Association of accelerated biological aging with obstructive sleep apnea symptoms and identification of a candidate biomarker gene signature

Abstract

Background

Obstructive sleep apnea (OSA) is a common sleep-related breathing disorder that is linked to cardiovascular, metabolic, and neurocognitive complications. However, its diagnosis relies on polysomnography, which is complex and resource-intensive, leading to frequent underdiagnosis. Emerging evidence suggests that accelerated biological aging may contribute to OSA pathophysiology, but systematic assessments using biological age metrics are limited.

Methods

Data from the National Health and Nutrition Examination Survey (NHANES) were analyzed to evaluate associations between biological age acceleration and symptom-based OSA risk. Weighted multivariable logistic regression was used to assess the relationships of KDM-Age and PhenoAge accelerations with symptom-based OSA risk. Bioinformatics analyses of the GSE135917 dataset identified aging-related differentially expressed genes (DEGs). Machine learning algorithms, including the least absolute shrinkage and selection operator (LASSO) and support vector machine–recursive feature elimination (SVM-RFE), were used to screen hub genes, which were validated in both external cohorts and a chronic intermittent hypoxia (CIH) mouse model.

Results

Higher KDM-Age and PhenoAge accelerations were independently associated with increased symptom-based OSA risk (both P < 0.001). Thirty aging-related DEGs were identified, which were mainly enriched in senescence, inflammatory, and immune pathways. Three hub genes-RBBP4, UCHL1, and ERRFI1-were selected by machine learning and exhibited favorable discriminative potential across validation datasets and the CIH model. In addition, an integrated three-gene predictive model demonstrated promising discriminative ability in the training set and acceptable predictive performance in independent validation datasets. A nomogram integrating these genes showed good calibration and demonstrated value as an exploratory analytical tool at this stage.

Conclusions

Accelerated biological aging is significantly associated with symptom-based OSA risk. The identified three-gene candidate biomarker signature links aging-related alterations to OSA and warrants further validation.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13148-026-02108-x.

Introduction

Obstructive sleep apnea (OSA) is a disorder characterized by recurrent episodes of partial (hypopnea) or complete (apnea) upper airway obstruction during sleep, with disease severity commonly measured by the apnea–hypopnea index (AHI). Substantial clinical evidence links OSA to serious public health consequences, including cardiovascular diseases, neurodegenerative disorders, malignancies, and motor vehicle accidents due to excessive daytime sleepiness [1, 2]. In clinical practice, OSA diagnosis primarily relies on polysomnography (PSG), the gold-standard method, or its simplified variant, respiratory polygraphy (RP-PSG). However, these diagnostic approaches face practical challenges, including prolonged testing duration (typically requiring overnight monitoring), high costs, and stringent technical requirements, collectively contributing to widespread underdiagnosis [3]. Consequently, many patients fail to receive timely diagnosis and treatment, thereby exacerbating health risks and societal burdens. This underscores an urgent need for more efficient and accessible screening-based approaches to identify individuals at high risk for OSA, particularly in large populations.

Epidemiologically, OSA prevalence is strongly age dependence. Specifically, the incidence in older adults (> 65 years) is approximately 2–3 times that in middle-aged individuals (30–65 years) [4]. OSA’s health impacts also vary by age. A pivotal finding is that OSA is significantly associated with cancer-related mortality in individuals younger than 65, but not in older populations [5]. This suggests that OSA may influence disease pathogenesis by accelerating biological aging processes beyond chronological aging alone. Studies indicate that OSA’s hallmark features—intermittent hypoxia and sleep fragmentation—can induce pathophysiological changes consistent with accelerated aging, potentially via telomere attrition, stem cell dysfunction, and epigenetic dysregulation [6]. Moreover, OSA-associated accelerated aging correlates with multiple end-organ diseases, and anti-aging interventions may enhance the reversibility of these changes [7], further supporting the mechanistic role of aging in OSA progression.

The aging process exhibits dual characteristics: it is related to, yet distinct from chronological age. Clinical observations of “premature aging” reveal that functional decline can occur independently of chronological age, underscoring the need for more precise aging assessment methods. However, unidimensional biomarkers face limitations: structural markers (e.g., telomere length) often require advanced detection techniques and change slowly, while functional markers (e.g., inflammatory cytokines) are prone to transient fluctuations [8]. These limitations have motivated the development of biological age estimation models, which integrate multidimensional clinical and molecular features using machine learning. Such models outperform chronological age in predicting disease risk and mortality, particularly in chronic diseases [9, 10]. Although OSA’s association with accelerated aging is established, systematic evaluations linking biological age to symptom-based OSA risk in population-based settings remain limited, thereby hindering both mechanistic understanding and clinical translation.

At the molecular level, growing evidence implicates OSA in accelerating cellular aging through multiple pathways. Intermittent hypoxia-OSA’s defining feature-triggers oxidative stress, inflammatory cascades, and circadian disruption, leading to hallmarks of aging such as telomere shortening [11], advanced glycation end-product accumulation [12], and senescence-associated secretory phenotype (SASP) activation [6]. These findings suggest a potential role for aging-related gene (ARG) networks in OSA pathogenesis. However, current knowledge of ARG regulatory mechanisms in OSA remains incomplete, limiting the development of targeted therapies.

To address these gaps, this study employs a multidimensional strategy: we establish quantitative links between biological age and symptom-based OSA risk through epidemiological analysis; then, we systematically decoding ARG regulatory networks and evaluate their potential biomarker value using machine learning. The findings are expected to advance screening-oriented risk stratification and deepen mechanistic understanding of OSA, while providing a foundation for future validation in clinically diagnosed cohorts.

Materials and methods

Study design and population

The National Health and Nutrition Examination Survey (NHANES) employs a complex, multistage probability sampling design to collect nationally representative data encompassing demographics, socioeconomic status, physical examinations, biospecimen collection, and questionnaire responses. This cross-sectional analysis used data from the 2007–2008 NHANES cycle, which included questionnaire-derived indicators of OSA-related symptoms and essential parameters for calculating biological age. Initially, 4025 participants aged ≥ 40 years were enrolled from this cycle. Figure 1 illustrates the study’s inclusion and exclusion criteria. Detailed information about the NHANES database is available at https://www.cdc.gov/nchs/nhanes/index.htm.

Fig. 1.

Participant selection flowchart from the NHANES 2007–2008 cross-sectional study

Symptom-based OSA risk classification

Symptom-based OSA risk classification (“high-risk” and “low-risk”) was evaluated using a sleep disorder questionnaire, following established methods [13]. Specifically, the questionnaire assessed snoring, daytime sleepiness, witnessed apnea (gasping/choking), and hypertension—four signs/symptoms recommended by the American Academy of Sleep Medicine for OSA screening [14]. Participants reporting two or more of these symptoms (frequently), or reporting a prior physician diagnosis of sleep apnea were classified as high risk based on symptoms [13. Others were categorized as low-risk.

