Photoplethysmography as a single source for analysis of sleep-disordered breathing in patients with severe cardiovascular disease


Offer Amir, MD, Heart Failure Center, Cardiology Department, Lady Davis Carmel Medical Center, Haifa 34323, Israel. Tel.: +972-50-626-5567; fax: +972-48625056; e-mail:


Sleep-disordered breathing and Cheyne–Stokes breathing are often not diagnosed, especially in cardiovascular patients. An automated system based on photoplethysmographic signals might provide a convenient screening and diagnostic solution for patient evaluation at home or in an ambulatory setting. We compared event detection and classification obtained by full polysomnography (the ‘gold standard’) and by an automated new algorithm system in 74 subjects. Each subject underwent overnight polysomnography, 60 in a hospital cardiology department and 14 while being tested for suspected sleep-disordered breathing in a sleep laboratory. The sleep-disordered breathing and Cheyne–Stokes breathing parameters measured by a new automated algorithm system correlated very well with the corresponding results obtained by full polysomnography. The sensitivity of the Cheyne–Stokes breathing detected from the system compared to full polysomnography was 92% [95% confidence interval (CI): 78.6–98.3%] and specificity 94% (95% CI: 81.3–99.3%). Comparison of the Apnea Hyponea Index with a cutoff level of 15 shows a sensitivity of 98% (95% CI: 87.1–99.6%) and specificity of 96% (95% CI: 79.8–99.3%). The detection of respiratory events showed agreement of approximately 80%. Regression and Bland–Altman plots revealed good agreement between the two methods. Relative to gold-standard polysomnography, the simply used automated system in this study yielded an acceptable analysis of sleep- and/or cardiac-related breathing disorders. Accordingly, and given the convenience and simplicity of its application, this system can be considered as a suitable platform for home and ambulatory screening and diagnosis of sleep-disordered breathing in patients with cardiovascular disease.


Sleep-disordered breathing (SDB) and Cheyne–Stokes breathing (CSB) occur frequently in patients with cardiovascular diseases, including coronary artery disease, hypertension, stroke and heart failure (Hanly et al., 1989). The prevalence of SDB in patients with heart failure is approximately 50% (Amir et al., 2010a–c; Hanly et al., 1989; Javaheri et al., 1998; Lanfranchi et al., 2003). Moreover, SDB and CSB were shown to have prognostic value for hospitalization and mortality in heart failure patients (Amir et al., 2010a–d; Ancoli-Israel et al., 2003; Hanly and Zuberi-Khokhar, 1996; Javaheri et al., 2007; Lanfranchi et al., 1999). Studies have shown that treatment of SDB can improve sleep quality as well as cardiovascular parameters (Flemons et al., 2004; Khayat et al., 2009). Accordingly, for purposes of both diagnosis and treatment, considerable effort has been devoted to the detection of patients with sleep disorders (Flemons et al., 2003).

A key parameter in assessing SDB is the apnea–hypopnea index (AHI). Based on overnight monitoring, the American Academy of Sleep Medicine (AASM) Task Force classified the sleep apnea–hypopnea syndrome as moderate in individuals with an AHI higher than 15 (American Academy of Sleep Medicine Task Force., 2007; Iber et al., 2007). The gold-standard diagnostic test for SDB, as well as for measuring AHI and several other prognostic parameters such as total sleep time and CSB, is overnight multi-channel polysomnography (PSG). This, however, is a time-consuming, expensive and labor-intensive method, requiring a certified technician for its operation. It is also uncomfortable for the patient because of the large number of sensors needed for data recording. These disadvantages seriously limit the suitability of full PSG for assessing the true prevalence of SDB, especially in patients with heart failure and other debilitating diseases. Many such patients are thus kept from receiving optimal treatment. Not surprisingly, therefore, SDB is undetected in the majority of affected patients. In the Wisconsin Sleep Cohort Study, for example, it was reported that as many as 93% of women and 82% of men with moderate to severe sleep apnea were previously undiagnosed (Hanly and Zuberi-Khokhar, 1996). There is therefore a genuine need for a simpler and less expensive procedure than full PSG for SDB screening, especially for sick individuals such as those with heart failure.

