Abstract
Aims
Cardiovascular risk assessment in obstructive sleep apnoea (OSA) has traditionally relied on the Apnea–Hypopnea Index (AHI). However, AHI primarily reflects respiratory event burden and may not adequately capture inter-individual heterogeneity in sleep-related cardiorespiratory regulation. This study aimed to identify sleep stage-specific heart rate variability (HRV) phenotypes and evaluate their associations with long-term cardiovascular outcomes.
Methods and results
Data were obtained from the Sleep Heart Health Study (SHHS). After exclusion of participants with missing clinical covariates or outcome data, 4029 individuals were included. Linear and non-linear HRV features were extracted from stage-specific electrocardiographic segments during N2, N3, and rapid eye movement (REM) sleep. K-means clustering was performed independently for each sleep stage after Z-score standardization, and the optimal cluster number was determined using silhouette coefficients and the elbow method. Multivariable Cox proportional hazards models and Kaplan–Meier analyses were used to evaluate associations between HRV-informed phenotypes and all-cause mortality, coronary heart disease, and cardiovascular disease (CVD). Three HRV-informed phenotypes were identified within the selected clustering framework across N2, N3, and REM sleep stages: Stable (Cluster 0), Burden (Cluster 1), and High risk (Cluster 2). Compared with the Stable phenotype, the High-risk phenotype was consistently associated with increased cardiovascular risk, with the strongest association observed for incident CVD in the N2 stage [hazard ratio (HR) 1.468, 95% confidence interval (CI) 1.280–1.684, P < 0.001]. Similar associations were observed in the N3 (HR 1.426, 95% CI 1.249–1.629, P < 0.001) and REM (HR 1.394, 95% CI 1.216–1.598, P < 0.001) stages. Sequential adjustment for AHI, oxygen desaturation index, arousal index, and sleep efficiency resulted in only minimal attenuation of hazard estimates, indicating that the prognostic associations were largely independent of conventional respiratory severity measures. In contrast, the Burden phenotype was consistently associated with lower all-cause mortality across all three sleep stages (HR 0.83–0.85, all P < 0.001). Subgroup analyses stratified by OSA severity demonstrated broadly consistent prognostic associations across AHI strata.
Conclusion
Incorporating sleep stage-specific HRV features into conventional clinical phenotyping provides complementary prognostic information beyond traditional respiratory severity metrics. HRV-informed phenotyping may improve long-term cardiovascular risk stratification in patients with OSA.
Keywords: Obstructive sleep apnoea, Heart rate variability (HRV), Cardiovascular risk, Sleep stages, Risk stratification, Autonomic regulation
Graphical Abstract
Graphical Abstract.
Background
Sleep is not only a critical process for physiological recovery but also a vital window reflecting cardiovascular autonomic regulatory function. Substantial evidence has demonstrated that sleep architecture fragmentation is closely associated with adverse cardiovascular events, including hypertension, heart failure, atherosclerosis, and acute coronary syndrome.1–4 Among these, sleep-disordered breathing—characterized by recurrent intermittent hypoxia and disrupted sleep structure—persistently activates the sympathetic nervous system and suppresses parasympathetic tone, leaving the heart in a chronic state of stress.5–8 However, traditional assessment frameworks rely heavily on the Apnea–Hypopnea Index (AHI),9 a metric that solely quantifies the external pathological burden imposed on the heart. This approach fails to accurately capture the physiological dysregulation underlying neurocognitive, metabolic, and cardiovascular effects,10 thereby overlooking the intrinsic regulatory resilience of the heart when facing pathological stress. Given that individuals with identical AHI levels exhibit significant heterogeneity in hypoxic load and subsequent autonomic stress,11 this risk variance necessitates a paradigm shift from simple event counting towards a deep evaluation of intrinsic cardiac regulatory resilience. This study aims to identify collapse patterns within the cardiac autonomic nervous system through stage-specific heart rate variability (HRV), facilitating a transition from burden intensity quantification to functional regulatory assessment.
As a non-invasive indicator of cardiac autonomic function, HRV sensitively reflects the sympathovagal balance.12 Under physiological conditions, the evolution of HRV across sleep depths exhibits distinct stage-specific characteristics; these physiological fluctuations inherently contain vital information regarding physiological reserve.13–15 HRV is not a simple periodic oscillation but a complex non-linear system driven by the antagonism between sympathetic and parasympathetic nerves. Based on the loss of physiological complexity hypothesis,16 a healthy physiological system should possess high fractal characteristics and adaptive capacity. This theory posits that a decline in the flexibility of the cardiac regulatory system manifests not only as numerical alterations but also as ‘centralized rigidity’ and reduced entropy in signal dynamics. Such erosion of complexity weakens the cardiovascular system's defensive resilience against pressure perturbations during sleep. Non-linear algorithms, such as approximate entropy17–19 and multifractal analysis,20 can sensitively capture the drastic evolution of the autonomic nervous system during specific sleep stages. This non-linear dynamics-based feature extraction effectively compensates for the information loss inherent in traditional frequency-domain metrics when processing non-stationary physiological signals, providing electrophysiological features with greater pathophysiological significance for early cardiovascular warning. Studies have indicated that haemodynamic and sympathetic changes during rapid eye movement (REM) sleep may trigger ischaemic events in patients with vascular diseases.21 Recently, clustering analysis has been widely applied to identify latent phenotypes in complex diseases.22,23 However, within the cross-disciplinary field of sleep and cardiovascular medicine, several questions remain unresolved: whether HRV metrics are essential variables for achieving high-resolution stratification, and whether different sleep stages possess consistent representative capabilities in identifying cardiac decompensation signals under autonomic fluctuations driven by varying sleep depths.
Given that the N2 stage dominates total sleep duration (approximately 50%) and presents relatively stable parasympathetic tone,24 its features may reflect the baseline homeostatic regulation of the cardiac autonomic system. In contrast, the REM stage, accompanied by intense autonomic fluctuations, reflects regulatory elasticity under simulated stress conditions.21 Identifying clustering patterns with high physiological consistency and prognostic value across different stages would provide a robust scientific basis for precision cardiovascular risk stratification in sleep medicine.
Accordingly, this study utilizes the large-scale community-based SHHS cohort to systematically compare the clustering performance of traditional clinical metrics vs. the integration of HRV features across N2, N3, and REM stages. We aim to demonstrate that the incorporation of HRV not only optimizes phenotypic classification mathematically but also consistently identifies cardiac autonomic sub-phenotypes—overlooked by traditional evaluation systems—that carry a significantly high risk of mortality, thereby providing empirical evidence for precision risk assessment throughout the entire sleep cycle.
