Beyond the AHI–pulse wave analysis during sleep for recognition of cardiovascular risk in sleep apnea patients

Recent evidence supports the use of pulse wave analysis during sleep for assessing functional aspects of the cardiovascular system. The current study compared the influence of pulse wave and sleep study‐derived parameters on cardiovascular risk assessment. In a multi‐centric study design, 358 sleep apnea patients (age 55 ± 13 years, 64% male, body mass index 30 ± 6 kg m−2, apnea–hypopnea index 13 [5–26] events per hr) underwent a standard overnight sleep recording. A novel cardiac risk index was computed based on pulse wave signals derived from pulse oximetry, reflecting vascular stiffness, cardiac variability, vascular autonomic tone and nocturnal hypoxia. Cardiovascular risk was determined using the ESC/ESH cardiovascular risk matrix, and categorized to high/low added cardiovascular risk. Comparisons between cardiac risk index and sleep parameters were performed for cardiovascular risk prediction. Apnea–hypopnea index, oxygen desaturation index and cardiac risk index were associated with high cardiovascular risk after adjustment for confounders (p = .002, .001, < .001, respectively). In a nested reference model consisting of age, gender and body mass index, adding cardiac risk index but not apnea–hypopnea index or oxygen desaturation index significantly increased the area under the receiver operating characteristic curve (p = .012, .22 and .16, respectively). In a direct comparison of oxygen desaturation index and cardiac risk index, only the novel risk index had an independent effect on cardiovascular risk prediction (pCRI < .001, pODI = .71). These results emphasize the association between nocturnal pulse wave and overall cardiovascular risk determined by an established risk matrix. Thus, pulse wave analysis during sleep provides a powerful approach for cardiovascular risk assessment in addition to conventional sleep study parameters.


| INTRODUC TI ON
Obstructive sleep apnea (OSA) is characterized by repetitive upper airway obstruction, intermittent hypoxia, and frequent arousal from sleep. OSA is highly prevalent in the middle-aged population, and both cardiovascular (CV) function and sleep quality are negatively affected by the disorder (Benjafield et al., 2019). Episodic hypoxia and arousal from sleep activate the sympathetic nervous system and induce fundamental changes in autonomic and cardiovascular homeostasis during the course of an obstructive respiratory event (Hedner et al., 1988).
Polysomnography (PSG) is the gold-standard for OSA assessment (Berry et al., 2012). The PSG recording typically includes a pulse oximetry signal for quantification of hypoxia. Other relevant PSG metrics are the apnea-hypopnea index (AHI), oxygen desaturation index (ODI), distribution of sleep stages, and the frequency of arousals from sleep. Recent studies have challenged the value of the AHI as a universal expression for OSA severity and associated CV risk (Pevernagie et al., 2020;Randerath et al., 2018). Hence, there is a need for more precise markers for clinically relevant OSA and its associated CV risk.
Pulse wave analysis has been applied for the assessment of vascular properties like stiffness and aging (Laurent et al., 2016). For example, oscillometric assessment of the pulse wave contour has been used to prospectively evaluate peripheral vascular properties.
Central vascular stiffness was shown to be a strong predictor of CV outcomes in large epidemiological studies (Ben-Shlomo et al., 2014).
In line with these findings, we developed an oximeter-based pulse wave analysis algorithm that integrates quantification of multiple aspects including autonomic, vascular and cardiac activity during sleep (Grote et al., 2011). This method, which also includes traditional measures of hypoxia, was found to predict high CV risk as determined by established CV risk matrices (e.g. Framingham, SCORE, PROCAM) in an overnight recording of the finger pulse wave signal in patients with suspected OSA. Importantly, the association with CV risk prediction was considerably more precise with the overnight pulse wave analysis than with traditional daytime blood pressure assessment (Sommermeyer et al., 2014).
The current study aimed to compare the capacity of conventional sleep parameters, such as AHI and ODI, and the newly developed finger pulse wave analysis to identify individuals with high CV risk. Different methods were simultaneously applied during an overnight assessment in a large multi-centric sleep lab cohort. We hypothesized that pulse wave analysis was superior to measures of sleep quality and OSA severity for recognition of high CV risk.