The aim of the present study was to evaluate an automated analysis for SDB detection using signals derived from a pulse oximeter attached to the patient’s finger. The regular pulse oximeter is based on two raw data signals: photoplethysmograph (PPG) and saturation. The PPG signal is obtained using optical methods and measures the volumetric changes of the organ. The saturation is the common estimation of the oxygen saturation level. Both signals are measured and recorded with the pulse oximeter device. The assumption was that if such a system could detect SDB reliably, it might obviate the need for a complex sleep laboratory study while still providing an early indication of SDB requiring treatment. The sleep apnea algorithm (SAA) tested here for SDB diagnosis is based on an analytical software platform developed by Morpheus Ox, WideMed Ltd (Herzliya, Israel). To validate the SAA with the corresponding analyses obtained by the gold standard, full PSG was carried out by a certified technician.



The study population comprised 74 patients (54 men and 20 women above the age of 18 years), of whom 60 had been admitted to the cardiology department at Morristown Memorial Hospital, Morristown, NJ, or Overlook Hospital, Summit, NJ because of cardiac complaints that included acute decompensation heart failure. The remaining 14 were ambulatory patients referred to the Atlantic Health Sleep Center, the sleep laboratory at Morristown Memorial Hospital, for clinical overnight PSG because of clinically suspected SDB. Each participant patient underwent a complete sleep study of 8 h connected to recording electrodes. The inclusion criteria were: age 18 years or older, admission into either the cardiology department or the sleep center, ability to perform a PSG full study test for a complete night and willingness to sign the informed consent. Patients were excluded from the study if they were unable to undergo overnight PSG, were receiving oxygen or had a life expectancy of <1 month at the time of recruitment. The study was approved by the Institutional Review Board of Morristown Memorial Hospital. All patients signed informed consent prior to their participation in the study.

Experimental procedures

The computer is connected to a monitoring device that records the patient’s physiological activity, which is reflected by saturation and PPG raw signals, and on which the diagnostic analysis of the data is based. In the present study the AHI and CSB parameters were measured and recorded by the SAA. The analysis of respiratory events and periodic breathing patterns is derived from the oxygen saturation signal of a NoninOEM III module. This signal is associated with the respiration of a patient. The signal processing is indicative of start and end times of apnea episodes. For that purpose the analysis produces values for each event or episode, the start and end of every event, which are the maximum and minimum of the signal, reduction duration and the decrease amount of the saturation in percentages. Only reduction episodes which are ≥3% and detected as physiological are taken into account for the next phase of the analysis. The second-phase analysis tries to locate periodicity between the events detected in the previous phase. Time shifts between recurrences of the features are computed and processed in order to generate an output including a plurality of bands corresponding to different multiples of a cycle length of the apnea episodes. Periodic events were classified as central, while all others were classified as obstructive events. The sleep and wake algorithm is based on the morphology and rhythm of breaths detected from the PPG signal. In order to determine the sleep and wake area the algorithm extracts a number of independent parameters that are taken into consideration. The measure for similarity between two adjacent breaths based on morphology and rhythm is the first parameter. A strong correlation between adjacent breaths is an indication for sleep episode. The second parameter is the complexity of the rhythm; achieving a constant rhythm is associated with sleep. Periodic desaturation events are the third parameter, and are regarded as sleep episode. Noise event detection serves as the final parameter, while all noise and movement events are tagged as wake. Each parameter or feature is modeled using the Gaussian mixture model (GMM) probability density function (PDF) during sleep and during wake. A Bayesian classifier using the likelihood ratio test (LRT) is executed to discriminate sleep and wake epochs based on their GMM PDFs.