Methods
Study population
This study utilized cohort data from the Sleep Heart Health Study (SHHS).25,26 A total of 5804 participants with baseline polysomnography recordings were initially identified. Participant exclusion was conducted in two stages. First, individuals with missing baseline clinical covariates or follow-up outcome data were excluded. The excluded variables included age, sex, race, smoking status, body mass index (BMI), systolic and diastolic blood pressure, lipid profiles, medication history (beta-blockers, calcium channel blockers, Class 1B antiarrhythmics, digitalis preparations, vasodilators, and antihypertensive treatment), AHI, total sleep time, sleep efficiency, arousal index, and longitudinal outcome indicators. After exclusion of 1775 participants with incomplete data, 4029 participants remained eligible for further analyses. Second, stage-specific signal quality control was performed separately for N2, N3, and REM sleep analyses. Participants with excessive electrocardiogram (ECG) artefacts or insufficient stage duration preventing extraction of a valid 5-min ECG segment were excluded from the corresponding sleep-stage analysis. The final analytical samples included 3980 participants for N2 analyses, 3709 participants for N3 analyses, and 3785 participants for REM analyses (Figure 1). Median follow-up durations were generally comparable across HRV-informed phenotypes within each sleep stage, reducing the likelihood that survival differences were primarily attributable to unequal follow-up time. The overall median follow-up durations were 4234.5 [3895.8, 4498.0] days for N2 analyses, 4248.0 [3946.0, 4500.0] days for N3 analyses, and 4267.0 [3878.0, 4509.0] days for REM analyses (see Supplementary material online, Table S7).
Figure 1.
Flowchart of participant selection. (A) Participant screening process for N2 stage. (B) Participant screening process for N3 stage. (C) Participant screening process for rapid eye movement stage.
Signal processing and heart rate variability feature extraction
All physiological signal processing procedures were implemented in Python version 3.9. Sleep stage annotations for N1, N2, N3, and REM sleep were obtained from the SHHS dataset. ECG recordings were segmented according to sleep stage, and stage-specific analyses were performed independently for N2, N3, and REM sleep.
For each sleep stage, continuous ECG signals were first processed using the NeuroKit2 toolbox for R-peak detection and RR interval (RRI) series construction. Subsequently, the RRI series were segmented into non-overlapping 5-min windows using a sliding window approach implemented in Python. Prior to HRV feature extraction, automatic artefact correction was applied to the RRI series to reduce the influence of ectopic beats and signal noise. Specifically, abnormal RRIs were identified and corrected using built-in NeuroKit2 preprocessing routines based on physiological plausibility thresholds.
HRV features covered three complementary domains of autonomic regulation, including time domain, frequency domain, and non-linear indices. Feature selection was guided by established HRV analysis guidelines and previous physiological studies rather than data-driven optimization alone.18,27,28 To reduce feature redundancy before clustering, pairwise correlation analysis was performed and highly collinear variables were excluded.
To improve measurement stability, the final HRV profile for each participant within a given sleep stage was defined as the arithmetic mean of all valid 5-min window-derived HRV features from that stage.
Subtype identification and clustering model construction
To evaluate the incremental contribution of HRV features to participant stratification, we constructed two clustering frameworks. Model A included only baseline demographic and clinical variables, whereas Model B additionally incorporated stage-specific HRV features. All continuous variables were standardized using Z-score normalization prior to clustering analysis.
Participant subgroups were identified using the K-means clustering algorithm implemented in the Scikit-learn library. The optimal number of clusters was determined by jointly considering the silhouette score and the elbow method, with results shown in Figure 2. Clustering analyses were conducted independently for the N2, N3, and REM sleep stages.
Figure 2.
Determination of the optimal number of clusters using silhouette analysis and the elbow method across N2 (A), N3 (B), and rapid eye movement sleep stages (C).
To visualize the high-dimensional clustering structure, t-distributed stochastic neighbor embedding (t-SNE) was applied solely for visualization purposes. Importantly, all clustering assignments were derived from the original standardized feature space, and t-SNE projections were not used in the clustering procedure. The resulting visualizations are shown in Figure 3.
Figure 3.
t-distributed stochastic neighbor embedding visualization of the integrated clinical and heart rate variability feature space across N2 (A), N3 (B), and rapid eye movement (C) sleep stages.
Cross-stage stability and feature visualization
To assess the consistency of cluster assignments across sleep stages, a cross-stage stability matrix was constructed to quantify agreement in participant cluster labels among N2, N3, and REM sleep. Agreement was evaluated using pairwise cluster similarity metrics. In addition, t-SNE was used solely for visualization purposes to project the high-dimensional feature space into two dimensions.
Feature heatmaps based on Z-score-standardized variables were generated to illustrate the relative distributions of clinical, sleep-related, and HRV features across the identified clusters.
Clinical association analysis and statistical validation
Baseline characteristics were compared across clusters. Continuous variables were compared using one-way analysis of variance (ANOVA) or the Kruskal–Wallis test, as appropriate according to data distribution, whereas categorical variables were compared using the chi-square test.
Survival analyses were performed to evaluate the associations between HRV-informed phenotypes and long-term clinical outcomes. Kaplan–Meier survival curves were generated for all-cause mortality, any cardiovascular disease (CVD), and coronary heart disease (CHD), and differences between groups were compared using log-rank tests.
Multivariable Cox proportional hazards models were subsequently constructed to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for each outcome. Model 1 was adjusted for baseline clinical covariates, including age, sex, race, BMI, systolic blood pressure, lipid levels, smoking status, and hypertension medication use. Model 2 additionally adjusted for AHI and oxygen desaturation index (ODI). Model 3 further adjusted for arousal index and sleep efficiency.
All statistical analyses were performed using two-sided tests with a significance threshold of P <0.05.
Results
Limitations of traditional clinical phenotyping
Before incorporating HRV features, we first evaluated a clustering framework constructed exclusively from baseline demographic and clinical variables. Unsupervised clustering based on these conventional metrics identified two participant groups (Figure 4). However, after multivariable adjustment, the clinical-only clustering framework demonstrated limited independent prognostic value for cardiovascular outcomes.
Figure 4.
Two-dimensional t-distributed stochastic neighbor embedding visualization of K-means clustering based on only clinical baseline features.
Using the N2-stage analysis as an example, the clinical-only two-cluster model failed to significantly predict either CHD or incident CVD. The HR for CHD was 0.977 (95% CI: 0.800–1.192, P = 0.816), while the HR for incident CVD was 0.968 (95% CI: 0.813–1.153, P = 0.718) (Table 1). Although the model showed a statistically significant association with all-cause mortality (HR: 1.179, 95% CI: 1.043–1.334, P = 0.009), its discriminatory capacity for cardiovascular-specific outcomes remained limited.
Table 1.
Association of clinical phenotypes derived from conventional data with risks of mortality and cardiovascular outcomes in a multivariable cox proportional hazards analysis
| HR (95% CI) | P | |
|---|---|---|
| Any coronary heart disease | ||
| Clinical Cluster 1 vs. Cluster 0 | 0.977 (0.800–1.192) | 0.816 |
| Any cardiovascular disease | ||
| Clinical Cluster 1 vs. Cluster 0 | 0.968 (0.813–1.153) | 0.718 |
| All-cause mortality | ||
| Clinical Cluster 1 vs. Cluster 0 | 1.179 (1.043–1.334) | 0.009 |
Clinical phenotypes were derived from baseline demographic and clinical data only, without inclusion of cardiac autonomic features. Cluster 0 was used as the reference group (HR = 1.00). Cox models were adjusted for baseline demographic and clinical covariates including age, sex, race, body mass index, blood pressure, lipid profile, and smoking status.
CI, confidence interval, HR, hazard ratio.
