Novel Sleep Phenotypic Profiles Associated With Incident Atrial Fibrillation in a Large Clinical Cohort

Abstract

BACKGROUND

While sleep disorders are implicated in atrial fibrillation (AF), the interplay of physiologic alterations and symptoms remains unclear. Sleep-based phenotypes can account for this complexity and translate to actionable approaches to identify at-risk patients and therapeutic interventions.

OBJECTIVES

This study hypothesized discrete phenotypes of symptoms and polysomnography (PSG)-based data differ in relation to incident AF.

METHODS

Data from the STARLIT (sleep Signals, Testing, And Reports LInked to patient Traits) registry on Cleveland Clinic patients (≥18 years of age) who underwent PSG from November 27, 2004, to December 30,2015, were retrospectively examined. Phenotypes were identified using latent class analysis of symptoms and PSG-based measures of sleep-disordered breathing and sleep architecture. Phenotypes were included as the primary predictor in a multivariable-adjusted Cox proportional hazard models for incident AF.

RESULTS

In our cohort (N = 43,433, age 51.8 ± 14.5 years, 51.9% male, 74.9% White), 7.3% (n = 3,166) had baseline AF. Over a 7.6- ± 3.4-year follow-up period, 8.9% (n = 3,595) developed incident AF. Five phenotypes were identified. The hypoxia subtype (n = 3,245) had 48% increased incident AF (HR: 1.48; 95% CI: 1.34–1.64), the apneas + arousals subtype (n = 4,592) had 22% increased incident AF (HR: 1.22; 95% CI: 1.10–1.35), and the short sleep + nonrapid eye movement subtype (n = 6,126) had 11% increased incident AF (HR: 1.11; 95% CI: 1.01–1.22) compared with long sleep + rapid eye movement (n = 26,809), the reference group. The hypopneas subtype (n = 2,661) did not differ from reference (HR: 0.89; 95% CI: 0.77–1.03).

CONCLUSIONS

Consistent with prior evidence supporting hypoxia as an AF driver and cardiac risk of the sleepy phenotype, this constellation of symptoms and physiologic alterations illustrates vulnerability for AF development, providing potential value in enhancing our understanding of integrated sleep-specific symptoms and physiologic risk of atrial arrhythmogenesis.

CENTRAL ILLUSTRATION

Filtered Relevant Variables by Cluster Followed by Outcome Model

This depicts a list of all variables filtered via latent class analysis, the 5 clusters described by the most relevant and differentiating variables, and last, a person depicting each sleep profile and their association with atrial fibrillation (AF) development. AHI = apnea hypopnea index; ESS = Epworth Sleepiness Scale; EtCO₂ = end-tidal carbon dioxide; NREM = nonrapid eye movement sleep; PLM = periodic limb movement; PSG = polysomnography; REM = rapid eye movement; SaO₂= oxygen saturation; SpO₂ = peripheral oxygen saturation.

Atrial fibrillation (AF) is a serious cardiac arrhythmia that affects over 5.2 million individuals in the United States,¹ adding a significant burden on health care systems,² with costs of $6.7 billion annually,³ and is associated with ischemic stroke⁴ and increased mortality.⁵ Identifying modifiable risk factors specific to sleep disruption that predict AF development is therefore crucial for early detection, risk stratification, and targeted interventions.

Sleep-disordered breathing is a driver of AF, with severe sleep-disordered breathing conferring a 4-fold increased AF risk compared with without.⁶ It encompasses obstructive sleep apnea (OSA), which may afflict up to 1 billion people globally,⁷ and is unrecognized in 85% of AF patients.⁸ Mechanistic factors include inflammation,^9,10 intermittent hypoxia leading to oxidative stress,^11–14 intrathoracic pressure alterations,¹⁵ and atrial remodeling,^16,17 with subsequent autonomic fluctuations. Despite this, randomized controlled trials have not yet identified risk mitigation with treatment, potentially due to older-aged cohorts, poor treatment adherence, or exclusion of sleepy patients and those with greater hypoxia.^18–20 Moreover, inclusion criteria to define OSA have utilized the apnea hypopnea index (AHI), a simple frequency measure.^18–20 To better capture the salient clinical and pathophysiologic features of sleep-disordered breathing, cluster analyses have focused on clinical outcomes such as cardiovascular disease,^21,22 mortality,^23,24 and treatment responsiveness, such as impact of²⁵ and adherence to²⁶ positive airway pressure; however, no cluster analyses have specifically examined symptoms and polysomnography (PSG) features in a longitudinal study of AF.