Each of the 60 patients in the hospitals and 14 patients in the sleep laboratory underwent overnight full PSG. Standard full PSG signals were acquired with the SOMNOscreen System Sleep-Monitoring System (SOMNOmedics GmbH, Am Sonnenstuhl, Randersacker, Germany). In addition to PPG and saturation signals, the full PSG included electroencephalography (EEG), electrocardiography (ECG), electromyography (EMG), eyes (EOG), respiratory flow, respiratory thermistor, respiratory abdominal/thorax piezo-electric bands, leg movement and body position. The raw data were analyzed independently by two certified technicians, and in cases of disagreement the decision was made by a third scorer. The interscorer variability between the two scorers for the AHI parameter was 75–80% and for CSB was 85%. The intrascorer variability was approximately 85–90% for both AHI and CSB. Sleep stages, respiratory events (apnea and hypopnea), event classification (central, obstructive or mixed) and CSB were all scored according to the most recent recommended guidelines of the AASM (AASM 2007; Iber et al., 2007) and were referred to as ‘gold standard’. Of note, these guidelines define a sleep apnea event as a reduction of more than 90% in nasal flow for at least 10 s, and a hypopnea event as a reduction of at least 30% in nasal flow for at least 10 s accompanied by a decrease of at least 4% in saturation. The AHI was calculated as the number of apneas and hypopneas that occurred during sleep, divided by the total sleep time. For classification into central and obstructive events, the technician systematically scored CSB and sporadic central episodes as ‘central’ and all other respiratory events as ‘obstructive’.

During the PSG procedure, the subset of PPG and saturation signals was extracted simultaneously and analyzed using the SAA. Scoring by this software platform enabled us to detect and measure the duration of CSB, detect respiratory events (AHI), classify these events as ‘central’ or ‘obstructive’, measure total sleep time and identify sleep/wake epochs. To determine the accuracy of the SAA in assessing SDB and the feasibility of using it in clinical practice, the SDB analysis acquired from this automated PSG subsystem was compared with that obtained by full PSG. Based on a comparison of the data derived from the SAA with the corresponding data obtained from the gold standard test, we calculated the sensitivity and specificity of the SAA for the various parameters measured.

Data were analyzed using sas® version 9.1 (SAS Institute, Cary, NC, USA). Descriptive statistics are presented for demographic data, whereas continuous data are represented by mean values ± standard deviation (SD), and categorical data by a numerical value and a percentage.


Data sets were obtained for all 74 patients. Their ages ranged from 27 to 100 years (mean ± SD: 64.6 ± 14.3; median 64 years), and their body mass index values ranged from 18 to 79 kg m2 (mean ± SD: 30.9 ± 8.7; median 30 kg m2). Of the 74 patients of the study population, 37 patients (50%) had coronary artery disease (CAD), 22 (30%) had diabetes mellitus, 14 (19%) had renal insufficiency (defined as serum creatinine > 1.5 mg dL−1) and 41 (55%) had hystory of hypertension.

All 74 patients underwent CSB analysis. In 10 patients, however, the EEG recording was incomplete because the apparatus became detached. Gold-standard (PSG) analyses of apnea, hypopnea, AHI, sleep/wake epochs, total sleep time and event classifications could not be derived for those subjects, and accordingly the data used for comparisons with the Morpheus Ox results were from 64 patients.

Detection of Cheyne–Stokes breathing

Each patient underwent a sleep study in which the presence or absence of CSB was determined by the two experimental methods. Table 1 presents a cross-tabulation of CSB detection by SAA and by gold-standard PSG in our patient population (= 74). Results are presented in a 2 × 2 matrix to facilitate sensitivity and specificity calculations. CSB was detected by PSG in 37 patients (50%). As can be seen from the table, the sensitivity of CSB detection by SAA was 34 of 37 = 92% [95% confidence interval (CI): 78.6–98.3%) and specificity was 35 of 37 = 94% (95% CI: 81.3–99.3%).

Table 1.   Comparison between the number of patients in whom Cheyne–Stokes breathing was detected by the SAA system and by gold-standard PSG, based on 74 patients
 Full PSGTotals for SAA (n)
Positive (n)Negative (n)
  1. PSG, polysomnography; SAA, sleep apnea algorithm; n, number of patients.

 Positive (n)34236
 Negative (n)33538
Totals for full PSG373774

Figs 1 and 2 present regression and Bland–Altman diagrams depicting the comparison between the SAA and the PSG findings. The regression coefficient was 0.94 and the majority of points in the Bland–Altman diagram were within the range of ±2 SD, confirming that CSB detection by the SAA system was as accurate as its detection by the gold standard test.

Figure 1.

 Linear regression of Cheyne–Stokes breathing (CSB) duration obtained from assessments by the Morpheus Ox system and by full polysomnography (PSG). The correlation coefficient (r) of the two curves is 0.94. Most of the apnea–hypopnea values are located close to the regression line.