After incorporation of stage-specific HRV features, a third high-risk phenotype emerged and demonstrated consistently elevated risks for both CHD and CVD across N2, N3, and REM sleep stages. For example, in the N2-stage analysis, the High-risk phenotype (Cluster 2) showed significantly increased risks for CHD (HR: 1.280, 95% CI: 1.126–1.456, P < 0.001) and CVD (HR: 1.424, 95% CI: 1.269–1.598, P < 0.001).
These findings suggest that conventional clinical clustering alone may provide limited additional prognostic information beyond established baseline risk factors, whereas incorporation of stage-specific HRV features enabled additional subgroup differentiation associated with long-term cardiovascular outcomes.
Distribution of clinical characteristics across heart rate variability-informed phenotypes
Clustering performance evaluation across sleep stages demonstrated that silhouette coefficients for the N2, N3, and REM analyses peaked at K = 3 (Figure 2), with corresponding values of 0.1672, 0.1660, and 0.1866, respectively. In parallel, the reduction in the sum of squared errors began to plateau, supporting the selection of a three-cluster solution for the present clustering framework.
After integration of stage-specific HRV features, participants were classified into three clusters within the selected clustering framework across N2, N3, and REM sleep stages: Cluster 0 (Stable), Cluster 1 (Burden), and Cluster 2 (High risk) (Figure 3). Baseline demographic, clinical, sleep-related, and representative HRV characteristics of the N2-stage clusters are summarized in Table 2. Corresponding results for N3 and REM sleep are presented in Supplementary material online, Tables S2 and S3.
Table 2.
Baseline clinical, sleep, and key HRV characteristics of HRV-derived clusters during N2 sleep
| Variables | Total (n = 3980) | Stable (Cluster 0) (n = 1409) | Burden (Cluster 1) (n = 1675) | High risk (Cluster 2) (n = 896) | Statistical value | P |
|---|---|---|---|---|---|---|
| Age, mean (SD) | 65.1 (10.6) | 62.9 (10.2) | 66.8 (10.5) | 65.2 (10.7) | F = 54.23 | <0.001 |
| Gender, n (%) | χ2 = 25.63 | <0.001 | ||||
| Male | 1885 (47.4) | 736 (52.2) | 722 (43.1) | 427 (47.7) | ||
| Female | 2095 (52.6) | 673 (47.8) | 953 (56.9) | 469 (52.3) | ||
| Race, n (%) | χ2 = 29.70 | <0.001 | ||||
| White | 3707 (93.1) | 1346 (95.5) | 1543 (92.1) | 818 (91.3) | ||
| American Indian or Alaska Native | 204 (5.1) | 38 (2.7) | 109 (6.5) | 57 (6.4) | ||
| Black or African American | 69 (1.7) | 25 (1.8) | 23 (1.4) | 21 (2.3) | ||
| BMI, mean (SD) | 28.3 (5.0) | 27.8 (4.8) | 28.1 (5.0) | 29.3 (5.3) | F = 25.10 | <0.001 |
| SBP, mean (SD) | 125.3 (18.4) | 123.3 (17.8) | 125.2 (18.6) | 128.5 (18.2) | F = 22.53 | <0.001 |
| DBP, mean (SD) | 71.4 (10.9) | 71.5 (10.4) | 70.5 (10.8) | 73.1 (11.7) | F = 17.50 | <0.001 |
| Drinking, n (%) | 3452 (86.7) | 1240 (88.0) | 1461 (87.2) | 751 (83.8) | χ2 = 8.96 | 0.011 |
| Smoking, n (%) | 377 (9.4) | 122 (8.5) | 160 (9.6) | 95 (10.3) | χ2 = 5.38 | 0.250 |
| Total cholesterol, mean (SD) | 206.8 (37.6) | 206.2 (36.9) | 205.1 (37.2) | 211.0 (39.2) | F = 9.09 | <0.001 |
| HDL-C, mean (SD) | 50.9 (15.9) | 51.5 (16.7) | 51.5 (15.3) | 48.8 (15.3) | F = 9.41 | <0.001 |
| Triglycerides, median [Q1, Q3] | 128.0 [91.0, 183.0] | 120.0 [85.0, 173.0] | 126.0 [90.0, 179.8] | 147.5 [103.0, 201.0] | H = 70.64 | <0.001 |
| AHI, median [Q1, Q3] | 13.9 [7.4, 24.5] | 12.4 [6.5, 21.8] | 14.2 [7.5, 25.3] | 16.3 [8.8, 28.5] | H = 56.80 | <0.001 |
| ODI 4%, median [Q1, Q3] | 5.6 [2.0, 13.7] | 4.6 [1.6, 11.1] | 6.0 [2.0, 14.2] | 7.2 [2.6, 16.2] | H = 55.82 | <0.001 |
| Sleep_Duration, mean (SD) | 362.4 (63.2) | 367.8 (61.3) | 360.9 (63.1) | 356.7 (65.5) | F = 8.24 | <0.001 |
| Sleep_Efficiency, mean (SD) | 82.8 (10.4) | 83.7 (9.9) | 82.5 (10.7) | 81.9 (10.7) | F = 9.44 | <0.001 |
| REM_Percent, median [Q1, Q3] | 20.1 [16.0, 24.0] | 20.7 [16.9, 24.4] | 19.9 [15.8, 23.9] | 19.5 [15.3, 23.4] | H = 29.15 | <0.001 |
| Arousal_Index, mean (SD) | 19.5 (10.7) | 18.6 (9.4) | 19.8 (11.5) | 20.4 (11.2) | F = 10.01 | <0.001 |
| Beta-blocker use, n (%) | 507 (12.7) | 175 (12.4) | 207 (12.4) | 125 (14.0) | χ2 = 1.53 | 0.465 |
| Clinical outcomes, n (%) | ||||||
| All-cause mortality | 957 (24.0) | 239 (17.0) | 460 (27.5) | 258 (28.8) | χ2 = 60.48 | <0.001 |
| Any coronary heart disease | 663 (16.7) | 195 (13.8) | 291 (17.4) | 177 (19.8) | χ2 = 14.87 | <0.001 |
| Any cardiovascular disease | 1002 (25.2) | 254 (18.0) | 470 (28.1) | 278 (31.0) | χ2 = 61.90 | <0.001 |
| Mean_HR_bpm, mean (SD) | 96.3 (21.9) | 115.8 (14.2) | 91.6 (16.5) | 74.5 (13.8) | F = 2218.95 | <0.001 |
| HRV_SDNN, median [Q1, Q3] | 166.0 [124.4, 208.6] | 153.6 [122.2, 185.2] | 204.9 [175.6, 241.9] | 103.4 [69.6, 136.7] | H = 1678.25 | <0.001 |
| HRV_RMSSD, median [Q1, Q3] | 243.5 [177.5, 314.1] | 267.0 [206.7, 337.0] | 280.3 [232.9, 333.5] | 126.1 [79.1, 166.7] | H = 1664.47 | <0.001 |
| HRV_pNN50, mean (SD) | 69.8 (27.4) | 92.2 (9.6) | 73.3 (13.9) | 28.2 (16.2) | F = 6688.92 | <0.001 |
| HRV_CVNN, mean (SD) | 0.3 (0.1) | 0.3 (0.1) | 0.3 (0.1) | 0.1 (0.1) | F = 2378.54 | <0.001 |
| HRV_LF, median [Q1, Q3] | 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] | H = 1633.73 | <0.001 |