We therefore sought to identify patient subtypes based on sleep symptoms and PSG measures in a large clinical cohort and identify which subtypes confer increased association with longitudinal AF. Characterizing distinct sleep profiles that differ in incident AF can provide valuable insights into the role of sleep disturbance in AF pathogenesis, contribute to precision medicine via personalized risk stratification, and identify high-risk subgroups to inform targeted interventions for AF prevention.

METHODS

In this retrospective cohort study, we identified patient subtypes based on 23 sleep symptoms, Epworth Sleepiness Scale score, and 24 PSG measures. We utilized the STAR-LIT (sleep Signals, Testing, And Reports LInked to patient Traits) registry of patients who have undergone sleep studies at the Cleveland Clinic. The objective was to identify and then compare patient subtypes in terms of incident AF. The STARLIT registry and this study were granted waivers of informed consent by the Cleveland Clinic Institutional Review Board (19–864 and 22–1195, respectively).

STUDY POPULATION.

Patients ≥18 years of age who underwent in-laboratory PSG from November 27, 2004, to December 30, 2015, were examined. Exclusion criteria were <3 hours of recording time to ensure data quality;²⁷ home sleep apnea testing, as this approach does not measure sleep architecture; missing AHI values; and lack of follow-up. To restrict to incident AF, those with history of AF before PSG based on electronic health record (Epic) review of diagnosis codes were also excluded (Figure 1).

FIGURE 1. Flow Diagram of Cohort Selection.

Beginning with 46,223 patients who had sleep studies in the time frame of interest, 881 were excluded for recording time <3 hours, 1,629 were excluded for having home sleep apnea tests, and 280 were excluded for missing data on the apnea hypopnea index (AHI), leaving 43,433 patients in the analytic cohort for the cluster analysis. Then, 53 patients were excluded for lack of follow-up after the sleep study and 3,166 were excluded for having baseline atrial fibrillation, leaving 40,214 patients in the analytic cohort for the outcome model. Of these, 3,595 patients developed incident atrial fibrillation.

SLEEP TESTING.

Full-night diagnostic PSGs and split-night PSGs, which have a diagnostic and therapeutic portion with positive airway pressure, were included. These consisted of overnight recording of physiologic signals included on standard PSG with scoring based on the American Academy of Sleep Medicine Manual for the Scoring of Sleep and Associated Events²⁸ and generation of a standardized report by a physician that was subsequently deposited in Epic. PSG data collection, staging, scoring, and variable descriptions are detailed in the Supplemental Methods.

SLEEP SYMPTOMS.

Sleep symptom data were collected by a questionnaire at the sleep laboratory and expressed as binary variables (see the Supplemental Methods). Immediately preceding their PSG, patients completed the Epworth Sleepiness Scale questionnaire by indicating propensity to doze on a scale of 0 to 3 in 8 situations, which are summed for a total score of 0 to 24.²⁹ Higher scores indicate more severe daytime sleepiness.

DATA COLLECTION AND OUTCOME ASCERTAINMENT.

Comorbidities included hypertension, diabetes, dyslipidemia, ischemic stroke, chronic obstructive pulmonary disease, coronary artery disease, myocardial infarction, coronary artery bypass grafting history, and heart failure identified by International Classification of Diseases–9th and –10th Revision codes (Supplemental Table 1) and procedure codes (Supplemental Table 2). Antiarrhythmic medication use was collected (Supplemental Table 3).

AF was identified by International Classification of Diseases codes 427.31, 427.32, and I48 and Current Procedural Terminology codes 93650, 93653, 93656, and 93657. Internal validation of this approach yielded a positive predictive value of 90% ie, of all patients with these diagnosis codes, 90% had AF confirmed by 12-lead electrocardiogram, sleep study electrocardiogram, or echocardiogram. Of the remaining 10% that could not be confirmed by these modalities, 80% had AF confirmed by testing external to the Cleveland Clinic, with data not available for our use (see Supplemental Results).

SQL (Structured Query Language) queries were used to extract patient-level information from PSG reports and Epic encounter notes. A randomly generated subset (n = 100) was manually assessed by a content expert (C.M.H.) to identify and correct language processing errors. Missing data were handled using full-information maximum likelihood (cluster analysis) and multiple imputation by chained equations (outcome model) (see the Supplemental Results).³⁰ Internal validation utilized bootstrap techniques.