Biological aging

PhenoAge was calculated using established methods [9] based on aging-related variables: albumin, alkaline phosphatase, creatinine, glucose, C-reactive protein concentrations, lymphocyte percentage (proportion of white blood cells), mean cell volume, red cell distribution width, and white blood cell count. KDM-Age was derived from clinical biomarkers including forced expiratory volume in 1 s, systolic blood pressure, albumin, alkaline phosphatase, blood urea nitrogen, creatinine, C-reactive protein, glycated hemoglobin, and total cholesterol [15].

We employed two distinct methodological approaches to quantify accelerated aging: (1) The direct difference method, where KDM-Age advance and PhenoAge advance were calculated through simple arithmetic subtraction of chronological age from algorithm-estimated biological age (KDM-Age or PhenoAge). This unadjusted metric directly reflects the absolute discrepancy between biological and chronological age, with positive values indicating accelerated aging. (2) The residual adjustment approach, wherein KDM-Age Acceleration and PhenoAge Acceleration were derived as residuals from linear regression models (with either KDM-Age or PhenoAge as the dependent variable and chronological age as the independent variable). This method effectively eliminates the linear dependence between chronological age and biological age estimates, with positive residuals representing faster-than-average aging rates relative to age-matched peers. Both approaches were implemented to provide complementary perspectives on aging acceleration while addressing different aspects of age-related variance.

Covariates

Covariates were selected based on biological relevance and published literature and included: (1) Demographic characteristics: age, sex (male/female), race/ethnicity (Mexican American, non-Hispanic White, non-Hispanic Black, other Hispanic, other races), BMI, education level; (2) Lifestyle factors: poverty-income ratio (PIR), smoking status, and alcohol user; (3) Clinical comorbidities: hypertension, diabetes mellitus, and dyslipidemia. Hypertension and diabetes mellitus were determined based on self-report [16]. The diagnosis of hyperlipidemia required meeting any of the following criteria: (1) triglycerides ≥ 150 mg/dL; (2) total cholesterol ≥ 200 mg/dL; (3) LDL-C ≥ 130 mg/dL; (4) HDL-C level (< 50 mg/dL in women or < 40 mg/dL in men); (5) use of lipid-lowering medication [17].

Microarray data acquisition and aging-related genes

Gene expression profiles from human transcriptomic datasets comparing normal controls and OSA patients were obtained from the Gene Expression Omnibus (GEO), including GSE135917 (subcutaneous adipose tissue) [18], GSE38792 (visceral adipose tissue) [19], and GSE75097 (peripheral blood mononuclear cells) [20]. In GSE135917, after excluding 24 OSA subjects who had received continuous positive airway pressure (CPAP) treatment, 8 controls and 34 OSA patients were included. GSE38792 included 8 controls and 10 OSA patients, and GSE75097 included 6 controls and 28 OSA patients after excluding CPAP-treated subjects. Probe names were converted to gene names using Perl with platform annotation files, including GPL6244 for GSE135917 and GSE38792 (NetAffx build 35; official annotation available up to July 1, 2016) and GPL10904 for GSE75097 (annotation version available up to January 28, 2016); when a single probe mapped to multiple genes, only the first annotated gene symbol was retained, and when multiple probes mapped to the same gene, probe-level expression values were averaged at the gene level using the avereps function in the limma package. All datasets were downloaded as processed series matrix files; therefore, no additional background correction was applied, and between-array normalization was performed using the limma package with default quantile normalization (normalizeBetweenArrays).

Aging-related genes (ARGs) were compiled from The Molecular Signatures Database (MSigDB), a publicly available resource for functional gene sets [21]. After removing duplicates, 1153 unique ARGs were retained for analysis (Supplementary Table 1).

Identification and functional analysis of aging-related DEGs

The expression profiles of aging-related genes (ARGs) were extracted from the GSE135917 training dataset, and differential expression analysis was performed using the limma package. The aging-related differentially expressed genes (OSA-ARDEGs) in OSA patients were identified with screening criteria of |log2 fold change (logFC)| > 0.585 and adjusted P value < 0.05. This threshold was selected to balance biological relevance and statistical sensitivity in the context of a relatively small sample size, allowing detection of moderate but potentially meaningful expression changes associated with OSA. The expression patterns of these DEGs were subsequently visualized through a heatmap.

Pearson correlation coefficients between DEGs were calculated using the corrplot package. Functional enrichment analyses, including Gene Ontology (GO; biological process, cellular component, and molecular function) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses, were performed with clusterProfiler, with top pathways visualized using ggplot2 and GOplot (top 8 for each GO category and top 10 for KEGG pathways). Protein–protein interaction networks were constructed via STRING and visualized in Cytoscape (v3.9.1), with hub genes identified by degree centrality using CytoHubba (top 10 nodes).

Immune infiltration analysis

CIBERSORT estimated abundances of 22 immune cell types. Wilcoxon tests compared infiltration levels between OSA and control groups. Spearman correlations linked ARDEGs to immune cells. Lollipop plots highlighted significant hub gene-immune cell pairs (red labels indicate P < 0.05).

Consensus clustering

Unsupervised clustering (ConsensusClusterPlus) partitioned OSA patients into subtypes using 30 OSA-ARDEGs (k = 2–9, PAM algorithm, Euclidean distance). Optimal clusters were determined via consensus matrices and cumulative distribution functions. PCA quantified transcriptional differences between the identified clusters. This analysis provides a data-driven framework for molecular stratification of OSA patients.

Machine learning algorithms

A multi-algorithm integrative approach was employed to identify key genes: (1) Least Absolute Shrinkage and Selection Operator (LASSO) regression (glmnet) with leave-one-out cross-validation (LOOCV) (optimal λ selected as lambda.min, corresponding to the minimum cross-validated deviance), and (2) Support Vector Machine-Recursive Feature Elimination (SVM-RFE) implemented based on the e1071 package and a published mSVM-RFE algorithm, with feature ranking based on linear SVM-derived weight vectors and cross-validation–based evaluation of feature subsets through iterative feature elimination. The final candidate hub genes were determined by taking the intersection of significant genes identified through both approaches. Nomograms were constructed using the rmda package as an exploratory modeling approach, with model performance assessed through calibration curves, decision curve analysis, and clinical impact curves.

ROC analysis

The pROC package generated ROC curves for candidate hub genes in training/validation datasets, with AUC values quantifying discriminative performance.

GSEA analysis

Gene set enrichment analysis (GSEA) was performed using the clusterProfiler package with MSigDB gene sets (c5.go. symbols and c2.cp.kegg. symbols). Samples were stratified by median expression of hub genes. Top 6 significant pathways (P < 0.05) were visualized.