| HRV_HF, median [Q1, Q3] | 0.0 [0.0, 0.1] | 0.0 [0.0, 0.0] | 0.1 [0.0, 0.1] | 0.0 [0.0, 0.1] | H = 2056.83 | <0.001 |
| HRV_LFHF, median [Q1, Q3] | 0.5 [0.4, 0.8] | 0.6 [0.5, 1.0] | 0.4 [0.4, 0.5] | 0.7 [0.5, 1.2] | H = 978.26 | <0.001 |
| HRV_TP, median [Q1, Q3] | 0.1 [0.0, 0.1] | 0.0 [0.0, 0.1] | 0.1 [0.1, 0.1] | 0.1 [0.1, 0.1] | H = 2033.38 | <0.001 |
| HRV_LFn, mean (SD) | 0.2 (0.1) | 0.3 (0.1) | 0.2 (0.0) | 0.3 (0.1) | F = 289.79 | <0.001 |
| HRV_HFn, mean (SD) | 0.5 (0.1) | 0.5 (0.1) | 0.6 (0.1) | 0.4 (0.1) | F = 731.16 | <0.001 |
| HRV_SD1SD2, mean (SD) | 1.9 (1.7) | 3.5 (2.0) | 1.1 (0.4) | 0.8 (0.3) | F = 1955.53 | <0.001 |
| HRV_DFA_alpha1, median [Q1, Q3] | 0.6 [0.4, 0.7] | 0.3 [0.2, 0.4] | 0.6 [0.5, 0.7] | 0.8 [0.7, 0.9] | H = 2646.74 | <0.001 |
| HRV_DFA_alpha2, median [Q1, Q3] | 0.5 [0.5, 0.7] | 0.5 [0.4, 0.5] | 0.5 [0.5, 0.6] | 0.7 [0.7, 0.9] | H = 1749.94 | <0.001 |
| HRV_SampEn, median [Q1, Q3] | 0.7 [0.5, 1.0] | 0.4 [0.3, 0.6] | 0.8 [0.7, 1.1] | 1.0 [0.8, 1.2] | H = 1743.97 | <0.001 |
| HRV_FuzzyEn, median [Q1, Q3] | 0.8 [0.6, 1.0] | 0.5 [0.4, 0.7] | 1.0 [0.8, 1.2] | 0.8 [0.7, 1.0] | H = 2007.21 | <0.001 |
Data are presented as mean (SD) for normally distributed continuous variables, median [Q1, Q3] for non-normally distributed continuous variables, and n (%) for categorical variables. Statistical differences were assessed using one-way analysis of variance for normally distributed continuous variables, the Kruskal–Wallis H test for non-normally distributed continuous variables, and the chi-square test for categorical variables. Underlying statistical assumptions, including normality and homogeneity of variance, were verified prior to analysis. Binary mutually exclusive variables are reported as the presence of the specific category. F, F-statistic from analysis of variance; H, H-statistic from Kruskal–Wallis test; χ2, chi-square statistic.
AHI, apnea–hypopnea index; BMI, body mass index; DBP, diastolic blood pressure; DFA, detrended fluctuation analysis; ECG, electrocardiogram; HF, high frequency; HFn, normalized high-frequency power; HRV, heart rate variability; LF, low frequency; LFHF, low-frequency-to-high-frequency ratio; LFn, normalized low-frequency power; ODI, oxygen desaturation index; pNN50, percentage of successive normal-to-normal intervals differing by more than 50 ms; REM, rapid eye movement; RMSSD, root mean square of successive differences; SampEn, sample entropy; SBP, systolic blood pressure; SDNN, standard deviation of normal-to-normal intervals; TP, total power.
Overall, Cluster 0 (Stable) demonstrated comparatively lower age, BMI, systolic blood pressure, and apnoea burden, together with relatively preserved sleep efficiency. Cluster 1 (Burden) was characterized primarily by older age with intermediate levels of cardiometabolic and sleep-related burden. In contrast, Cluster 2 (High risk) exhibited higher BMI, elevated systolic blood pressure, more severe sleep-disordered breathing indices, and lower sleep efficiency across sleep stages.
Several representative HRV features also differed significantly among clusters, further supporting physiological differences among the identified phenotypes.
Heart rate variability feature profiles across clusters after heart rate variability integration
After incorporation of stage-specific HRV features, three HRV-informed clusters were identified across the N2, N3, and REM sleep stages (Figure 3). Distinct distributions of HRV features were observed among clusters, indicating heterogeneity in cardiac rhythm dynamics across participant subgroups (Figure 5).
Figure 5.
Cardiac autonomic phenotypic profiles based on integrated clinical and heart rate variability features across N2 (A), N3 (B), and rapid eye movement (C) sleep stages.
Cluster 0 (Stable) demonstrated relatively higher time-domain HRV indices, including RMSSD and SDNN, together with comparatively lower entropy-related measures across sleep stages. In contrast, Cluster 1 (Burden) exhibited relatively higher entropy- and complexity-related HRV features, including sample entropy and multiscale entropy metrics, while maintaining intermediate levels of conventional clinical burden indicators.
Cluster 2 (High risk) showed lower values across multiple time-domain and non-linear HRV features, accompanied by higher cardiometabolic burden and more severe sleep-disordered breathing characteristics. This cluster also demonstrated the highest long-term risks for cardiovascular outcomes across N2, N3, and REM analyses.
Overall, incorporation of HRV features enabled improved physiological subgroup differentiation beyond conventional clinical and sleep-related characteristics, with consistent feature distributions observed across different sleep stages.
Validation of independent prognostic value
After adjustment for baseline clinical covariates, HRV-informed phenotypes remained significantly associated across N2, N3, and REM sleep stages. To further evaluate whether these associations could be explained by conventional respiratory disease severity, sequential sensitivity analyses were performed by progressively adjusting for respiratory burden and sleep-related variables (Table 3).
Table 3.