STATISTICAL ANALYSES.

Latent class analysis (LCA) was used to identify patient clusters by estimating model-based probabilities of observing each response pattern, assuming that the observed variables are influenced by cluster membership. All variables were initially included (see the Supplemental Methods). Those with high content overlap or high correlations (r > 0.8) were reviewed by content experts (C.M.H., R.M.); the most clinically relevant were retained. A single-class model was built; successive models were built increasing the number of classes by 1. The optimal number of classes was determined by iteratively comparing the model with k classes with the model with (k - 1) classes using multiple fit indices (see the Supplemental Results and Supplemental Table 4). Variable discrimination across classes was inspected visually in line graphs. For variables with little discrimination, one was removed at a time until a parsimonious group remained. Patients were grouped into their class using estimated posterior membership probabilities, with thresholds >0.80 indicating high accuracy. LCA was conducted using MPlus version 8.4.

Descriptive statistics were computed for the entire cohort, stratified by cluster, and expressed as mean ± SD for continuous variables and count and percentage for categorical variables. Across-cluster differences were compared by analysis of variance for continuous variables and Pearson’s chi-square test for categorical variables.

Cox proportional hazards models constructed with cluster as the independent variable and time from PSG to AF as the dependent variable were adjusted for age, sex, race, body mass index (BMI), tobacco use, hypertension, diabetes, dyslipidemia, coronary artery disease, heart failure, myocardial infarction, coronary artery bypass grafting, chronic obstructive pulmonary disease, antiarrhythmic medication, and prescription for positive airway pressure, which was time-varying. HRs with 95% CI were computed to estimate association between cluster and AF. Plots of Schoenfeld residuals tested the proportional hazards assumption to infer that the HR remains proportional across follow-up. Computations were performed using R version 4.0.3 (R Foundation for Statistical Computing). Statistical significance was determined by 2-sided tests and P values below 0.05.

A sensitivity analysis of incident AF excluding split studies was conducted to determine if results from full-night diagnostic PSGs aligned with overall findings. Last, a stratified analysis comparing the recommended (3% desaturation or arousal) and accepted (4% desaturation) hypopnea definitions was conducted to determine if hypopnea definition had an impact on overall results.

RESULTS

There were 43,433 adult patients (age 52 ± 15 years, 48% female, 75% White) included in the cluster analysis (Table 1). After excluding 3,166 (7.3%) patients for baseline AF and 53 (0.1%) for lack of followup, 40,214 patients remained for the outcome model (Figure 1). The follow-up period was 7.6 ± 3.4 years through January 17, 2022.

TABLE 1.