| Study subjects and data collection
A study flow chart is presented in Figure 1. The study design has previously been reported in detail (Sommermeyer et al., 2014). In brief, the study comprised 520 subjects with suspected OSA investigated at one of five tertiary sleep centres in Germany and Sweden. Twenty-five recordings were excluded due to irregular pulse wave signals (e.g. atrial fibrillation). A random subset of recordings (n = 115) was used to enable the training of a novel cardiac risk index (CRI; Sommermeyer et al., 2016). The current analysis contained the remaining subjects in the cohort (n = 380). Incomplete recordings lacking any parameter were removed prior to the analysis (n = 22). All subjects underwent a standard full-night sleep study.
Clinical routines differed between centres, and a subgroup of 92 patients was studied with a polygraphic sleep test. The final analysis was, therefore, performed in two cohorts: the main analysis cohort (n = 358, assessing AHI, ODI and pulse wave parameters); and the PSG-subcohort (n = 266, assessing AHI, ODI, pulse wave parameters and additional sleep variables). Written informed consent was obtained from each participant prior to entry into the study, and the protocol was approved by the respective local Ethics Committee at each centre.

| Sleep study
The polygraphic (PG) montage included finger pulse oximetry, respiratory effort and nasal-oral pressure. Signals used for pulse wave analysis were collected via a modified pulse oximeter module (ChipOx, Corscience GmbH&Co.KG) integrated into a PG device (Somnolab, SomnCheck/R&K, or SomnCheck II, Weinmann Geraete fuer Medizin GmbH&Co.KG). Further details of the technical set-up have been described elsewhere .
The full night in-lab PSG was performed in accordance with standards set by the American Academy of Sleep Medicine (AASM ;Iber et al., 2007). The PSG recordings were based on a complete montage, which included recordings of electroencephalography (EEG), electrooculography (EOG), chin and left/right anterior tibialis electromyography (EMG), and electrocardiography (ECG). The derived sleep variables included total sleep time (TST), parameters of sleep architecture (expressed as percent time spent in each sleep stage of TST), and periodic leg movement index.
All recordings were manually scored by trained sleep technicians in accordance with the AASM criteria (Iber et al., 2007). For classification of hypopnea events, a ≥ 4% desaturation criterion (PG/PSG), or an arousal (only PSG) were used. The frequency of apnea/hypopnea and hypoxic events (≥ 4%) was calculated as the AHI and the ODI, respectively.

| Pulse wave analysis during sleep
The overnight photoplethysmographic recording included a single unfiltered pulse wave signal that was used for all subsequent analyses. The following eight pulse wave parameters were computed according to the methodology earlier described in detail .
• The pulse wave attenuation index (PWA-I) -an index describing amplitude attenuations of 10%-30% from baseline per hour.
• The pulse propagation time (PPT) -the mean time-interval between the peak systolic pulse wave and the dicrotic notch.
• The respiration-related pulse oscillation (RRPO) -pulse rate variability in the frequency range of typical physiological breathing rate (0.15-0.4 Hz).
• The pulse rate acceleration index (PR-I) -number of pulse rate accelerations, defined as > 10% increases from baseline, per hour.   • The difference between the PR-I and the SpO 2 -I -reflecting the chronotropic response to hypoxia.
Subsequently, the overall CRI was computed as a composite metric of these parameters . The computed CRI value varied between 0 (average CV risk) and 1 (increased CV risk).