Respiratory-adjusted sensitivity analyses of HRV-derived cluster associations with long-term clinical outcomes across N2, N3, and REM sleep stages
| Model 1 | Model 2 | Model 3 | ||||
|---|---|---|---|---|---|---|
| HR (95% CI) | P | HR (95% CI) | P | HR (95% CI) | P | |
| N2 stage | ||||||
| Any coronary heart disease | ||||||
| Burden (Cluster 1) | 1.083 (0.945–1.241) | 0.2498 | 1.079 (0.941–1.236) | 0.275 | 1.078 (0.941–1.235) | 0.2799 |
| High risk (Cluster 2) | 1.289 (1.105–1.503) | 0.0012 | 1.282 (1.099–1.495) | 0.0016 | 1.282 (1.099–1.495) | 0.0016 |
| Any cardiovascular disease | ||||||
| Burden (Cluster 1) | 1.226 (1.086–1.385) | 0.001 | 1.219 (1.079–1.376) | 0.0014 | 1.218 (1.078–1.375) | 0.0015 |
| High risk (Cluster 2) | 1.468 (1.280–1.684) | <0.001 | 1.457 (1.270–1.671) | <0.001 | 1.453 (1.267–1.666) | <0.001 |
| All-cause mortality | ||||||
| Burden (Cluster 1) | 0.836 (0.784–0.891) | <0.001 | 0.833 (0.782–0.888) | <0.001 | 0.837 (0.785–0.892) | <0.001 |
| High risk (Cluster 2) | 1.078 (1.001–1.161) | 0.0475 | 1.073 (0.996–1.156) | 0.0638 | 1.081 (1.004–1.165) | 0.04 |
| N3 stage | ||||||
| Any coronary heart disease | ||||||
| Burden (Cluster 1) | 1.079 (0.949–1.227) | 0.2467 | 1.075 (0.946–1.223) | 0.2682 | 1.075 (0.945–1.222) | 0.2726 |
| High risk (Cluster 2) | 1.262 (1.086–1.466) | 0.0024 | 1.257 (1.081–1.460) | 0.0029 | 1.255 (1.080–1.459) | 0.003 |
| Any cardiovascular disease | ||||||
| Burden (Cluster 1) | 1.195 (1.065–1.341) | 0.0024 | 1.190 (1.061–1.335) | 0.003 | 1.187 (1.058–1.332) | 0.0034 |
| High risk (Cluster 2) | 1.426 (1.249–1.629) | <0.001 | 1.418 (1.241–1.620) | <0.001 | 1.409 (1.233–1.610) | <0.001 |
| All-cause mortality | ||||||
| Burden (Cluster 1) | 0.850 (0.797–0.908) | <0.001 | 0.848 (0.794–0.905) | <0.001 | 0.850 (0.796–0.907) | <0.001 |
| High risk (Cluster 2) | 1.075 (0.999–1.157) | 0.0518 | 1.070 (0.995–1.152) | 0.0696 | 1.068 (0.992–1.149) | 0.0809 |
| REM stage | ||||||
| Any coronary heart disease | ||||||
| Burden (Cluster 1) | 1.105 (0.968–1.260) | 0.1396 | 1.102 (0.966–1.258) | 0.1477 | 1.102 (0.966–1.258) | 0.1475 |
| High risk (Cluster 2) | 1.253 (1.074–1.462) | 0.0041 | 1.246 (1.068–1.454) | 0.0052 | 1.247 (1.068–1.454) | 0.0051 |
| Any cardiovascular disease | ||||||
| Burden (Cluster 1) | 1.212 (1.077–1.363) | 0.0014 | 1.208 (1.074–1.359) | 0.0016 | 1.207 (1.073–1.358) | 0.0017 |
| High risk (Cluster 2) | 1.394 (1.216–1.598) | <0.001 | 1.382 (1.205–1.585) | <0.001 | 1.380 (1.203–1.582) | <0.001 |
| All-cause mortality | ||||||
| Burden (Cluster 1) | 0.835 (0.784–0.890) | <0.001 | 0.833 (0.782–0.888) | <0.001 | 0.837 (0.786–0.892) | <0.001 |
| High risk (Cluster 2) | 1.075 (1.000–1.156) | 0.0506 | 1.070 (0.995–1.151) | 0.0686 | 1.069 (0.994–1.150) | 0.0724 |
HRs and 95% CIs were estimated using multivariable Cox proportional hazards models. Cluster 0 was used as the reference group. Model 1 was adjusted for baseline demographic and clinical covariates; Model 2 additionally adjusted for respiratory burden metrics, including Apnea–Hypopnea Index and oxygen desaturation index; Model 3 further adjusted for sleep-related metrics, including arousal index and sleep efficiency. The stability of hazard estimates across models was evaluated to assess the robustness of HRV-derived cluster associations independent of conventional sleep-disordered breathing severity measures.
CI, confidence interval; HR, hazard ratio; HRV, heart rate variability; REM, rapid eye movement.
Across all sleep stages, sequential adjustment for respiratory burden and sleep-related variables resulted in only minimal changes in HRs. The magnitude and direction of the observed associations remained largely unchanged after adjustment for AHI, ODI, arousal index, and sleep efficiency, supporting the robustness of the prognostic value of HRV-informed phenotyping beyond conventional respiratory severity measures. Consistent with these findings, the overall clustering patterns were largely preserved across pairwise sleep-stage comparisons (Figure 6), further supporting the stability and reproducibility of the identified HRV-informed phenotypes.
Figure 6.
Clinical profiles of composite sleep phenotypes and robustness validation across pairwise sleep stage combinations: (A) N2 to N3, (B) N2 to REM, (C) N3 to REM.
Kaplan–Meier survival analyses also demonstrated clear separation of cumulative event-free survival across the three HRV-informed phenotypes for CHD, CVD, and all-cause mortality in the N2, N3, and REM sleep stages (Figure 7), providing complementary visual evidence of their prognostic discrimination.
Figure 7.
Kaplan–Meier survival curves for all-cause mortality and cardiovascular outcomes stratified by stage-specific autonomic phenotypes: N2 (A), N3 (B), and rapid eye movement (C) sleep stages.
Among all analyses, the N2-stage High-risk phenotype (Cluster 2) demonstrated the strongest association with cardiovascular outcomes. For incident CVD, the HR decreased only marginally from 1.468 (95% CI: 1.280–1.684, P < 0.001) in Model 1 to 1.453 (95% CI: 1.267–1.666, P < 0.001) after additional adjustment for respiratory burden and sleep-related variables (Model 3). Similar stability of hazard estimates was observed across the N3 and REM sleep stages, as well as for CHD outcomes, indicating that the prognostic associations of the HRV-derived phenotypes were largely independent of conventional respiratory severity measures.
In contrast, the Burden phenotype (Cluster 1) showed only modest increases in cardiovascular risk despite older age and intermediate clinical burden profiles. For CHD, several associations in Cluster 1 did not remain statistically significant after adjustment. For all-cause mortality, distinct patterns were observed across clusters. The Burden phenotype consistently demonstrated lower all-cause mortality across all three sleep stages, with HRs remaining consistently below 0.90 after sequential adjustment. In contrast, mortality associations for the High-risk phenotype were comparatively weaker and became attenuated after additional adjustment.
Overall, these findings indicate that integration of sleep stage-specific HRV features enables identification of physiologically distinct phenotypes with prognostic relevance beyond conventional clinical and respiratory severity indicators.
Among the three sleep stages, HRV-informed phenotypes derived from N2 sleep generally exhibited slightly stronger associations with cardiovascular outcomes than those derived from N3 or REM sleep, although the overall patterns were highly consistent across stages. Given that N2 sleep represents the dominant proportion of total sleep duration and reflects relatively stable physiological conditions during overnight sleep, HRV features derived from this stage may better capture baseline cardiorespiratory regulatory patterns. However, because no formal interaction testing between sleep stages was performed, these findings should be interpreted cautiously.