Cohort Description Stratified by Cluster

	Full Cohort (N = 43,433)	Hypoxia (n = 3,245)	Apneas + Arousals (n = 4,592)	Short Sleep + NREM (n = 6,126)	Long Sleep + REM (n = 26,809)	Hypopneas (n = 2,661)	P Value
Age, y	51.8 ± 14.5	56.7 ± 13.6	54.6 ± 13.4	55.0 ± 15.1	49.9 ± 14.4	51.9 ± 13.1	<0.001a
Sexb							<0.001c
Female	20,882 (48.1)	1,357 (41.8)	1,094 (23.8)	3,455 (56.4)	13,861 (51.7)	1,115 (41.9)
Male	22,548 (51.9)	1,888 (58.2)	3,498 (76.2)	2,671 (43.6)	12,946 (48.3)	1,545 (58.1)
Raceb							<0.001c
Black	8,527 (19.8)	585 (18.2)	852 (18.7)	868 (14.2)	5,781 (21.7)	441 (16.7)
White	32,313 (74.9)	2,516 (78.1)	3,503 (76.8)	4,949 (81.2)	19,268 (72.4)	2,077 (78.5)
Other	2,284 (5.3)	119 (3.7)	206 (4.5)	279 (4.6)	1,552 (5.8)	128 (4.8)
Body mass index, kg/m2,b	35.0 ± 10.7	40.1 ± 11.5	38.6 ± 11.5	34.9 ± 11.1	33.5 ± 10.1	37.5 ± 10.1	<0.001a
Tobacco useb							<0.001c
Current	5,126 (12.8)	3,128 (12.6)	631 (11.0)	522 (17.5)	292 (12.1)	553 (13.1)
Former	14,001 (34.9)	8,056 (32.6)	2,172 (37.9)	1,245 (41.8)	833 (34.4)	1,695 (40.2)
Never	20,618 (51.4)	13,339 (53.9)	2,900 (50.6)	1,180 (39.6)	1,270 (52.4)	1,929 (45.8)
Passive	332 (0.83)	207 (0.84)	31 (0.54)	30 (1.01)	28 (1.2)	36 (0.85)
Witnessed apneas	16,368 (37.7)	1,419 (43.7)	2,553 (55.6)	1,924 (31.4)	9,262 (34.5)	1,210 (45.5)	<0.001c
Epworth Sleepiness Scale scoreb	9.2 ± 5.5	9.8 ± 5.9	9.8 ± 5.6	8.2 ± 5.3	9.2 ± 5.5	9.7 ± 5.6	<0.001a
Apnea hypopnea index	30.0 ± 30.5	55.0 ± 37.7	86.5 ± 23.3	22.1 ± 19.0	17.2 ± 15.9	49.1 ± 27.3	<0.001a
Obstructive apneasb	16.3 ± 32.7	31.2 ± 50.9	44.5 ± 54.8	10.2 ± 20.6	9.1 ± 17.6	19.0 ± 29.5	<0.001a
Hypopneasb	20.0 ± 42.0	16.2 ± 32.5	7.6 ± 18.1	12.2 ± 22.3	11.2 ± 21.3	148.6 ± 54.4	<0.001a
Total sleep time, minb	327.8 ± 80.2	313.0 ± 85.3	294.0 ± 88.1	278.7 ± 88.9	346.4 ± 69.4	330.5 ± 69.1	<0.001a
Minimum SaO2, %b	83.2 ± 8.0	72.2 ± 10.0	78.4 ± 8.2	85.6 ± 5.8	85.0 ± 6.7	81.3 ± 7.3	<0.001a
Mean SaO2, %b	93.1 ± 3.0	87.4 ± 3.8	92.0 ± 2.1	93.6 ± 3.0	93.9 ± 2.1	92.7 ± 2.1	<0.001a
Sleep time SaO2 <90%, %b	10.8 ± 19.3	66.2 ± 16.2	19.7 ± 14.8	4.8 ± 8.4	3.8 ± 6.9	12.6 ± 16.2	<0.001a
Sleep efficiency, %b	75.2 ± 15.9	72.1 ± 17.2	68.1 ± 18.4	64.0 ± 18.2	79.3 ± 12.9	76.0 ± 13.7	<0.001a
Wake after sleep onset, minb	75.4 ± 55.1	87.8 ± 57.8	103.7 ± 61.3	110.8 ± 66.0	60.7 ± 44.2	77.5 ± 51.0	<0.001a
REM latency, minb	145.5 ± 90.1	151.7 ± 92.9	171.4 ± 89.6	294.3 ± 62.6	109.9 ± 55.7	162.4 ± 93.3	<0.001a
Percent REM, %b	16.9 ± 8.8	17.7 ± 10.8	16.9 ± 10.1	7.9 ± 6.1	19.0 ± 7.4	16.1 ± 8.5	<0.001a
Arousal indexb	34.3 ± 22.8	44.0 ± 26.6	70.8 ± 26.8	37.2 ± 18.9	25.1 ± 13.1	43.6 ± 21.7	<0.001a
Hypertension	23,286 (53.6)	2,249 (69.3)	3,045 (66.3)	3,633 (59.3)	12,826 (47.8)	1,533 (57.6)	<0.001c
Diabetes	9,184 (21.1)	1,081 (33.3)	1,288 (28.0)	1,503 (24.5)	4,707 (17.6)	605 (22.7)	<0.001c
Dyslipidemia	23,499 (54.1)	1,999 (61.6)	2,891 (63.0)	3,608 (58.9)	13,464 (50.2)	1,537 (57.8)	<0.001c
Coronary artery diseased	5,800 (13.4)	624 (19.2)	827 (18.0)	988 (16.1)	3,003 (11.2)	358 (13.5)	<0.001c
Ischemic stroke	1,774 (4.1)	161 (5.0)	200 (4.4)	313 (5.1)	994 (3.7)	106 (4.0)	<0.001c
Myocardial infarctiond	1,185 (2.7)	114 (3.5)	178 (3.9)	199 (3.2)	617 (2.3)	77 (2.9)	<0.001c
Coronary artery bypass graftingd	1,460 (3.4)	165 (5.1)	215 (4.7)	255 (4.2)	725 (2.7)	100 (3.8)	<0.001c
Chronic obstructive pulmonary disease	1,442 (3.3)	269 (8.3)	132 (2.9)	294 (4.8)	688 (2.6)	59 (2.2)	<0.001c
Heart failured	2,825 (6.5)	450 (13.9)	398 (8.7)	520 (8.5)	1,287 (4.8)	170 (6.4)	<0.001c
Study type							<0.001c
Full night diagnostic	32,0 64 (73.8)	1,696 (52.3)	1,779 (38.7)	4,804 (78.4)	21,831 (81.4)	1,954 (73.4)
Split night study	11,369 (26.2)	1,549 (47.7)	2,813 (61.3)	1,322 (21.6)	4,978 (18.6)	707 (26.6)