Chronic intermittent hypoxia animal model

Eight male SPF-grade C57BL/6 J mice (6–7 weeks old, 18–22 g) were obtained from the Animal Experimental Center of Renmin Hospital, Wuhan University. Mice were housed under standard conditions (12-h light/dark cycle) with free access to food and water. After one week of acclimation, they were randomly divided into control and CIH groups (n = 4 per group). The CIH model was induced using an intermittent hypoxia system (ProOx-100, TOW-INT TECH, China). Oxygen concentration was cycled every 120 s, with FiO₂ reduced to 6% for 60 s by nitrogen infusion and then restored to 21% within 60 s using a mixture of air and oxygen. This exposure was performed for 8 h/day for 6 consecutive weeks, while controls were kept under normoxia. After modeling, mice were euthanized, and adipose tissues were collected for further analysis. All procedures were approved by the Institutional Animal Care and Use Committee of Renmin Hospital of Wuhan University (IACUC No. 20220723 A) and followed NIH guidelines for animal care.

Real-time PCR analysis

Total RNA was extracted from mouse adipose tissues using TRIzol reagent (Thermo Fisher Scientific, USA), and RNA concentration was determined with a NanoDrop spectrophotometer (Thermo Fisher Scientific, USA). cDNA was synthesized using the Transcriptor First Strand cDNA Synthesis Kit (Roche, Switzerland), and real-time quantitative PCR was performed with SYBR Green qPCR Master Mix (High ROX) (Servicebio, Wuhan, China). GAPDH served as the internal control, and relative gene expression levels were calculated using the 2^ − ΔΔCt method. Primer sequences used for RT-qPCR are listed in Supplementary Table 2.

Hematoxylin and eosin (HE) staining

Adipose tissues were fixed overnight in adipose tissue fixative (Cat. No. CR2210071, Servicebio, China), embedded in paraffin, and sectioned at a thickness of 4 μm. The sections were stained with hematoxylin and eosin (HE) and examined under a light microscope. Adipocyte diameters were measured in five randomly selected fields (100 ×) per section using ImageJ software.

Statistical analysis

All analyses incorporated NHANES sampling weights to account for the complex survey design, using the full sample 2-year MEC examination weights, with weighted baseline characteristics stratified by symptom-based OSA risk status. Continuous variables were expressed as mean ± standard error (SE) while categorical variables were presented as percentages (95% confidence intervals). We used weighted multivariable logistic regression to examine the association between biological age (analyzed as a continuous and categorical variable) and symptom-based OSA risk. Three continuous models were constructed: Model 1 was a crude model; Model 2 adjusted for gender, race, education level (Less than 9th grade/9th-12th grade/High school graduate or GED/Some college or Associate degree/College graduate or above), poverty-income ratio, smoking status, and alcohol user; and Model 3 additionally adjusted for BMI (categorical variables), hyperlipidemia, and diabetes mellitus. Hypertension was excluded from adjustment as it constitutes part of the high-risk of OSA symptoms diagnostic criteria. The deadline for data acquisition and download is December 27, 2023, and the study design and data analysis time is from April 10, 2025 to June 5, 2025.

All analyses were performed using the statistical software packages R (http://www.R-project.org, The R Foundation), with two-tailed p-values < 0.05 considered statistically significant.

Results

Baseline characteristics of participants

The study comprised 2198 participants, with weighted proportions indicating 47.66% (95% CI: 41.06–54.26) as low symptom-based OSA risk and 52.34% (95% CI: 42.42–62.25) as high symptom-based OSA risk (Table 1). While chronological age differences between groups were minimal (low-risk: 54.25 ± 0.47 years; high-risk: 55.31 ± 0.32 years; P = 0.03), significant differences were observed in biological aging markers (all P < 0.0001). Participants classified as high symptom-based OSA risk exhibited elevated KDM-Age (59.36 ± 0.56 vs. 53.52 ± 0.76), PhenoAge (58.87 ± 0.43 vs. 55.50 ± 0.52), KDM-Age advance (4.05 ± 0.37 vs. − 0.73 ± 0.40), PhenoAge advance (3.56 ± 0.27 vs. 1.24 ± 0.21), KDM-Age acceleration (0.46 ± 0.38 vs. − 4.40 ± 0.42), PhenoAge acceleration (0.53 ± 0.27 vs. − 1.74 ± 0.21) (all P < 0.001). Metabolic parameters were markedly worse in the high-risk group, including higher BMI (31.10 ± 0.33 vs. 27.05 ± 0.22), greater obesity prevalence (48.85% vs. 22.89%), and increased rates of hypertension (59.03% vs. 17.08%) and diabetes (14.04% vs. 6.69%) (all P < 0.001). Gender distribution showed marginally more males in the high-risk group (51.94% vs. 45.73%; P = 0.02), while educational attainment was lower (college graduates: 24.07% vs. 33.94%; P = 0.004). No significant differences were detected in race (P = 0.21), smoking status (P = 0.97), or alcohol user (P = 0.65) between the high- and low-risk groups.

Table 1.