Obstructive sleep apnoea severity-stratified subgroup analysis
Subgroup analyses stratified by obstructive sleep apnoea (OSA) severity demonstrated broadly consistent prognostic associations of HRV-informed phenotypes across both AHI strata (Table 4). Among participants with AHI <15 events/h, the High-risk phenotype generally showed increased risks of CHD and consistently showed increased risks of CVD across the N2, N3, and REM sleep stages. For example, in the N2 stage, the High-risk phenotype was associated with a 31.4% increased risk of CHD (HR 1.314, 95% CI 1.050–1.645, P = 0.017) and a 47.3% increased risk of CVD (HR 1.473, 95% CI 1.206–1.799, P < 0.001). Similar associations were observed in the N3 and REM stages.
Table 4.
Sleep stage-specific subgroup analyses of HRV-derived cluster associations with long-term clinical outcomes stratified by OSA severity
| AHI < 15 | AHI ≥ 15 | |||
|---|---|---|---|---|
| HR (95% CI) | P | HR (95% CI) | P | |
| N2 stage | ||||
| Any coronary heart disease | ||||
| Burden (Cluster 1) | 1.081 (0.899–1.301) | 0.407 | 1.036 (0.866–1.238) | 0.702 |
| High risk (Cluster 2) | 1.314 (1.050–1.645) | 0.017 | 1.164 (0.952–1.424) | 0.140 |
| Any cardiovascular disease | ||||
| Burden (Cluster 1) | 1.215 (1.029–1.434) | 0.022 | 1.145 (0.977–1.342) | 0.096 |
| High risk (Cluster 2) | 1.473 (1.206–1.799) | <0.001 | 1.323 (1.108–1.580) | 0.002 |
| All-cause mortality | ||||
| Burden (Cluster 1) | 0.972 (0.881–1.073) | 0.577 | 0.904 (0.808–1.012) | 0.081 |
| High risk (Cluster 2) | 0.879 (0.776–0.995) | 0.042 | 1.018 (0.895–1.158) | 0.785 |
| N3 stage | ||||
| Any coronary heart disease | ||||
| Burden (Cluster 1) | 1.126 (0.929–1.365) | 0.226 | 1.018 (0.840–1.234) | 0.854 |
| High risk (Cluster 2) | 1.353 (1.083–1.691) | 0.008 | 1.173 (0.949–1.451) | 0.140 |
| Any cardiovascular disease | ||||
| Burden (Cluster 1) | 1.258 (1.058–1.496) | 0.0094 | 1.169 (0.985–1.386) | 0.0740 |
| High risk (Cluster 2) | 1.526 (1.249–1.863) | <0.001 | 1.350 (1.119–1.629) | 0.002 |
| All-cause mortality | ||||
| Burden (Cluster 1) | 1.008 (0.911–1.116) | 0.879 | 1.007 (0.893–1.135) | 0.911 |
| High risk (Cluster 2) | 0.978 (0.866–1.103) | 0.715 | 1.022 (0.894–1.168) | 0.750 |
| REM stage | ||||
| Any coronary heart disease | ||||
| Burden (Cluster 1) | 1.086 (0.899–1.311) | 0.393 | 1.063 (0.883–1.279) | 0.518 |
| High risk (Cluster 2) | 1.258 (1.000–1.582) | 0.050 | 1.182 (0.961–1.455) | 0.114 |
| Any cardiovascular disease | ||||
| Burden (Cluster 1) | 1.198 (1.012–1.419) | 0.036 | 1.172 (0.995–1.381) | 0.058 |
| High risk (Cluster 2) | 1.383 (1.126–1.698) | 0.002 | 1.323 (1.102–1.589) | 0.003 |
| All-cause mortality | ||||
| Burden (Cluster 1) | 0.927 (0.838–1.026) | 0.142 | 0.937 (0.833–1.053) | 0.274 |
| High risk (Cluster 2) | 0.888 (0.783–1.007) | 0.064 | 1.041 (0.912–1.189) | 0.550 |
Separate subgroup analyses were performed for N2, N3, and REM sleep stage-derived HRV clusters after stratification by OSA severity using an AHI threshold of 15 events/h. HRs and 95% CIs were estimated using multivariable Cox proportional hazards models with Cluster 0 as the reference group. Models were adjusted for baseline clinical covariates. Directional consistency across OSA strata was evaluated to assess the robustness of HRV-derived cluster associations independent of conventional sleep-disordered breathing severity.
AHI, Apnea–Hypopnea Index; CI, confidence interval; HR, hazard ratio; HRV, heart rate variability; OSA, obstructive sleep apnea; REM, rapid eye movement.
Among participants with AHI ≥15 events/h, the High-risk phenotype remained significantly associated with increased CVD risk across all three sleep stages (N2: HR 1.323, P = 0.002; N3: HR 1.350, P = 0.002; REM: HR 1.323, P = 0.003). Although associations with CHD were generally weaker and did not consistently reach statistical significance, the HRs remained directionally consistent across sleep stages.
The Burden phenotype showed modest associations with cardiovascular outcomes in several subgroup analyses, whereas its associations with all-cause mortality were less consistent across OSA severity strata and sleep stages. Overall, these findings indicate that the prognostic value of HRV-informed phenotypes is broadly preserved across different levels of OSA severity. The consistent associations with cardiovascular outcomes, particularly for the High-risk phenotype, suggest that HRV-informed phenotyping provides complementary prognostic information beyond conventional measures of respiratory disease severity rather than simply reflecting apnoea burden alone.
Discussion
Limitations of conventional apnoea burden-based phenotyping
In this study, clustering based solely on conventional clinical and sleep-related variables demonstrated limited independent prognostic utility after adjustment for baseline covariates. Although traditional indicators such as AHI, age, and obesity remain important markers of disease burden, these variables alone may not fully capture the substantial heterogeneity in long-term cardiovascular outcomes among patients with sleep-disordered breathing.
Our findings suggest that phenotyping approaches relying primarily on respiratory event frequency may not fully capture the physiological heterogeneity associated with OSA. In particular, individuals with similar levels of respiratory burden may demonstrate markedly different cardiovascular trajectories during long-term follow-up. This observation supports incorporation of physiological signal-derived features, such as sleep stage-specific HRV patterns, into cardiovascular risk stratification frameworks.
Importantly, the present findings do not imply that HRV-informed phenotypes or the specific number of identified clusters represent a definitive biological classification. Rather, they demonstrate that incorporating HRV-derived physiological information into conventional clinical phenotyping provides complementary prognostic value beyond traditional respiratory severity metrics alone.
Prognostic associations of heart rate variability-derived physiological phenotypes
Using a consistent clustering framework, the present study identified HRV-informed physiological phenotypes that were associated with distinct long-term cardiovascular risk profiles across multiple sleep stages. More importantly, incorporating HRV features into the clustering framework substantially improved prognostic stratification compared with clustering based solely on conventional clinical variables. Notably, the High-risk phenotype demonstrated consistently elevated risks of CVD and generally higher risks of CHD after multivariable adjustment, whereas the Burden phenotype exhibited a favourable all-cause mortality profile in the primary analyses despite older age and intermediate clinical burden.
These findings suggest that HRV-derived physiological patterns may capture aspects of cardiovascular vulnerability not fully reflected by conventional respiratory burden metrics alone. However, the interpretation of HRV complexity-related features requires caution. Nonlinear HRV measures, including entropy-based indices, likely reflect composite cardiorespiratory interactions involving autonomic regulation, respiratory instability, sleep fragmentation, and cyclic heart rate variation associated with sleep-disordered breathing.