Values are mean ± SDorn (%). A heat map of relevant variables is included to visualize the differences across clusters (Supplemental Figure 1).

^aP value via analysis of variance.

^bData not available for all subjects.

^cP value via Pearson’s chi-square test.

^dComposite variable.

NREM = nonrapid eye movement; REM = rapid eye movement; SaO₂ = oxygen saturation.

SUBTYPE IDENTIFICATION.

The subset of clinically meaningful variables that best differentiated patient clusters was presence of witnessed apneas, Epworth Sleepiness Scale score, AHI, number of hypopneas, total sleep time, percentage of sleep time with oxygen saturation <90% (T90), rapid eye movement (REM) latency, REM percentage, and arousal index. Sleepiness was retained given its clinical relevance in incident cardiovascular disease²¹ and biologic plausibility based on animal experimental data.³¹ Five clusters were best fit (Table 1, Figure 2, Supplemental Table 4): 1) the long sleep + REM subtype (n = 26,809 [62%]) had the longest sleep time (346 ± 69 minutes), shortest REM latency (110 ± 56 minutes), and highest REM proportion (19% ± 7% of sleep); 2) the short sleep + nonrapid eye movement (NREM) subtype (n = 6,126 [14%]) had the shortest sleep time (279 ± 89 minutes), longest REM latency (294 ± 63 minutes), and lowest REM proportion (8% ± 6% of sleep), and thus the highest NREM proportion; 3) the hypopneas subtype (n = 2,661 [6%]) had the most hypopneas (149 ± 54 hypopneas); 4) the apneas + arousals subtype (n = 4,592 [11%]) had the highest prevalence of witnessed apneas (56% of patients), highest AHI (87 ± 23 events/h), fewest hypopneas (8% ± 18% hypopneas), and highest arousal index (71 ± 27 arousals/h); and 5) the hypoxia subtype (n = 3,245, 8%) had the highest T90 (66% ± 16%) and lowest minimum (72% ± 10%) and mean (87% ± 4%) oxygen saturation. The model classified patients well; average posterior probabilities were 0.94 ± 0.11, 0.84 ± 0.16, 0.94 ± 0.13, 0.90 ± 0.15, and 0.95 ± 0.12, for clusters 1 to 5, respectively.

FIGURE 2. Sleep Variables by Patient Clusters (N = 43,333).

Data are presented as standardized means. Error bars represent ±1 SD. A higher value on the y-axis indicates a worse value of the variable (ie, more witnessed apneas, higher Epworth Sleepiness Scale score, higher apnea hypopnea index, more hypopneas, shorter total sleep time, higher percentage of sleep time with oxygen saturation <90%, longer rapid eye movement (REM) latency, lower REM percentage, or higher arousal index). NREM = nonrapid eye movement sleep.

CROSS-SUBTYPE DIFFERENCES.

The subtypes differed by demographics (Table 1). Members of the hypoxia subtype were the oldest (57 ± 14 years) with the highest BMI (40 ± 12 kg/m²), while long sleep + REM members were the youngest (50 ± 14 years) with the lowest BMI (34 ± 10 kg/m²). The hypoxia, hypopneas, and apneas + arousals subtypes were mostly male (58%, 58%, and 76%, respectively), while the long sleep + REM and short sleep + NREM subtypes were mostly female (52% and 56%, respectively). Four clusters had mostly full-night diagnostic PSGs; apneas + arousals had mostly split-night PSGs (61.3%), as expected because elevated AHI is an indication to convert a diagnostic PSG to a split-night PSG. A heat map of relevant variables is included to visualize the differences across clusters (Supplemental Figure 1).