Figure 2.

 Bland–Altman diagram showing Cheyne–Stokes breathing (CSB) duration measured by the Morpheus Ox system and by full polysomnography (PSG). Most values can be seen within the confidence interval of 2 standard deviations.

AHI measurement

The distribution of AHI values assessed by SAA compared to the gold-standard PSG. The gold standard test recorded a wide distribution of AHI values, ranging from 0 to 88 with a median of 29. The SSA test presented a distribution of AHI values ranging from 0 to 80 with a median of 31.

Table 2 lists measurements of AHI cutoff levels (as 15 or higher –‘positive’, or below 15 –‘negative’) for the SAA system and for gold-standard PSG. Comparison of the results shows that the sensitivity of the cutoff-level measurement by SAA was 39 of 40 = 98% (95% CI: 87.1–99.6%) and specificity was 23 of 24 = 96% (95% CI: 79.8–99.3%). The overlapping confidence limits demonstrate that the sensitivities and specificities of AHI cutoff level detection by the SAA were similar. Figs 3 and 4 present regression and Bland–Altman diagrams for the AHI, showing a regression coefficient of 0.92 and a majority of points within ±2 SD in the Bland–Altman presentation. Table 4, as well as Figs 3 and 4, thus confirm the accuracy of AHI assessment by the SAA when compared to the gold standard assessment.

Table 2.   Assessment by the SAA system and by full PSG of the sensitivity and specificity of the measured AHI cutoff level in 64 patients as ≥15 (‘positive’) or <15 (‘negative’)
 Full PSGTotals for SAA (n)
Positive (n)Negative (n)
  1. PSG, polysomnography; SAA, sleep apnea algorithm; n, number of patients.

 Positive (n)39140
 Negative (n)12324
Totals for full PSG402464
Figure 3.

 Linear regression of the apnea–hypopnea index (AHI) obtained from assessments by the Morpheus Ox system and by full polysomnography (PSG). The correlation coefficient (r) of the two curves is 0.92. Most of the AHI values are located close to the regression line.

Figure 4.

 Bland–Altman diagram showing the apnea–hypopnea index (AHI) measured by the Morpheus Ox system and by full polysomnography (PSG). Most values can be seen within a confidence interval of 2 standard deviations.

Table 4.   Comparison of respiratory event classification (‘central’ versus ‘obstructive’) between the SAA system and full polysomnography in 64 patients
 Full PSGTotals for SAA (n)
Obstructive (n)Central (n)
  1. PSG, polysomnography; SAA, sleep apnea algorithm; n, number of patients.

 Obstructive (n)183310882921
 Central (n)62550905715
Totals for full PSG245861788636

Detection and classification of respiratory events

Table 3 presents a comparison of the detection of respiratory events (apnea and hypopnea), on an event-by-event basis, by the SAA system and by the gold standard. Results are presented in a 2 × 2 matrix to facilitate sensitivity and specificity calculations. As can be seen from the table, the sensitivity of respiratory event detection by SAA was 8636 of 10 700 = 81% (95% CI: 80.0–81.5%) and the positive predictive value was 8636 of 10 609 = 81.4% (95% CI = 80.7–82.1%). The false positives (1973 respiratory events that can be seen in Table 5) are events that the SAA scored as respiratory events while the gold standard did not. Those events were scored by the SAA as events, as they had a saturation reduction of 4%; however, they did not show a reduction component in the full PSG flow and effort channels and for that reason were not scored as events by the gold standard. The false negatives (2064 respiratory events shown in Table 5) are apnea events which did not have a saturation reduction of 4% and therefore could not be scored by SAA.

Table 3.   Comparison between the number of respiratory events present in 64 patients, as detected by the SAA system and by full polysomnography
 Full PSGTotals for SAA (n)
Positive (n)Negative (n)
  1. PSG, polysomnography; SAA, sleep apnea algorithm; n, number of patients; NA, not available.

 Positive (n)8636197310 609
Totals for full PSG10 700197312 673
Table 5.   Epoch-based comparison of sleep/wake evaluation by the SAA system and by full polysomnography based on 64 patients
 Full PSGTotals for SAA
  1. PSG, polysomnography; SAA, sleep apnea algorithm.