Therefore, HRV complexity measures likely reflect integrated cardiorespiratory regulatory dynamics rather than isolated autonomic function alone. The identified phenotypes should thus be interpreted as data-driven physiological signatures associated with differential long-term outcomes.
Importantly, additional respiratory-adjusted sensitivity analyses demonstrated that the observed associations remained highly stable after adjustment for AHI, ODI, arousal index, and sleep efficiency. Furthermore, subgroup analyses stratified by OSA severity demonstrated that the prognostic associations of HRV-informed phenotypes were generally preserved across both AHI strata, particularly for cardiovascular outcomes, further supporting that these associations were not solely driven by conventional measures of respiratory disease severity. Collectively, these findings suggest that incorporating HRV-derived physiological features into conventional clinical phenotyping provides incremental prognostic information beyond respiratory severity measures alone, irrespective of the exact clustering solution.
Sleep stage-specific characteristics and obstructive sleep apnoea severity-dependent associations
The prognostic associations of the HRV-informed phenotypes remained generally consistent across N2, N3, and REM sleep stages, suggesting that these physiological signatures capture relatively stable patterns of cardiovascular risk stratification rather than isolated stage-dependent fluctuations. This consistency indicates that the prognostic contribution of sleep stage-specific HRV features is not restricted to a particular clustering solution or sleep stage, but instead reflects complementary physiological information beyond conventional clinical characteristics. Among the evaluated stages, N2-derived phenotypes showed slightly stronger associations with cardiovascular outcomes, possibly because N2 sleep represents the longest and relatively more stable sleep stage during overnight recording.
Subgroup analyses stratified by OSA severity demonstrated broadly consistent prognostic associations across both AHI strata, particularly for CVD outcomes. In particular, the High-risk phenotype remained consistently associated with increased CVD risk in both participants with AHI <15 and those with AHI ≥15 across all three sleep stages. Although effect estimates for CHD tended to be larger in the lower-AHI subgroup, the overall direction of the associations remained consistent following OSA severity stratification. These findings further support the robustness of the prognostic value of HRV-informed phenotyping across different levels of respiratory disease burden.
These findings further suggest that the prognostic information captured by HRV-derived physiological signatures cannot be fully explained by conventional respiratory disease severity alone. Although HRV measures inevitably incorporate respiratory-related influences, the generally consistent associations observed across OSA severity strata support the view that HRV-informed phenotyping captures complementary physiological information beyond apnoea burden rather than simply representing another expression of respiratory event frequency.
Another interesting observation was that the favourable all-cause mortality profile observed for the Burden phenotype in the primary analyses became less consistent after stratification by OSA severity. Although several subgroup analyses continued to suggest lower mortality risk, these associations were not uniformly statistically significant across sleep stages and OSA severity strata. This pattern may reflect underlying physiological heterogeneity not fully captured by conventional clinical characteristics. However, because residual confounding and selection bias cannot be fully excluded in this retrospective observational study, these findings should be interpreted cautiously and require further validation in prospective cohorts.
Study limitations and future clinical perspectives
The present study has several limitations. First, this was a retrospective observational analysis based on the SHHS cohort, and causal relationships cannot be established. Although the large sample size and long-term follow-up strengthened statistical reliability, external validation in independent prospective cohorts remains necessary.
Second, HRV-derived features are sensitive to ECG signal quality and may be influenced by respiratory-driven autonomic oscillations, including cyclic variation of heart rate during sleep-disordered breathing. Although additional respiratory-adjusted analyses demonstrated relatively stable associations after adjustment for conventional respiratory burden metrics, the identified HRV phenotypes likely reflect composite cardiorespiratory dynamics rather than isolated intrinsic autonomic function alone.
Third, non-linear HRV measures such as entropy-based metrics remain incompletely validated as direct physiological markers of autonomic regulation. Therefore, these features were interpreted as data-driven physiological signatures associated with long-term outcomes rather than definitive mechanistic markers of autonomic reserve.
In addition, residual confounding cannot be fully excluded in this retrospective observational analysis. Although cardiovascular medication history was incorporated into the revised baseline analyses, medication use may still represent a potential source of residual confounding because commonly prescribed cardiovascular drugs can influence HRV measurements. Notably, baseline beta-blocker use did not differ significantly across the identified clusters, making it less likely that differential beta-blocker therapy alone accounted for the observed prognostic differences. Nevertheless, because medication use reflects both treatment effects and the underlying CVD burden, its independent contribution cannot be completely disentangled in this observational study. Furthermore, although automated ECG quality control and artefact correction were performed before HRV feature extraction, manual beat-by-beat annotation was not feasible in this large population-based cohort. Consequently, residual measurement error related to ECG quality or undetected arrhythmias cannot be entirely excluded. In addition, unmeasured socioeconomic status, healthcare access, behavioural factors, and other lifestyle-related variables may also have partially influenced both HRV patterns and long-term clinical outcomes. Future studies incorporating direct autonomic measurements, multimodal physiological monitoring, and external validation cohorts are warranted to further clarify the biological and clinical significance of these findings.
Despite these limitations, the present findings suggest that incorporating sleep stage-specific HRV features into conventional clinical phenotyping provides complementary prognostic information beyond traditional respiratory burden metrics alone. Rather than emphasizing a single clustering solution, our findings support the incremental value of HRV-informed phenotyping for long-term cardiovascular risk stratification in patients with sleep-disordered breathing. Because the analytical framework relies primarily on ECG-derived signals, future integration with wearable monitoring systems may facilitate scalable and clinically applicable cardiovascular risk assessment in sleep-disordered breathing populations.
Conclusion
This study identified sleep stage-specific HRV-informed phenotypes associated with distinct long-term cardiovascular risk profiles across the N2, N3, and REM sleep stages. The observed associations remained stable after sequential adjustment for conventional respiratory burden and sleep-related variables.
These findings suggest that conventional measures of sleep-disordered breathing severity alone may not fully capture the heterogeneity of long-term cardiovascular risk among individuals with OSA. Rather than emphasizing a specific clustering solution, our findings demonstrate that incorporating sleep stage-specific HRV features into conventional clinical phenotyping provides complementary prognostic information beyond traditional clinical characteristics and respiratory severity metrics. Across all sleep stages, the High-risk phenotype consistently demonstrated elevated cardiovascular risk, whereas the Burden phenotype was associated with lower all-cause mortality despite older age and intermediate clinical burden.
The identified phenotypes likely reflect integrated physiological interactions involving autonomic regulation, respiratory instability, and sleep-related cardiorespiratory dynamics rather than isolated autonomic function alone. Overall, HRV-informed phenotyping represents a promising complementary strategy for long-term cardiovascular risk stratification in patients with OSA and warrants further validation in prospective studies.
Supplementary Material
Contributor Information
Xiao Chen, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing 102218, China.
Luqi Bai, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing 102218, China.
Chong Li, School of Biomedical Engineering, Tsinghua University, Beijing 100084, China.
Boda Zhou, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing 102218, China.
Supplementary material
Supplementary material is available at European Heart Journal – Digital Health.