INCIDENT AF.

The patient subtypes differed in longitudinal association with incident AF in the fully adjusted model (omnibus P < 0.001) (Table 2, Figure 3). The long sleep + REM subtype was used as the reference. The hypoxia subtype had the strongest association with incident AF at 48% compared with reference (HR: 1.48; 95% CI: 1.34–1.64). The apneas + arousals subtype had 22% increased incident AF compared with reference (HR: 1.22; 95% CI: 1.10–1.35). The short sleep + NREM subtype had 11% (HR: 1.11; 95% CI: 1.01–1.22) increased incident AF compared with the reference. The hypopneas subtype did not differ from the reference in terms of AF association (HR: 0.89; 95% CI: 0.77–1.03). Internal validation provided an optimism-corrected concordance index of 0.781.

TABLE 2.

Cox Proportional Hazards Model of Incident Atrial Fibrillation (n = 40,214)

Cluster (vs Long Sleep + REM)	HR (95% CI)	P Value
Hypopneas	0.89 (0.77–1.03)	0.124
Short sleep + NREM	1.11 (1.01–1.22)	0.035
Apneas + arousals	1.22 (1.10–1.35)	<0.001
Hypoxia	1.48 (1.34–1.64)	<0.001

Omnibus P < 0.001. The follow-up period was 7.6 ± 3.4 years. The model was adjusted for age, sex, race, body mass index, tobacco use, hypertension, diabetes, dyslipidemia, coronary arterydisease, heart failure, myocardial infarction, coronary artery bypass grafting, chronic obstructive pulmonary disease, antiarrhythmic medication use, and positive airway pressure use, which was considered time- varying over the follow-up period.

Abbreviations as in Table 1.

FIGURE 3. Cumulative Incidence Plot From Sleep Study to Atrial Fibrillation (n = 40,214).

The incidence of atrial fibrillation within the follow-up period of mean 7.6 ± 3.4 years was 17.1% (n = 496 of 2,903 patients) for hypoxia, 12.9% (n = 540 of 4,173 patients) for apneas + arousals, 10.1% (n = 561 of 5,559 patients) for short sleep + NREM, 8.5% (n = 210 of 2,484 patients) for hypopneas, and 7.1% (n = 1,788 of 25,095 patients) for long sleep + REM. Omnibus P < 0.001. Abbreviations as in Figure 2.

The subtypes remained differentially associated with incident AF in the sensitivity analysis excluding split-night PSGs (omnibus P < 0.001) (Supplemental Table 5). In this subset (n = 30,082), the hypoxia subtype had 41% increased incident AF compared with reference (HR: 1.41; 95% CI: 1.23–1.61). The association of the apneas + arousals and short sleep + NREM subtypes with incident AF was no longer significant compared with reference (HR: 1.08; 95% CI: 0.93–1.25; and HR: 1.11; 95% CI: 1.00–1.24, respectively). The hypopneas subtype again did not differ from reference in association with incident AF (HR: 0.90; 95% CI: 0.76–1.07).

Results aligned with the full cohort in the subset of patients with hypopnea defined by 3% desaturation or arousal (n = 32,068; omnibus P < 0.001). However, in the subset of patients with hypopnea defined by 4% desaturation (n = 4,178), results were no longer statistically significant (omnibus P = 0.670), possibly due to decreased sample size (Supplemental Table 6).

DISCUSSION

We performed this large clinical cohort study of more than 43,000 patients to identify symptom-based and objective overnight sleep physiology–based patient subtypes associated with longitudinal AF over an average 8-year follow-up period. Symptoms, sleep-disordered breathing measures, and sleep architecture combined can discern 5 distinct patient subtypes, each characterized by a unique sleep profile differentially associated with incident AF (Central Illustration). Key findings include that: 1) the hypoxia phenotype was most strongly associated with AF, 48% increased from reference; 2) sleep architectural alterations (ie, REM characteristics and sleep duration) were relevant in AF association; and 3) apneas may be more strongly associated with AF than hypopneas. Findings were independent of obesity and age, which differed across the clusters. By illustrating that latent sleep-specific subtypes differ in AF development, findings carry high clinical relevance and provide novel insights into relationships of subjective and objective sleep and evolution of atrial arrhythmia over time.