Baseline characteristics of participants by symptom-based OSA risk status

Variable	Total	Low-risk group	High-risk group	P - value
N %	NA	47.66(41.06,54.26)	52.34(42.42,62.25)
Age, years	54.81(0.34)	54.25(0.47)	55.31(0.32)	0.03
BMI, kg/m2	29.17(0.21)	27.05(0.22)	31.10(0.33)	< 0.001
KDM-Age	56.58(0.56)	53.52(0.76)	59.36(0.56)	< 0.001
PhenoAge	57.26(0.44)	55.50(0.52)	58.87(0.43)	< 0.001
KDM-Age advance	1.77(0.31)	− 0.73(0.40)	4.05(0.37)	< 0.001
Phenoage advance	2.45(0.20)	1.24(0.21)	3.56(0.27)	< 0.001
KDM-Age acceleration	− 1.85(0.33)	− 4.40(0.42)	0.46(0.38)	< 0.001
Phenoage acceleration	− 0.55(0.20)	− 1.74(0.21)	0.53(0.27)	< 0.001
BMI group, n%				< 0.001
Normal/underweight	26.46(21.30,31.62)	36.04(31.54,40.54)	17.74(13.85,21.62)
Overweight	37.07(31.53,42.60)	41.07(37.27,44.87)	33.42(29.61,37.22)
Obesity	36.47(29.59,43.35)	22.89(19.69,26.09)	48.85(42.88,54.81)
PIR ratio, n%				0.86
< 1	9.24(7.39,11.08)	9.27(6.99,11.55)	9.20(6.48,11.93)
1- < 3	31.27(23.87,38.68)	30.59(23.67,37.52)	31.90(25.74,38.06)
> = 3	59.49(47.75,71.22)	60.14(52.10,68.17)	58.90(51.50,66.31)
Gender, n%				0.02
Female	51.02(43.46,58.57)	54.27(51.18,57.37)	48.06(44.69,51.42)
Male	48.98(41.00,56.96)	45.73(42.63,48.82)	51.94(48.58,55.31)
Race, n%				0.21
Mexican american	5.73(3.68, 7.78)	6.42(4.07,8.77)	5.10(2.79,7.42)
Non-hispanic black	8.78(6.06,11.50)	8.28(5.21,11.35)	9.24(4.98,13.50)
Non-hispanic white	76.98(60.66,93.30)	75.71(69.82,81.60)	78.15(71.02,85.27)
Other hispanic	3.63(1.95, 5.30)	3.79(1.89,5.68)	3.48(1.70,5.27)
Other race—including multi-racial	4.87(2.85, 6.90)	5.80(2.85,8.76)	4.03(2.37,5.69)
Smoking status, n%				0.97
No	49.83(41.91,57.76)	49.75(43.48,56.02)	49.91(45.92,53.89)
Yes	50.17(41.37,58.97)	50.25(43.98,56.52)	50.09(46.11,54.08)
Alcohol user, n%				0.65
No	25.03(19.95,30.11)	24.41(19.48,29.34)	25.59(22.45,28.73)
Yes	74.97(63.49,86.45)	75.59(70.66,80.52)	74.41(71.27,77.55)
Hypertension, n%				< 0.001
No	60.96(52.20,69.73)	82.92(79.86,85.99)	40.97(37.60,44.34)
Yes	39.04(31.51,46.56)	17.08(14.01,20.14)	59.03(55.66,62.40)
Diabetes mellitus, n%				< 0.001
No	87.49(73.93,101.05)	91.58(89.31,93.85)	83.76(80.21,87.32)
Borderline	1.97(1.31, 2.64)	1.72(0.82,2.63)	2.20(1.34,3.06)
Yes	10.54(7.78, 13.29)	6.69(4.82, 8.56)	14.04(10.65,17.42)
Hyperlipidemia, n%				0.07
No	17.67(15.59,19.76)	20.38(16.62,24.14)	15.21(12.90,17.51)
Yes	82.33(69.01,95.64)	79.62(75.86,83.38)	84.79(82.49,87.10)
Education level, n%				0.004
College graduate	28.78(22.86,34.70)	33.94(29.07,38.82)	24.07(18.56,29.58)
Less than college	71.22(59.10,83.35)	66.06(61.18,70.93)	75.93(70.42,81.44)

Significant results are in bold. Continuous variables were presented as mean (SE), and categorical variables as percentage (95% CI)

SE standard error, BMI body mass index, OSA obstructive sleep apnea

Association of biological aging measures with symptom-based OSA risk

As shown in Table 2, KDM-age advance (continuous) was significantly associated with symptom-based OSA risk in unadjusted Model 1 (OR = 1.04 per unit increase; 95% CI: 1.03–1.05; P < 0.001). This association persisted after adjusting for education level, PIR ratio, smoking status, alcohol user (Model 2: OR = 1.04; 95% CI: 1.02–1.07; P < 0.001) and further adjustment for BMI and comorbidities (Model 3: OR = 1.03; 95% CI: 1.02–1.04; P < 0.001). When analyzed as a binary variable (> 0 vs. < 0), accelerated aging conferred a 2.22-fold higher risk in Model 1 (95% CI: 1.67–2.95), which remained significant in fully adjusted models (Model 3: OR = 1.75; 95% CI: 1.30–2.35). The gradual attenuation of effect sizes (Model 1 → 3: 2.22 → 2.25 → 1.75) suggests partial confounding while confirming the independent effect of biological aging. Quartile analysis in the fully adjusted model demonstrated significantly elevated odds of high symptom-based OSA risk for higher KDM-age advance groups compared to Q1 (reference), including Q3 (OR = 1.90; 95% CI: 1.26–2.85, P = 0.004) and Q4 (OR = 1.84; 95% CI: 1.28–2.64, P = 0.002). Consistent results were observed for KDM-Age acceleration, further reinforcing the robust relationship between accelerated biological aging and symptom-based OSA risk.

Table 2.

Association of KDM-Age measures with high symptom-based OSA risk

Variables	Model 1		Model 2		Model 3
	OR (95% CI)	P - value	OR (95% CI)	P - value	OR (95% CI)	P - value
KDM-Age advance	1.04(1.03,1.05)	< 0.001	1.04(1.02,1.07)	0.02	1.03(1.02,1.04)	< 0.001
*KDM-Age advance
< 0	Ref		Ref		Ref
> 0	2.22(1.67,2.95)	< 0.001	2.25(1.17,4.32)	0.03	1.75(1.30,2.35)	0.001
KDM-Age advance quartile
Q1	Ref		Ref		Ref
Q2	1.62(1.13,2.32)	0.01	1.63(1.13,2.35)	0.01	1.36(0.99,1.87)	0.06
Q3	2.45(1.66,3.62)	< 0.001	2.56(1.65,3.97)	< 0.001	1.90(1.26,2.85)	0.004
Q4	2.64(1.88,3.72)	< 0.001	2.78(1.89,4.09)	< 0.001	1.84(1.28,2.64)	0.002
KDM-Age acceleration	1.04(1.03,1.05)	< 0.001	1.04(1.02,1.07)	0.02	1.03(1.02,1.04)	< 0.001
KDM-Age acceleration quartile
Q1	Ref		Ref		Ref
Q2	1.70(1.28,2.26)	0.001	1.70(1.27,2.27)	0.001	1.47(1.14,1.90)	0.01
Q3	2.36(1.57,3.55)	< 0.001	2.44(1.56,3.81)	< 0.001	1.85(1.21,2.83)	0.01
Q4	2.70(1.88,3.88)	< 0.001	2.79(1.91,4.07)	< 0.001	1.86(1.27,2.72)	0.003

Model 1: Unadjusted. Model 2: Adjusted for gender, race, education level, PIR ratio (categorical variables), smoking status, alcohol user. Model 3: Adjusted for gender, race, education level, PIR ratio (categorical variables), smoking status, alcohol user, BMI (categorical variables), hyperlipidemia, and diabetes mellitus

OR odds ratio, 95% CI 95% confidence interval, BMI body mass index, OSA obstructive sleep apnea

^*KDM-Age advance: Values = 0 set as missing (NA)

Table 3 shows the association between PhenoAge measures and high symptom-based OSA risk. PhenoAge advance demonstrated a positive association with high symptom-based OSA risk. Each unit increase corresponded to 7% higher odds of high risk in Model 1 (OR = 1.07; 95% CI: 1.05–1.09; P < 0.001), remaining significant after full adjustment (Model 3: OR = 1.03; 95% CI: 1.02–1.05; P < 0.001). Binary analysis showed accelerated individuals (> 0) had 2.15-fold greater risk in Model 1 (95% CI: 1.72–2.69), attenuating to 1.44-fold in Model 3 (95% CI: 1.13–1.85; P = 0.01). Quartile analysis revealed that compared to Q1 (reference), Q2 (OR = 1.50; 95% CI: 1.17–1.92; P = 0.004), Q3 (OR = 2.14; 95% CI: 1.71–2.69; P < 0.001) and Q4 (OR = 2.95; 95% CI: 2.05–4.24; P < 0.001) all showed significantly elevated OSA symptoms risk in Model 1. However, only Q3 (OR = 1.39; 95% CI: 1.03–1.87; P = 0.03) and Q4 (OR = 1.68; 95% CI: 1.18–2.39; P = 0.01) maintained statistical significance in the fully adjusted Model 3. Similar patterns were observed for PhenoAge acceleration.