| CV risk assessment according to conventional CV risk factor
The ESC/ESH risk matrix (Mancia et al., 2007) was used as a goldstandard to classify overall CV risk. This matrix is based on an F I G U R E 1 Study flow chart. Subjects with irregular pulse wave signals (n = 25) have been excluded prior to the study. A subset of subjects was used to train the cardiac risk index (n = 115) and, therefore, excluded from further validation. Final analysis was performed in two cohorts. Recordings in the polysomnography (PSG)-subcohort contained electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG) traces to assess sleep stages and periodic limb movements assessment of various risk factors, including demographics, anthropometrics, history of CV disease, blood pressure, smoking habits, glucose and lipid variables, and associated clinical conditions, to assign each subject to one out of five risk classes, ranging from "average cardiovascular risk" to "very high added cardiovascular risk". These classes are calibrated to indicate a 10-year added risk of CV disease of 0%, > 0%-15%, > 15%-20%, > 20%-30%, and > 30%, according to Framingham criteria (European Society of Hypertension-European Society of Cardiology Guidelines, 2003). We allocated each subject in one of two groups (low/high risk), containing those with average, low and moderate added risk, and those with high or very high added CV risk, respectively.

| Statistical analysis
The analysis was performed using R, version 3.5.3 (R Foundation for Statistical Computing). Descriptive data are presented in mean ± standard deviation or median with interquartile range, depending on the distribution pattern. Pearson's Correlation Coefficient was used to examine dependencies in the set of candidate parameters to identify possible confounding. For evaluating the predictive value of added parameters, we followed the guideline for evaluating novel risk markers, published by the American Heart Association (Hlatky et al., 2009). This is realized by the following steps.
• Reporting statistical relevance by assessing parameters' p-values in a single variable model.
• Evaluating the additional value of the tested parameter after statistical adjustment for established risk factors.
• Evaluation of the models' fit on the present dataset after adding the tested parameters, using the Akaike Information Criterion (AIC) together with a model comparison using the Likelihood Ratio Test.
• Evaluating the predictive value and capability of discrimination by analysing changes in the area under the ROC curve (receiver operating characteristic curve).
All PSG-variables were used as continuous variables, and data were tested for distribution patterns. For the between-group comparison of different study samples, we used Student's t-test for normally distributed parameters, and Mann-Whitney U-test for non-normally distributed data. Pearson's Chi-squared test was used for comparison of categorical variables. All following analyses have been performed using logistic regression models, predicting a binary outcome of low versus high added CV risk based on the gold-standard ESC/ESH risk matrix. Analyses have been performed to assess single parameter impact, as well as additional predictive value, when controlling for established risk factors. For this purpose, two different models have been fitted for each tested parameter. The first model contained a single predictor to assess its individual impact on CV risk. The second model controlled for age, gender and body mass index (BMI) to evaluate the additional predictive value. The AIC was used to assess the model's likelihood of the resulting classification (Akaike, 1973). To evaluate the significance of changes in the model's likelihood, a likelihood ratio (LR) test using chi-square statistics was performed. Area under the ROC curve (AUC) served as a marker for general model performance in terms of sensitivity and specificity. Differences in AUC between the different modelling approaches were tested using DeLong's nonparametric test (DeLong et al., 1988). To address limitations of the study design, we conducted a sensitivity analysis to determine the effect of different sleep study protocols (PG and PSG) and study sites.

| Group differences in anthropometric and sleep test-derived parameters
Subjects with elevated CV risk were older, had a higher BMI and differed significantly in terms of analysed pulse wave and PSG-derived parameters (Table 1).

| Correlation analysis of pulse wave and sleep test-derived parameters
In the set of PSG parameters, we observed a high inter-correlation between the AHI and ODI (r = .85, p < .001), as well as for sleep stages (N1 and N2, r = −.55; N1 and N3, r = −.44). Overall, pulse wave parameters showed weaker or no significant inter-correlation: the strongest correlation was found for the two pulse rate-related parameters RRPO and PR-I (r = .62, p < .001), followed by PPT and RRPO (r = .30, p < .001; Table 2).

| Logistic regression models
All tested pulse wave parameters, as well as the ODI and the AHI, were significantly associated with CV risk classification (Table 3).
However, time spent in sleep stages N2 and N3 was not associated with risk. When adjusting for age, gender and BMI, the associations with the risk classification were slightly weaker, while the time spent in specific sleep stages no longer influenced the outcome in these models.