 Wake17 444554622990
 Sleep973428 64838 382
Totals for full PSG27 17834 19461 372

Table 4 presents the results of respiratory event classification (‘central’ or ‘obstructive’) by SAA and by the gold standard in 64 patients. Results are presented in a 2 × 2 matrix to facilitate sensitivity calculations. As can be seen from the table, the sensitivity of detection by SAA was 5090 of 6178 = 82% (95% CI = 78.6–98.3%) for central events and 1833 of 2458 = 74.6% (95% CI = 72.81–76.25%) for obstructive events. The overall agreement between the two systems was (1833 + 5090 of 8636) = 80%.

Evaluation of sleep/wake epochs and total sleep time

Table 5 shows an epoch-by-epoch comparison of the ability of the SAA system and of the gold-standard system to determine whether the subject was asleep or awake. Agreement between the findings obtained by the two systems was (28 648 + 17 444) of 61 372 = 75%.


The main finding of this study was that the SAA detect SDB reliably in both and out of hospital settings.

Conventional testing by PSG requires a complex array of equipment and the presence of a team of personnel. This often makes its use impractical, especially for patients suffering from severe conditions such as acute or chronic heart failure. To overcome these difficulties, research in recent years has focused on examining the feasibility of using portable or subset PSG systems to test SDB (Young et al., 1997). In most of these studies the devices have been tested in isolated laboratory settings concomitantly with full PSG monitoring, while in some the patients have used the devices at home. The portable devices are studied by one of two methods. In the first, the saturation signal is tested in conjunction with one or two respiratory signals (Abraham et al., 2006; Amir et al., 2010a–d; De Chazal et al., 2004; Larsen et al., 1984; Levy et al., 1996; Stoohs and Guilleminault, 1992). These systems are relatively easy to validate, because the respiratory events can be detected and classified clearly. In the second method (which is similar to our mode of analysis), the saturation signal is tested in conjuction with PPG, ECG, accelerometry or snoring (Amir et al., 2010d; Brouillette et al., 1987; Hajduk et al., 2000; Heneghan et al., 2008; Kiely et al., 1996; Mayer et al., 1998; Nijima et al., 2007; Raymond et al., 2003; Whitelaw et al., 2005). The additional signal monitors a range of movements, allowing total sleep time and wake episodes to be identified. In most of the studies referred to above, SDB parameters show good correlation when measured by portable recordings and by full PSG.

In the present study, the SAA system was tested in two different settings, ambulatory and in hospital. Under these conditions, and although the tested population comprised patients with severe cardiovascular disease, including heart failure, the data yielded by SAA were as reliable as those obtained by full PSG. Moreover, in addition to results of the conventional SDB parameters discussed above, our system was able to classify the respiratory events as well as to differentiate between sleep and wake. Validation of this classification against gold-standard PSG showed that it was accurate and reliable. The new sleep and wake algorithm used for the SAA enables the system to detect AHI without correction for an estimated sleep time. The sleep and wake comparison between the system to the gold standard yielded 75% agreement. Taking into consideration the low interscorer variability in sleep and wake episodes and the AHI diagnosis parameter, which takes the sleep period into consideration, the sleep and wake episodes give a reliable result.

Because SDB is known to be associated significantly with cardiovascular morbidity and mortality, patients admitted with a cardiovascular condition are likely to be more prone to SDB than healthy individuals. Paradoxically, these patients may not be suitable candidates for evaluation by formal testing in a sleep laboratory because of their poor health status. Thus, they could benefit from having a reliable sleep study carried out via an automated system in a more convenient way. For a large proportion of this sick population, this might be the only suitable means of undergoing any sleep study assessment.

In summary, our results, by demonstrating that the SAA can conveniently provide an acceptable analysis of sleep-related and/or cardiac-related breathing disorders, suggests that it may serve as a future platform for screening and diagnosis of SDB in patients with significant cardiovascular disease.

Declarations of Interest

Offer Amir is an advisor of WideMed Ltd; Deganit Barak-Shinar is an employee of WideMed Ltd.


The research was sponsored by WideMed Ltd, and was conducted at Morristown Memorial Hospital, NJ, USA.