Table 3.

Association of PhenoAge measures with high symptom-based OSA risk

Variables	Model 1		Model 2		Model 3
	OR (95% CI)	P - value	OR (95% CI)	P - value	OR (95% CI)	P - value
PhenoAge advance	1.07(1.05,1.09)	< 0.001	1.07(1.02,1.11)	0.02	1.03(1.02,1.05)	< 0.001
#PhenoAge advance
< 0	Ref		Ref		Ref
> 0	2.15(1.72,2.69)	< 0.001	2.06(1.25,3.40)	0.02	1.44(1.13,1.85)	0.01
PhenoAge advance quartile
Q1	Ref		Ref		Ref
Q2	1.50(1.17,1.92)	0.004	1.43(1.09,1.88)	0.01	1.18(0.87,1.61)	0.27
Q3	2.14(1.71,2.69)	< 0.001	2.02(1.51,2.70)	< 0.001	1.39(1.03,1.87)	0.03
Q4	2.95(2.05,4.24)	< 0.001	2.81(1.87,4.23)	< 0.001	1.68(1.18,2.39)	0.01
PhenoAge acceleration	1.07(1.05,1.09)	< 0.001	1.06(1.02,1.11)	0.03	1.03(1.01,1.05)	< 0.001
PhenoAge acceleration quartile
Q1	Ref		Ref		Ref
Q2	1.64(1.26,2.13)	0.001	1.55(1.15,2.07)	0.01	1.26(0.93,1.70)	0.13
Q3	2.18(1.73,2.75)	< 0.001	2.07(1.56,2.73)	< 0.001	1.47(1.10,1.97)	0.01
Q4	2.97(2.07,4.26)	< 0.001	2.83(1.91,4.19)	< 0.001	1.69(1.17,2.43)	0.01

Model 1: Unadjusted. Model 2: Adjusted for gender, race, education level, PIR ratio (categorical variables), smoking status, alcohol user. Model 3: Adjusted for gender, race, education level, PIR ratio (categorical variables), smoking status, alcohol user, BMI (categorical variables), hyperlipidemia, and diabetes mellitus

OR odds ratio, 95% CI 95% confidence interval, BMI body mass index, OSA obstructive sleep apnea

^#PhenoAge advance: Values = 0 set as missing (NA)

Identification of OSA-ARDEGs and functional enrichment analysis

In the training dataset GSE135917, we identified 30 OSA-ARDEGs using the R package limma with predefined filtering criteria. Among these, 12 genes were upregulated and 18 were downregulated in OSA (Fig. 2A). Correlation analysis revealed significant interrelationships (either positive or negative) among the 30 OSA-ARDEGs (Fig. 2B). GO enrichment analysis demonstrated that biological processes (BP) were primarily associated with cellular stress responses (e.g., to chemical stimuli or radiation), hormonal regulation (e.g., glucocorticoid and corticosteroid responses), smooth muscle cell proliferation, and transcriptional regulation (including miRNA-mediated positive regulation). Cellular components (CC) were enriched in transcriptional regulatory complexes, kinase signaling complexes, and organelle-specific functions. Molecular functions (MF) involved RNA polymerase II-specific transcription factor activity, protein-modifying binding (e.g., ubiquitin/ubiquitin-like ligase binding), and signal transducer binding (Fig. 2C). KEGG analysis highlighted pathways related to cellular senescence, viral infection, and oncogenesis, including cellular senescence, TNF and Jak-STAT signaling, Kaposi sarcoma-associated herpesvirus infection, and transcriptional misregulation in cancer. Disease-associated pathways such as hepatitis B and Salmonella infection were also enriched, indicating potential biological relevance to systemic inflammatory and stress-related processes (Fig. 2D). PPI network analysis identified 23 interconnected OSA-ARDEGs after removing isolated nodes (Fig. 2E), with the top 10 hub genes (JUN, ATF3, IL6, EGR1, MYC, FOS, PTGS2, CDKN1A, MCL1, and CCN2) determined by degree algorithm (Fig. 2F).

Fig. 2.

Identification and functional analysis of aging-related differentially expressed genes (DEGs). A Heatmap of 30 aging-related DEGs between OSA samples and control samples. B Correlation analysis of the 30 OSA-associated aging-related DEGs (OSA-ARDEGs). C Bubble plot showing the top 8 significantly enriched Gene Ontology (GO) terms. D Bar plot displaying the top 10 enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. E Protein–protein interaction (PPI) network of 23 OSA-ARDEGs constructed using STRING. F PPI network of the top 10 hub genes, with node color indicating rank (red: highest; orange: lowest)

Immunoinfiltration analysis revealed significant differences between OSA patients and controls (Fig. S1A). OSA patients exhibited increased infiltration of memory B cells, resting NK cells, and M1 macrophages but reduced activated dendritic cells compared to controls (Fig. S1B). Further analysis showed that OSA-ARDEGs correlated significantly with activated dendritic cells, M0/M2 macrophages, and partially with resting NK cells and regulatory T cells (Tregs) (Fig. S1C). Together, these findings suggest that OSA is accompanied by altered immune cell composition, characterized by increased pro-inflammatory cell infiltration and changes in immune regulatory cell populations, and that these alterations are associated with aging-related gene expression patterns.

