be reached, to cardiorespiratory-based staging's performance.As sleep originates at the brain level, its expression on the cardiac and respiratory activity depends on the complex interaction between the central and autonomous nervous systems and the innervated organs.Internal factors, such as age and pathologies, and external factors, such as medications or alcohol, influence sleep at central and peripheral levels.A better understanding of cardiorespiratory coupling mechanisms is necessary to solve this issue.Concerning the CReSS algorithm, it is useful to differentiate the contribution of respiratory effort and IHR to the model.While the IHR signal is agnostic to the source used, being it the SSP, ECG etc., the original model was trained on RIP-derived respiratory effort.It is not only plausible, but probable that the RIP-and the SSP sensorderived effort signals show differences in the relationships across features employed by the neural network, which then lead to performance degradation.Additionally, it will be important to verify the reliability of the IHR signal and its contribution to the overall estimation performance.For example, in the S RespjSSP , IHR SSP combination, the four classes Cohen's kappa increases compared to S RespjSSP alone but in 24 out of 91 recordings it worsened the performance.In those cases, it would be valuable to determine the reliability of portions (or the entirety) of the IHR signal to be excluded with automatic methods, and employ only the respiratory effort signal.5 | CONCLUSIONSThis paper presents a novel application of the SSP sensor in sleep stage estimation using cardiorespiratory signals, expanding its current field of usage, and demonstrating the usefulness of respiratory-based sleep staging systems.The performance of automatic sleep staging is promising and with a moderate to substantial agreement with manually scored hypnograms, independently from the level of sleep stage discrimination desired, and increasing following the improvement in quality of the respiratory and cardiac signals.Preliminary results on TST estimation demonstrate its usefulness for future wearable devices and the necessity to update guidelines for Type 3 diagnostic devices, particularly for a population with prevalent concomitant insomnia, such as people with SDB disorders.
Abbreviations: IHR, instantaneous heart rate; PPV, positive predictive value; REM, rapid eye movement; SSP, suprasternal pressure sensor signal.aIndicatesa statistically significant improvement ( p < 0.05) versus previous combination (left to right).bIndicates a significant worsening.widespread,thedirect observation of the physiological phenomena of interest, enriched by reliable sleep stage estimates, is a highly desirable goal.Here we demonstrate that it is possible to estimate sleep stages using only the respiratory effort signal measured by the SSP sensor and how filtering the cardiogenic oscillations improved the performance considerably.Also, we demonstrate how the aforementioned cardiogenic oscillations are not mere measuring artefacts but contain valuable information.Including the IHR signal extracted fromF I G U R E 5