Author contributions
Xiao Chen (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review & editing [equal]), Luqi Bai (Conceptualization, Methodology, Software, Writing—review & editing [equal]), Chong Li (Conceptualization, Funding acquisition, Project administration, Supervision, Writing—review & editing [equal]), and Boda Zhou (Conceptualization, Funding acquisition, Project administration, Supervision, Writing—original draft, Writing—review & editing [equal])
Artificial intelligence usage statement
Artificial intelligence (AI) tools were utilized exclusively for language refinement, English translation, and proofreading. No generative AI was used for data analysis, data interpretation, or content generation.
Funding
This work was supported by Beijing Municipal Health Commission (2024-3-034), Beijing Hospitals Authority (ZLRK202518), and Tsinghua University (ZY01_07 and 2611T0812).
Data availability
All data used in this study are available from the SHHS upon approved application (https://sleepdata.org/datasets/shhs). Relevant R and python code is available from the corresponding author upon reasonable request.
References
- 1. Ravichandran R, Gupta L, Singh M, Nag A, Thomas J, Panjiyar BK. The interplay between sleep disorders and cardiovascular diseases: a systematic review. Cureus 2023;15:e45898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Sarode R, Nikam PP. The impact of sleep disorders on cardiovascular health: mechanisms and interventions. Cureus 2023;15:e49703. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Gan TM, Yang YJ, Mo GL, Wang SR, Li SH, Li JY. Sleep disorders are closely associated with coronary heart disease in US adults (≥20 years): a cross-sectional study. Medicine (Baltimore) 2024;103:e39698. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Kiss MG, Cohen O, McAlpine CS, Swirski FK. Influence of sleep on physiological systems in atherosclerosis. Nat Cardiovasc Res 2024;3:1284–1300. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Eckert DJ, Sands SA. Hypoxia and sleep-disordered breathing: friend or foe? Am J Respir Crit Care Med 2022;205:869–872. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Gilmartin GS, Lynch M, Tamisier R, Weiss JW. Chronic intermittent hypoxia in humans during 28 nights results in blood pressure elevation and increased muscle sympathetic nerve activity. Am J Physiol Heart Circ Physiol 2010;299:H925–H931. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Mehra R, Chung MK, Olshansky B, Dobrev D, Jackson CL, Kundel V, et al. Sleep-disordered breathing and cardiac arrhythmias in adults: mechanistic insights and clinical implications: a scientific statement from the American Heart Association. Circulation 2022;146:e119–e136. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Solhjoo S, Haigney MC, Punjabi NM. Sleep-disordered breathing destabilizes ventricular repolarization: cross-sectional, longitudinal, and experimental evidence. Heart Rhythm 2025;22:808–816. [DOI] [PubMed] [Google Scholar]
- 9. Guilleminault C, Tilkian A, Dement WC. The sleep apnea syndromes. Annu Rev Med 1976;27:465–484. [DOI] [PubMed] [Google Scholar]
- 10. Malhotra A, Ayappa I, Ayas N, Collop N, Kirsch D, McArdle N, et al. Metrics of sleep apnea severity: beyond the apnea-hypopnea index. Sleep 2021;44:zsab030. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Azarbarzin A, Sands SA, Stone KL, Taranto-Montemurro L, Messineo L, Terrill PI, et al. The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the osteoporotic fractures in men study and the Sleep Heart Health Study. Eur Heart J 2019;40:1149–1157. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Task force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Circulation 1996;93:1043–1065. [PubMed] [Google Scholar]
- 13. Bonnet MH, Arand DL. Heart rate variability: sleep stage, time of night, and arousal influences. Electroencephalogr Clin Neurophysiol 1997;102:390–396. [DOI] [PubMed] [Google Scholar]
- 14. Lyu J, Yeh CH, Shi W, Penzel T. Dynamic brain-heart interaction in sleep characterized by variational phase-amplitude coupling framework. Commun Biol 2025;8:1235. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Shinar Z, Akselrod S, Dagan Y, Baharav A. Autonomic changes during wake-sleep transition: a heart rate variability based approach. Auton Neurosci 2006;130:17–27. [DOI] [PubMed] [Google Scholar]
- 16. Lipsitz LA, Goldberger AL. Loss of ‘complexity’ and aging. Potential applications of fractals and chaos theory to senescence. JAMA 1992;267:1806–1809. [PubMed] [Google Scholar]
- 17. Pincus SM. Approximate entropy as a measure of system complexity. Proc Natl Acad Sci U S A 1991;88:2297–2301. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Richman JS, Moorman JR. Physiological time-series analysis using approximate entropy and sample entropy. Am J Physiol Heart Circ Physiol 2000;278:H2039–H2H49. [DOI] [PubMed] [Google Scholar]
- 19. Goldberger AL, Amaral LA, Hausdorff JM, Ivanov P, Peng CK, Stanley HE. Fractal dynamics in physiology: alterations with disease and aging. Proc Natl Acad Sci USA 2002;99:2466–2472. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Ivanov PC, Amaral LA, Goldberger AL, Havlin S, Rosenblum MG, Struzik ZR, et al. Multifractality in human heartbeat dynamics. Nature 1999;399:461–465. [DOI] [PubMed] [Google Scholar]
- 21. Somers VK, Dyken ME, Mark AL, Abboud FM. Sympathetic-nerve activity during sleep in normal subjects. N Engl J Med 1993;328:303–307. [DOI] [PubMed] [Google Scholar]
- 22. Haldar P, Pavord ID, Shaw DE, Berry MA, Thomas M, Brightling CE, et al. Cluster analysis and clinical asthma phenotypes. Am J Respir Crit Care Med 2008;178:218–224. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Burgel PR, Roche N, Paillasseur JL, Tillie-Leblond I, Chanez P, Escamilla R, et al. Clinical COPD phenotypes identified by cluster analysis: validation with mortality. Eur Respir J 2012;40:495–496. [DOI] [PubMed] [Google Scholar]
- 24. Ohayon MM, Carskadon MA, Guilleminault C, Vitiello MV. Meta-analysis of quantitative sleep parameters from childhood to old age in healthy individuals: developing normative sleep values across the human lifespan. Sleep 2004;27:1255–1273. [DOI] [PubMed] [Google Scholar]
- 25. Zhang GQ, Cui L, Mueller R, Tao S, Kim M, Rueschman M, et al. The national sleep research resource: towards a sleep data commons. J Am Med Inform Assoc 2018;25:1351–1358. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Quan SF, Howard BV, Iber C, Kiley JP, Nieto FJ, O'Connor GT, et al. The Sleep Heart Health Study: design, rationale, and methods. Sleep 1997;20:1077–1085. [PubMed] [Google Scholar]
- 27. Kleiger RE, Miller JP, Bigger JT, Moss AJ. Decreased heart rate variability and its association with increased mortality after acute myocardial infarction. Am J Cardiol 1987;59:256–262. [DOI] [PubMed] [Google Scholar]
- 28. van Ravenswaaij-Arts C, Kollee L, Hopman J, Stoelinga GBA, Geijn H. Heart rate variability. Ann Intern Med 1993;118:436–447. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
All data used in this study are available from the SHHS upon approved application (https://sleepdata.org/datasets/shhs). Relevant R and python code is available from the corresponding author upon reasonable request.