Molecular subtyping of OSA based on aging-related genes

Unsupervised consensus clustering of 30 OSA-ARDEGs evaluated cluster stability across k = 2–9 (Fig. 3A–C). By comprehensively evaluating the silhouette coefficient and cumulative distribution function curve, we found that when k = 2, consensus clustering showed the best subtype discrimination (Fig. 3D), and this conclusion was further verified in the consensus matrix heat map of k = 3 (Fig. 3E). Patients were classified into C1 (n = 10, 33.3%) and C2 (n = 24, 66.7%) subtypes. Differential expression analysis revealed distinct OSA-ARDEGs expression profiles between subtypes (Fig. 3F). Specifically, UCHL1, CCND2, GLB1, LAPTM5, SQSTM1, SLC9A7, MIR10A, MIR29B2, EEF1E1, DDIT3, RBBP4, and GCA were significantly upregulated in C2, while the remaining 18 genes were higher in C1 (Fig. 3G). To visualize the spatial separation of the identified clusters, PCA was performed based on the global expression patterns of ARDEGs. The results demonstrated a clear separation between the two clusters in the reduced dimensional space(Fig. 3H). Additionally, C2 showed higher infiltration of resting NK cells and M0 macrophages but lower M2 macrophages and activated dendritic cells compared to C1 (Fig. S2A, B), providing new insights into the molecular heterogeneity of OSA.

Fig. 3.

Identification of aging-related expression patterns based on OSA-ARDEGs. A Cumulative distribution function (CDF) curves for k = 2 to 9. B Relative change in area under CDF curves for k = 2–9. C Consensus scores for each subtype (k = 2 to 9). D, E Consensus clustering matrices for k = 2 and k = 3. F, G Heatmap and boxplots showing expression patterns of 30 OSA-ARDEGs in subtypes C1 and C2 (^*P < 0.05, ^**P < 0.01, ^***P < 0.001). H Principal component analysis (PCA) showing the distribution of samples based on the identified clusters

Machine learning identifies OSA-specific aging biomarkers

From the initial pool of 30 OSA-ARDEGs, we employed two machine learning algorithms for key gene selection. The LASSO regression algorithm (Fig. 4A, B) identified five signature genes, while the SVM-RFE algorithm (Fig. 4C, D) selected four candidate genes. Intersection analysis revealed three consensus genes: RBBP4, UCHL1, and ERRFI1 (Fig. 4E). To validate their discriminative potential, the three feature genes were systematically evaluated across multiple datasets. In the GSE135917 training set, all three genes demonstrated promising discriminative performance: RBBP4 (AUC = 0.949, 95% CI: 0.875–0.996), UCHL1 (AUC = 0.908, 95% CI: 0.809–0.985), and ERRFI1 (AUC = 0.765, 95% CI: 0.485–0.982) (Fig. 4F). Furthermore, the three feature genes exhibited reasonable predictive performance in two independent validation datasets, GSE38792 (Fig. 4G) and GSE75097 (Fig. S3A). Subsequently, an integrated three-gene model was constructed to assess the combined discriminative ability of these features. In the GSE135917 training set, the model achieved an AUC of 0.971 (95% CI: 0.912–1.000) (Fig. S3B). In two independent validation datasets, the model yielded an AUC of 1.000 (95% CI: 1.000–1.000) in GSE38792 (Fig. S3C) and an AUC of 0.756 (95% CI: 0.548–0.929) in GSE75097 (Fig. S3D).

Fig. 4.

Feature gene identification using machine learning algorithms. A Cross-validated binomial deviance of least absolute shrinkage and selection operator (LASSO) regression, with the optimal regularization parameter (λ = lambda.min) indicated. B Coefficient profiles of the 30 OSA-ARDEGs generated by LASSO regression. C Cross-validation accuracy curve of support vector machine-recursive feature elimination (SVM-RFE). D SVM-RFE cross-validation error curve. E Venn diagram showing overlapping feature genes identified by LASSO and SVM-RFE. F Receiver operating characteristic (ROC) curves demonstrating discriminative value of the three selected feature genes in the training dataset GSE135917. G ROC curves for the three feature genes in the validation dataset GSE38792

Furthermore, in vivo experiments provided additional confirmation: a CIH-based OSA mouse model was established by 6 weeks of intermittent hypoxia exposure, and HE staining showed significantly larger adipocytes in the CIH group compared with controls (Fig. S4A, B). qPCR analysis further revealed that RBBP4 and UCHL1 mRNA levels were significantly upregulated, while ERRFI1 was downregulated in CIH mice relative to the control group (Fig. S4C–E), consistent with our bioinformatics results. These findings highlight the discriminative potential of these genes, though their clinical applicability requires further validation.

Using these three key genes, we developed a nomogram model for OSA risk prediction (Fig. 5A). The model integrates standardized expression scores to calculate individual disease probabilities. Calibration curve analysis indicated good agreement between predicted probabilities and observed outcomes (Fig. 5B). Decision curve analysis (DCA) revealed superior net benefit across the 0–1 threshold probability range compared to alternative approaches (Fig. 5C). Clinical impact curve (CIC) analysis further confirmed the model's stable predictive performance in risk stratification (Fig. 5D). These findings suggest that our three-gene nomogram provides both discriminative accuracy and clinically actionable risk assessment.

Fig. 5.

Construction and evaluation of a three-gene nomogram model. A Nomogram integrating RBBP4, UCHL1, and ERRFI1 for OSA distinguishing. B Calibration curve for nomogram validation. C Clinical impact curve evaluating model utility. D Decision curve analysis of the nomogram model

GSEA and immune infiltration analysis of signature genes

GSEA revealed distinct biological pathways associated with each gene. RBBP4 showed significant enrichment in immune-related pathways including adaptive immune response, regulation of cell activation, and T cell receptor signaling (Fig. 6A, D), suggesting its immunomodulatory role. UCHL1 was primarily associated with cell proliferation, smooth muscle contraction regulation, and oxidative phosphorylation (Fig. 6B, E), indicating involvement in growth and energy metabolism. ERRFI1 demonstrated enrichment in antigen presentation, lysosomal function, and steroid biosynthesis pathways (Fig. 6C, F), implicating its dual role in immune and metabolic regulation.

Fig. 6.

Enrichment and immune cell correlation analyses of feature genes. A-C Gene Set Enrichment Analysis (GSEA) of GO terms for RBBP4 A, UCHL1 B, and ERRFI1 C high/low expression groups. D-F GSEA of KEGG pathways for RBBP4 D, UCHL1 E, and ERRFI1 F groups. G-I Lollipop plots showing correlations between RBBP4 G, UCHL1 H, ERRFI1 I expression and immune cell infiltration

Immune correlation analysis substantiated these findings. RBBP4 exhibited positive correlation with M0 macrophages (p = 0.032) but negative association with activated dendritic cells (p = 0.003) (Fig. 6G). UCHL1 showed positive correlations with both M0 macrophages (p = 0.005) and regulatory T cells (p = 0.020), while maintaining negative association with activated dendritic cells (p = 0.017) (Fig. 6H). ERRFI1 displayed a unique pattern, strongly correlating with activated dendritic cells (P < 0.001) but inversely associated with M0 macrophages (p = 0.044) and resting NK cells (p = 0.020) (Fig. 6I). These results collectively highlight the genes’ specific roles within immune regulatory networks.