| Akaike Information Criterion (AIC)
The differences in AIC, resulting from any of the parameters of interest added to a reference model, are reported for the parameters previously found significant (Table 4). In addition, the p-value of the corresponding LR test is presented. All approaches resulted in significant improvements in the model. The model containing the pulse wave-derived CRI had the highest likelihood of association.

| Area under the ROC curve
The AUC of the ROC curve for the model containing age, gender and BMI was 0.821. When CRI was added into the model, the AUC increased to 0.853 (p = .012). The various investigated sleep parameters or the single pulse wave parameters did not add any further improvement over the reference model. Figure 2 visualizes the improvements in AUC for a selection of parameters from pulse wavederived and conventional sleep-derived parameters.

| Nested analysis for the direct comparison of pulse wave and PSG-derived parameters
In the final evaluation, we fitted a model containing the strongest parameters from both methodologies, the novel risk index CRI from pulse wave and the ODI from conventional PSG. In order to control for a possible relation between the predictors, an interaction term was introduced in the model. However, while the CRI was highly significant (p = .001) in this model, neither the single variable influence of the ODI nor the joint interaction term were significant (p = .708 and p = .776, respectively). Figure 3 illustrates that the CRI further improved the model quality when the ODI was already entered (AIC change 16.5, p < .001), while the ODI TA B L E 1 Descriptive data of the study cohort: data presented as median [IQR]  Asterisks indicating parameters' significance (0 '***' 0.001 '**' 0.01 '*' 0.05 '•' 0.1 ' '1). a Main-cohort (n = 358, with n = 215 for low-risk group, n = 143 for high-risk group).
has no significant additive effect on top of the CRI (AIC change −2.0, p = .356).

| Sensitivity analysis
We performed sensitivity analyses to address factors that may have influenced the sleep scoring results. A binary variable to define the type of sleep study (PG or PSG) and a categorical variable that separated the contribution of sleep centres were added to the model.
The results confirmed that the AHI was slightly underestimated in PG recordings, but this difference did not impact on the outcome.
Sleep centres showed a minor difference on average severity of CV risk, but no significant difference on the effect of sleep parameters.
The CRI was independent to these influences in all tested models.