APPROACH.

Cluster analysis provides an unsupervised machine learning approach to understand the interplay of symptoms and PSG measures including continuous and categorical variables. This is important in leveraging sleep as a risk factor due to inconsistent associations of these outcomes with physiologic alterations and symptoms independently. Prior work on sleep-based clusters has also used model-based approaches like LCA,^21,23,26,32 in addition to distance measures like k-means/modes^22,24 and hierarchical^22,25 clustering methods. We chose a model-based approach to utilize probabilities, which are more clinically applicable than distance measures,³³ to allow for nonspherical clusters, which is important for our high-dimensional clinic-based data,³⁴ and to use established model fit criteria for model comparisons.

CHRONIC INTERMITTENT HYPOXIA.

While AHI has historically been used to examine the association between sleep-disordered breathing and clinical outcomes, it does not account for mechanistic complexities (eg, intermittent hypoxia and autonomic fluctuations), varies night to night, and is dependent on scoring accuracy.³⁵ Therefore, alternative metrics have been applied, such as the hypoxia metrics in the current study: T90 and minimum and mean oxygen saturation. Novel metrics have also been put into practice more recently (eg, hypoxic burden),^13,36 which are out of the scope of this paper.

Chronic intermittent hypoxia has lasting effects on the cardiovascular system including endothelial dysfunction leading to hypertension and atherosclerotic disease, metabolic effects leading to insulin resistance and/or dysfunctional glucose processing, inflammation and oxidative stress leading to broader health implications, psychological and mood effects, and relevant to the hypoxia cluster with the highest Epworth Sleepiness Scale score (10 ± 6), neurocognitive deficits³⁷ such as excessive daytime sleepiness. Animal studies have shown that hypoxia/reoxygenation events during sleep can irreversibly damage wake-promoting neurons.³¹ Oxidative stress from sleep-disordered breathing and chronic intermittent hypoxia³⁸ may mediate this permanent neuronal loss that leads to persistent daytime sleepiness.^31,39 This could explain residual sleepiness after OSA treatment.

Chronic intermittent hypoxia has biologic plausibility in terms of direct influences on the atria. Animal models have elucidated this relationship; an experimental rat model used programmed electrical stimulation to alter atrial electrophysiology resulting in an increased atrial effective refractory period compared with normoxic control animals.⁴⁰ Another rat model showed chronic intermittent hypoxia caused a shortened atrial refractory period and atrial enlargement, fibrosis, upregulated inflammatory pathways, and thus increased AF inducibility.⁴¹ Epidemiologic studies have been less consistent. Two studies that defined sleep-related hypoxia with T90, as in the current study, did not observe an association of hypoxia and incident AF in older cohorts.⁴² However, in a clinical cohort, an association of hypoxia and AF was observed only in patients under 65 years of age.⁴³ Consistent with our findings, a larger clinical cohort study of patients with a median age of 47 years showed a strong association of hypoxia and AF.¹² Overall, these data suggest that clinical vs epidemiologic cohorts may have greater burden of hypoxia, enabling detection of significant associations with incident arrhythmia. Supporting this, Genuardi et al⁴⁴ found that sleep-related hypoxia may increase risk of venous thromboembolism in a clinical cohort, while the association of AHI and venous thromboembolism was confounded by obesity.

SLEEP ARCHITECTURE.

Frequency of arousals (ie, arousal index) has been inconsistent in epidemiologic studies despite mechanistic studies suggesting detrimental influences. Unlike hypoxia, this may be due to scoring agreement inconsistency between polysomnologists and institutions, or the multifactorial etiopathology of arousals. An experimental study (n = 20 healthy individuals) suggested that arousals cause cardiovascular consequences via endothelial dysfunction and increased arterial stiffness.⁴⁵ Cross-sectional analysis has observed that the arousal index is inversely associated with AF in the MESA (Multi-Ethnic Study of Atherosclerosis) cohort.⁴⁶ However, a longitudinal study (n = 6,000) with 11-year follow-up showed that low arousal threshold predicted mortality.⁴⁷

While sleep duration was not associated with AF in cross-sectional analysis of the MESA cohort,⁴⁶ larger longitudinal studies have shown that shorter sleep duration is associated with incident AF⁴⁸ and cardiovascular disease,⁴⁹ consistent with our findings. Lower REM proportion was associated with AF in a subset of the Cardiovascular Health Study⁵⁰—a relationship not observed in cross-sectional analysis of the MESA cohort.⁴⁶ In this context, our findings suggest that longer sleep duration with higher REM proportion may protect against AF development.