Discussion

This study employed an integrative analytical framework combining epidemiological cohorts, multi-omics profiling, and machine learning to explore the association between biological aging processes and symptom-based OSA risk at the population level, and to characterize aging-related molecular features in OSA using transcriptomic data. Our analyses revealed statistically significant correlations between accelerated biological aging (KDM-Age acceleration/PhenoAge acceleration) and a higher symptom-based OSA risk, independent of confounding factors such as BMI. The Candidate biomarker gene signature (RBBP4, UCHL1, and ERRFI1) suggests potential mechanistic links between OSA OSA-related systemic alterations and cellular aging pathways.

Aging is a complex biological process characterized by progressive degenerative changes in tissue structure and physiological function. Numerous studies have demonstrated close associations between OSA and aging processes. Previous literature reports that compared with healthy controls, OSA patients show significant impairments across multiple cognitive domains including alertness, coordination, executive function, as well as verbal and visuospatial memory [22]. Emerging evidence further indicates that untreated OSA is not only closely associated with aging-related cognitive dysfunction [23] but may also increase the risk of early dementia and neurodegenerative diseases [22, 24]. Given that accelerated biological aging has been clearly identified as a significant risk factor for multiple malignancies [25], in-depth exploration of the relationship between biological aging and OSA holds substantial value for reducing public health burdens. BA, as a comprehensive indicator integrating multidimensional factors including genetic background, environmental exposures, lifestyle, and psychological states, provides a more holistic quantitative tool for assessing individual aging processes [26].

This study first conducted a cross-sectional analysis based on large-scale population data to systematically evaluate the association between two biological age indicators (KDM-Age and PhenoAge) and symptom-based OSA risk. The findings revealed that although the high symptom-based OSA risk group and low-risk group had similar chronological ages, the former showed significantly higher biological age indicators along with poorer metabolic health (reflected by higher BMI, and higher prevalences of hypertension and diabetes). This result strongly aligns with previous research showing that OSA patients generally exhibit typical clinical features of accelerated aging (such as chronic low-grade inflammation and metabolic disorders) [24]. Multivariate regression analyses further confirmed that accelerated biological age (including KDM-Age and PhenoAge) showed positive associations with symptom-based OSA risk, and these associations remained statistically significant after adjusting for potential confounders such as chronological age and BMI. Particularly noteworthy, quartile analysis revealed a clear dose–response relationship between them, indicating a gradient association pattern between the degree of biological aging and symptom-based OSA risk. Notably, biological age indicators such as KDM-Age and PhenoAge intrinsically incorporate metabolic- and inflammation-related components, many of which overlap with established OSA risk factors including obesity and metabolic disorders. To address this potential overlap, we applied stepwise adjustment models with and without metabolic covariates, allowing us to distinguish aging-related signals that are independent of traditional metabolic risk factors. Although the strength of the associations was attenuated after further adjustment for BMI and metabolic comorbidities, the persistence of statistically significant associations suggests that accelerated biological aging may capture additional pathophysiological information beyond obesity-related mechanisms alone. This important finding suggests that biological age indicators may provide a quantitative framework for the symptom-based OSA risk.

Although the aging process exhibits significant stochastic characteristics [27], recent advances in aging biology reveal that aging rate is at least partially regulated by evolutionarily highly conserved gene networks and biochemical pathways [28]. This insight provides a theoretical foundation for developing intervention aimed at delaying or mitigating age-related pathological changes and extending healthspan [29]. Given the continuously rising incidence of OSA and its aging-related complications, elucidating the pathophysiological mechanisms underlying cellular aging in OSA is of particular importance. Based on this, our study identified 30 OSA-ARDEGs through transcriptomic analysis, primarily derived from adipose tissue samples. It should be noted that, compared with whole blood or peripheral blood mononuclear cells (PBMCs), adipose tissue is not an optimal or readily accessible sample type for routine clinical screening or biomarker development. However, accumulating evidence indicates that chronic intermittent hypoxia, a hallmark of OSA, can induce inflammatory and metabolic remodeling in adipose tissue, promoting immune cell infiltration, adipokine dysregulation, and transcriptional reprogramming, thereby contributing to systemic metabolic and inflammatory alterations in patients with OSA [30–33]. Because adipose tissue dysfunction triggers the release of pro-inflammatory cytokines into the systemic circulation, it actively cross-talks with circulating immune cells, including PBMCs. This shared hypoxic and inflammatory milieu provides a biological rationale for why an adipose-derived gene signature may retain partial transcriptional relevance and detectable signals in peripheral blood. The OSA-ARDEGs identified in this study exhibited significant enrichment in immune regulation, cellular stress response and metabolic pathways, with particularly notable enrichment in cellular senescence, TNF signaling pathway and Jak-STAT pathway-patterns that are consistent with the established pathophysiological features of OSA, including intermittent hypoxia and chronic inflammation [34, 35]. Accordingly, the three-gene signature identified here should be regarded as a candidate biomarker gene signature, whose primary value lies in providing mechanistic insight into systemic molecular alterations associated with OSA, rather than serving as a directly translatable clinical screening tool.

At the molecular level, aging-related transcriptional alterations were accompanied by distinct immune cell phenotypic changes in OSA. Specifically, our study observed decreased infiltration levels of activated dendritic cells in OSA patients, a finding consistent with previously reported reductions of circulating dendritic cells in moderate-to-severe OSA patients [36]. More crucially, activated dendritic cells exhibited significant correlations coordinated changes between OSA-ARDEGs expression patterns and immune cell functional states. Alterations in macrophage polarization were also evident in OSA samples. Increased infiltration of M1 macrophages was observed, along with significant associations between OSA-ARDEGs and M0/M2 macrophage subsets. These observations highlight macrophage polarization abnormalities as a prominent immunological feature co-occurring with aging-related transcriptional changes in OSA. Consistent with this notion, previous studies have reported dysregulation of the monocyte–macrophage system in untreated OSA patients, characterized by reduced proportions of classical monocytes and increased pro-inflammatory CD16⁺ subsets, including intermediate and non-classical monocytes [37]. In addition, compensatory immune activation has been described in OSA. For example, a study involving 48 patients with severe OSA reported increased numbers of cytotoxic NK and NKT cells despite impaired adaptive immune function, further underscoring the complexity of immune alterations associated with the disease [38]. Taken together, these findings depict an immune microenvironment in OSA characterized by concurrent aging-related transcriptional changes and immune cell remodeling, reflecting the heterogeneous and multifaceted nature of immune dysregulation under intermittent hypoxia. Furthermore, unsupervised clustering based on OSA-ARDEGs delineated distinct immune infiltration patterns between the C1 and C2, suggesting the presence of molecularly heterogeneous aging-related expression profiles in OSA rather than discrete disease categories.