| DISCUSS ION
The main findings in this cohort of sleep apnea patients are the All pulse wave-derived parameters showed significant differences between low and high CV risk groups, and contributed significantly to single variable prediction models, confirming their independent relevance for CV risk prediction. Because the inter-correlations between those parameters were limited, combined information from these parameters, defined as the CRI, appeared to be more useful.
Our data support other reports suggesting a strong association between hypoxic patterns determined by an advanced analysis of signals recorded in patients with OSA (Terrill, 2020 Pulse wave parameters assessed during daytime rest, such as pulse wave velocity or vascular stiffness determined by peripheral arterial tonometry, have been extensively used to assess CV risk in prospective studies and in various clinical protocols (Kim & Kim, 2019). For example, application of a photoplethysmographic signal to assess blood pressure was described in several studies, highlighting the link between hypertension and pulse wave aberrations (Elgendi et al., 2019). Vascular stiffness measured by applanation tonometry (Arteriograph) or oscillometric pulse wave velocity (Spygmocor) were validated against gold-standard methodology in prospective, large-scale epidemiological studies (Townsend et al., 2015). However, these assessments reflect short time windows in awake patients. We argued that photoplethysmography during the sleep period may enable a unique, undisturbed and extended period to assess vascular reactivity. During nighttime, overall sympathetic activity is decreased and, compared with awake state during daytime, the influence of non-standardized stressors is likely to be reduced by sleep. Fluctuations of the pulse wave amplitude and heart rate can be observed, suggesting an opportunity to capture autonomic nervous and CV function not seen during daytime. Indeed, the individual pulse wave parameters have been extensively validated in our previous studies, showing their potential to reflect independent properties of the vascular system (Grote & Zou, 2017). The analysis of nocturnal PWA may also be applied to predict daytime blood pressure and hypertension (Zou et al., 2009), or the analysis of PPT changes during sleep has been shown to reflect functional vascular properties (Svedmyr et al., 2016). The methodology also enables assessment of CV function during various stages of sleep, as well as identifying vascular stiffness during rapid eye movement sleep in patients with chronic obstructive pulmonary disease . Other applications include effects on nasal positive airway pressure therapy in OSA patients . Finally, the HypnoLaus study, a population-based cohort study including ambulatory PSG assessment of a representative Swiss adult population (Hirotsu et al., 2020), recently demonstrated that pulse wave attenuation frequency, duration and slope were independent predictors of prevalent hypertension in this population whereas conventional PSG variables were not. In addition, pulse rate accelerations related to apneic events predicted all cause and cardiovascular mortality in two large population-based cohorts (Azarbarzin et al., 2021).
Despite the different methodologies used, results from these TA B L E 3 Logistic regression models for the prediction of high CV risk class-each line shows coefficient and significance level for a single predictor in an unadjusted or adjusted model AHI, apnea-hypopnea index; BMI, body mass index; CRI, cardiac risk index; ODI, oxygen desaturation index; PPT, pulse propagation time; PR-I, pulse rate variability; PWA-I, pulse wave attenuation index; REM, rapid eye movement; RRPO, respiratory-related pulse oscillations; SpO 2 -I, hypoxia index; T < 90, time below 90% SpO 2 ; TSD, time in symmetric desaturation; TST, total sleep time.
prospective population-based studies strongly support our approach and suggest that incorporating multiple different features of a vascular signal may be a useful tool for CV risk classification.  (Mancia et al., 2007). In line with the recent update of the ESC/ESH recommendations for CV risk classifications (Piepoli et al., 2016), the pulse wave-derived CRI showed a strong association with other risk classifier systems like the Framingham Score or the SCORE system (Sommermeyer et al., 2014). Using risk scores to evaluate the predictive value of the CRI in cross-sectional study design, rather than using prospective outcome data, is a major lim-  (Zapater et al., 2020). Indeed, we showed previously that the pulse wave-derived CRI was significantly higher in OSA patients with a previous CV event (Sommermeyer et al., 2014). However, our study was underpowered to model a prediction for the entire CV risk spectrum, but we identified a clear trend of stepwise increasing AHI/ODI and pulse wave parameters from low to intermediate to high CV risk classes (data not shown). Further studies using advanced pulse wave analysis in population-based studies will be needed to clarify this issue.

| Clinical relevance
There is an evident potential for application of an advanced pulse wave analysis in clinical sleep recordings. According to current standards, oximetry devices by default only record oxygen saturation or pulse rate, but we advocate plethysmographic pulse wave signals to be included in sleep recordings to take full advantage of the potential of such signals. There is a strong interest in simplified and cost-efficient methods for home sleep testing, especially in the analysis of parameters derived from finger pulse wave signal (Zou et al., 2006). For the task of CV risk assessment, our results suggest that PSG-specific parameters may be redundant if the pulse wave signal is analysed for the purpose of CV risk assessment.

| CON CLUS IONS
A pulse wave-derived CRI from sleep recordings was validated in patients with suspected OSA. Novel parameters derived from nocturnal pulse wave analysis provided a better approach for CV risk assessment compared with conventional parameters from standard sleep studies.
In addition, pulse wave signals contained information on top of that provided by the most prominent CV risk factors like age, BMI and gen- der. An oximeter-based photoplethysmographic signal is available in almost all instruments for diagnosis of sleep-disordered breathing. The use of an advanced pulse wave analysis for estimation of CV risk may add important information to the routine sleep diagnostic procedure.
Prospective evaluations of these novel biomarkers are warranted.

ACK N OWLED G EM ENTS
The authors would like to express their gratitude for the support throughout the study to Jeanette Norum (Sahlgrenska) and Lena Engelmark (Sahlgrenska) for their technical and intellectual input during different parts of the study. The study was sup-