STRENGTHS.

Our work has several strengths and differs from prior sleep-related cluster analyses. We looked at an outcome unique from other sleep-based cluster analyses, incident AF, and we identified clusters based on both symptoms and PSG-based measures combined. We had a larger sample size and leveraged a clinic-based cohort using our established registry with efforts to verify and validate the data used. Our findings add to the growing evidence shifting clinical paradigms to risk-stratify patients by more than the AHI and consider more of the rich neuro-cardio-respiratory data provided by the PSG to better capture the physiologic processes associated with symptoms and outcomes. This is especially important to inform patient selection in future clinical trials, to accurately investigate the impact of OSA treatment on outcomes.

STUDY LIMITATIONS.

Our analysis is based on preexisting records and not on prospectively collected data. While it provides a valuable resource to efficiently investigate longitudinal relationships in big data, the design precludes inferences of causality. The Epworth Sleepiness Scale score differs across subtypes statistically; however, further work will need to determine the extent to which these are clinically meaningful. The exclusion of home sleep apnea tests (which do not measure indices of sleep architecture of interest in our analyses) may have introduced selection bias. Given the long time span of our data, there may be temporal variations in PSG scoring, data collection, and software; however, this would be expected to bias toward the null. Although the geographic network of our sleep laboratories covers a diverse socioeconomic spectrum with the majority of referral from community-based providers, this was a single-center study; therefore, we may only generalize our findings to patients undergoing sleep studies at a quaternary care center. We also recognize that, as our institution is a large referral center, patients may have followed up elsewhere. This would also be expected to result in underestimation of the magnitude of association due to random misclassification. Residual confounding is possible despite careful selection of covariates, including change in covariate data over time. Last, we adjusted for positive airway pressure prescription but did not have data on adherence.

FUTURE DIRECTIONS.

This study opens several avenues for future research. Although we conducted internal validation, which offers reassurance to consistency of the findings, replicating these results in independent cohorts would enhance generalizability. Exploring alternative outcomes such as mortality or perioperative AF or predictors such as hypoxic burden or circulating biomarkers of inflammation implicated in sleep-disordered breathing pathophysiology may provide further mechanistic insights, allowing for a more comprehensive understanding of the relationship between sleep and AF.

CONCLUSIONS

Analyzing a large clinical registry of sleep symptoms, validated sleepiness scores, and overnight sleep physiologic measures allowed us to identify distinct subtypes or clusters of patients that differ in association with AF, revealing unique sleep profiles that can be leveraged in the clinical setting. Consistent with recent literature mainly focused on cardiovascular disease, our findings unique to AF implicated hypoxia, and perhaps sleepiness, as strong factors associated with incident AF. Findings hold potential implications for risk stratification, prognostication, earlier identification of patients at increased risk of AF, and interventional targets focused on sleep disturbances.

PERSPECTIVES.

COMPETENCY IN MEDICAL KNOWLEDGE:

Sleep-related hypoxia is associated with development of AF.

COMPETENCY IN PATIENT CARE AND PROCEDURAL SKILLS:

Patients with sleep-related hypoxia should be made aware that they are vulnerable to developing AF and screening for sleep-disordered breathing can be considered.

COMPETENCY IN INTERPERSONAL AND COMMUNICATION SKILLS:

It is important to discuss sleep-related symptoms with patients, as this may help determine cardiac risk.

TRANSLATIONAL OUTLOOK 1:

Sleep-related hypoxia may increase atrial arrhythmogenesis via oxidative stress and inflammation.

TRANSLATIONAL OUTLOOK 2:

Our results of increased longitudinal AF in relation to the hypoxia subtype are consistent with the described neurocognitive deficits including excessive daytime sleepiness attributable to chronic intermittent hypoxia that has been shown to be irreversible in animal models.

ABBREVIATIONS AND ACRONYMS

AF

atrial fibrillation

AHI

apnea hypopnea index

BMI

body mass index

LCA

latent class analysis

NREM

nonrapid eye movement

OSA

obstructive sleep apnea

PSG

polysomnography

REM

rapid eye movement

T90

percentage of sleep time with oxygen saturation <90%