To explore aging-related molecular features associated with OSA, this study employed machine learning algorithms to identify a candidate biomarker gene signature composed of three core genes (RBBP4, UCHL1, and ERRFI1). These genes demonstrated discriminative ability in the training dataset and showed reproducible performance across independent validation datasets, suggesting their potential relevance as aging-associated molecular features in OSA. Consistent expression trends of these genes were further supported by in vivo experiments using a CIH mouse model, providing preliminary biological plausibility. All three genes have unique mechanisms in aging regulation: Retinoblastoma-binding protein 4 (RBBP4), as a core component of chromatin remodeling complexes, plays pivotal roles in maintaining genomic stability, stem cell self-renewal and cell fate determination by regulating epigenetic modifications such as histone deacetylation and H3K27 methylation [39]. Ubiquitin carboxyl-terminal hydrolase L1 (UCHL1), the most abundant deubiquitinating enzyme in the nervous system, not only maintains neuronal functional integrity [40] but also participates in regulating proliferation-senescence balance in various cells through coordinating protein homeostasis and damage response mechanisms [41]. ERBB receptor feedback inhibitor 1 (ERRFI1/MIG6), as a “molecular brake” of EGFR signaling pathway, not only promotes tumorigenesis when downregulated [42] but may also lead to cellular premature senescence through sustained inhibition of mitotic signals [43].

In the context of OSA, chronic intermittent hypoxia is known to induce systemic inflammatory stress and immune dysregulation, processes that are closely linked to aging-related transcriptional alterations. Consistent with this biological background, the GSEA results observed for RBBP4, UCHL1, and ERRFI1-particularly enrichment in immune-related pathways, oxidative phosphorylation, lysosomal function, and cell proliferation-suggest that these genes may be involved in molecular responses to hypoxia-associated inflammatory and metabolic stress rather than disease-specific signaling pathways. This interpretation is further supported by immune cell correlation analyses, in which the three genes exhibited distinct but overlapping associations with macrophage subsets, dendritic cells, regulatory T cells, and NK cells, a pattern compatible with the heterogeneous immune remodeling observed under chronic intermittent hypoxia and systemic inflammation in OSA. Taken together, although direct associations between these genes and OSA have not been reported previously, integrative GSEA and immune cell correlation analysis indicate that the RBBP4-UCHL1-ERRFI1 molecular network may be involved in aging- and inflammation-related processes observed in OSA, potentially through coordinated effects on epigenetic regulation, protein homeostasis, and growth signal modulation. These observations should be interpreted as exploratory and hypothesis-generating, and further mechanistic validation using targeted experimental models will be required to clarify their functional roles.

Several limitations of this study should be acknowledged. First, the epidemiological analysis was based on cross-sectional data from NHANES, which precludes causal inference between accelerated biological aging and high symptom-based OSA risk. In addition, the classification of OSA in NHANES relied on a symptom-based screening questionnaire rather than polysomnography, and thus reflects classification into the high symptom-based OSA risk group rather than clinically diagnosed OSA. Second, although multiple biological age indicators were examined and adjusted for key confounders, residual confounding cannot be fully excluded, particularly given the complex overlap between biological aging metrics and metabolic comorbidities such as obesity, diabetes, and dyslipidemia. Third, the transcriptomic analyses were conducted using publicly available GEO datasets with relatively small sample sizes, which may increase the risk of overfitting and limit the stability and generalizability of the identified gene signature. Specifically, the perfect classification (AUC = 1.000) observed in the GSE38792 validation set is likely a biologically implausible artifact driven by these limitations. Moreover, the training and validation datasets were derived from different tissue sources. Because this three-gene signature was fundamentally derived from adipose tissue (the training set), it intrinsically captures localized adipose pathology. Consequently, its more modest predictive performance in PBMCs (GSE75097, AUC = 0.756) strongly suggests that these molecular features are largely tissue-specific rather than universally generalizable. The signal is likely “diluted” in the systemic circulation, which explains the reduced—albeit still biologically relevant—performance in blood. Fourth, although in vivo validation using a chronic intermittent hypoxia mouse model provided supportive biological evidence, gene expression patterns are inherently tissue- and species-specific, and the findings from adipose tissue may not directly translate to other clinically accessible tissues. Finally, the three-gene signature identified in this study should be regarded as an exploratory, hypothesis-generating molecular feature rather than a diagnostic tool, and its clinical relevance requires further validation in larger, well-powered, and prospectively phenotyped cohorts.

Conclusion

The present study demonstrates that accelerated biological aging is associated with an increased risk of symptom-based OSA. The identified aging-related gene signature may provide preliminary molecular insight into OSA-related biological alterations and warrants further validation in independent cohorts.

Abbreviations

AHI

Apnea–hypopnea index

ARGs

Aging-related genes

ARDEGs

Aging-related differentially expressed genes

BA

Biological age

PhenoAge

Phenotypic age

KDM-Age

Klemera-doubal method biological age

NHANES

National health and nutrition examination survey

BP

Biological process

CC

Cellular component

CIC

Clinical impact curve

DCA

Decision curve analysis

LASSO

Least absolute shrinkage and selection operator

MF

Molecular function

ROC

Receiver operating characteristic

SVM-RFE

Support vector machine-recursive feature elimination

SASP

Senescence-associated secretory phenotype

Author contributions

Ke Hu (Corresponding Author) oversaw Conceptualization, Funding acquisition, Writing (original draft and review), Supervision, Project administration, and Resources. Yixuan Wang was involved in Conceptualization, Investigation, Methodology, Formal analysis, and Writing (original draft and review). Yuhan Wang participated in Investigation, Methodology, Formal analysis, and Writing (original draft). Qingfeng Zhang contributed to Validation, Software, Formal analysis, and Writing (review). Jiali Xiong assisted in Investigation, Methodology, and Resources. Beini Zhou performed Validation, Visualization, Writing (review), and Project administration. Mengcan Wang and Shujuan Wu supported Validation, Software, and Resources.

Funding

This work was supported by the National Natural Science Foundation of China (82270101).

Data availability

Publicly available datasets from the Gene Expression Omnibus GEO and the National Health and Nutrition Examination Survey (NHANES, https://www.cdc.gov/nchs/nhanes/index.html) were analyzed in this study. The GEO datasets used include GSE135917, GSE38792, and GSE75097. All analysis scripts supporting the findings of this study are publicly available at GitHub (https://github.com/sleeprenmin/agingsleep/tree/master). Additional information related to the analysis is available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

All procedures were approved by the Institutional Animal Care and Use Committee of Renmin Hospital of Wuhan University (IACUC No. 20220723 A) and followed NIH guidelines for animal care.

Consent for publication

Not applicable.

Informed consent

Not applicable.

Competing interest

The authors declare that they have no competing interests.