Integrating epidemiological and transcriptomic data reveals novel lipid metabolic drivers of obstructive sleep apnea Authors: Yanhui Liu, Sisi Gong, Jiaqi Wang, Shanshan Chen, Yiming Zeng, Chunmei Fan Published: 2026 ABSTRACT AbstractBackgroundObstructive sleep apnea (OSA) is characterized by recurrent upper airway obstruction during sleep and is frequently accompanied by dyslipidemia. However, the molecular mechanisms linking lipid metabolism to OSA remain incompletely understood. This study aimed to investigate the association between blood lipid levels and OSA and explore potential underlying molecular pathways.MethodsData from the National Health and Nutrition Examination Survey (NHANES) 2005–2008 and 2015–2018 were analyzed to evaluate the associations between OSA and blood lipid parameters using multivariable regression and sensitivity analyses. Additionally, OSA-related transcriptomic data (GSE135917) were obtained from the Gene Expression Omnibus (GEO), and lipid metabolism–related genes were retrieved from the Molecular Signatures Database (MSigDB). Differentially expressed lipid metabolism–related genes (DELMRGs) were identified through data integration. Machine learning approaches, protein–protein interaction network analysis, and receiver operating characteristic analysis were applied to identify key genes. Gene set enrichment analysis (GSEA) was performed to elucidate associated biological pathways, and transcription factor–gene and gene–microRNA regulatory networks were constructed using NetworkAnalyst and Cytoscape.ResultsAnalysis of NHANES data showed that triglyceride (TG) and low-density lipoprotein cholesterol (LDL-C) levels were positively associated with OSA, whereas high-density lipoprotein cholesterol (HDL-C) levels was inversely associated (all p < 0.001). A nonlinear, inverted U-shaped association between TG levels and OSA risk was also observed (p for nonlinearity < 0.05). Transcriptomic analysis identified 34 DELMRGs, among which CYP3A4, CYP4A22, and MED18 emerged as key genes. GSEA revealed pathways potentially involved in lipid metabolism and OSA pathophysiology, while regulatory network analyses further supported the biological relevance of these genes.ConclusionsThis study demonstrates that dyslipidemia characterized by elevated TG and LDL-C levels and reduced HDL-C levels is associated with an increased likelihood of OSA, and identifies three DELMRGs that may be involved in OSA pathophysiology. These findings provide exploratory mechanistic insights and offer a basis for future studies to further investigate the role of lipid metabolism in OSA.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12986-026-01160-x.Keywords: Obstructive sleep apnea, Lipid metabolism, NHANES, Bioinformatics analysis, Triglyceride Background Obstructive sleep apnea (OSA) is a common yet frequently underdiagnosed sleep disorder characterized by recurrent partial or complete upper airway obstruction during sleep, resulting in repeated episodes of apnea or hypopnea, chronic intermittent hypoxemia, and sleep fragmentation [1]. If left untreated, OSA can result in serious health complications, including type 2 diabetes, insulin resistance, systemic inflammation, atherosclerosis, metabolic syndrome, and obesity, conditions that are well-established contributors to both morbidity and mortality [2]. Epidemiological studies have demonstrated marked geographical variation in the prevalence of OSA. For example, reported prevalence rates in Asian populations range from 3% to 97%, while global estimates suggest a prevalence between 9% and 38%, highlighting considerable heterogeneity across populations [3, 4]. Given its high prevalence, OSA has become a major global public health concern [5]. Nevertheless, due to its heterogeneous clinical manifestations and the limitations of current diagnostic approaches, OSA remains substantially underdiagnosed, underscoring the need for reliable biomarkers to facilitate early detection and risk stratification. Accumulating evidence suggests a close association between dyslipidemia and OSA. Intermittent hypoxia (IH), a hallmark feature of OSA, has been shown in animal models to promote dyslipidemia and hepatic oxidative stress by disrupting lipid metabolism, impairing lipoprotein clearance, and increasing lipid production [6]. Clinical studies have further demonstrated that even mild OSA, accompanied by sleep-related hypoxemia, is associated with dysregulated lipid profiles [7]. Notably, treatments for OSA, such as continuous positive airway pressure (CPAP) therapy and multilayer sleep surgery, have been shown to significantly improve metabolic parameters [8, 9]. These findings suggest a clear link between lipid metabolism and OSA. However, the exact mediating factors and underlying signaling pathways responsible for abnormal lipid metabolism in OSA patients remain unclear. Elucidating OSA pathogenesis from the perspective of lipid metabolism may therefore provide important insights into disease mechanisms and support the development of early diagnostic strategies. With advances in bioinformatics, large-scale datasets have increasingly been applied to identify disease-associated genes and molecular pathways [10]. In the present study, data from the U.S. National Health and Nutrition Examination Survey (NHANES) (2005–2008 and 2015–2018) were used to investigate the association between blood lipid levels and OSA. Furthermore, differentially expressed genes (DEGs) of OSA identified from the Gene Expression Omnibus (GEO) database were integrated with lipid metabolism–related genes (LMRGs) derived from public databases. Comprehensive bioinformatics analyses were performed to identify OSA-associated differentially expressed lipid metabolism–related genes (DELMRGs) and to elucidate their potential molecular mechanisms. By integrating epidemiological and transcriptomic data, this study aimed to clarify the relationship between blood lipid levels and OSA, and to provide preliminary insights into the mechanisms and clinical implications of lipid metabolism dysregulation in OSA. Methods Research design The NHANES was initiated in 1999 and is a continuous, nationally representative survey conducted in the United States to assess the health and nutritional status of the civilian, non-institutionalized population. The study protocols were reviewed and approved by the Ethics Review Committee of the National Center for Health Statistics, and all participants provided written informed consent [11]. Data from the 2005–2008 and 2015–2018 were included in the present study, as sleep-related questionnaire data were available only during these periods. A total of 39,722 participants were initially enrolled. Participants who did not complete the sleep questionnaire in full (n = 14,551) were excluded. In addition, 14,745 participants with missing lipid profile data, including total cholesterol (TC), triglycerides (TG), low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C), were excluded. After further excluding 1,191 individuals younger than 20 years and 1,229 participants with incomplete demographic or health-related information, including sex, race, and diabetes status, a total of 8,006 participants were included in the final analysis (Fig. 1). Study flowchart. Blue boxes represent the NHANES analysis, purple boxes indicate transcriptomic data sources, green boxes correspond to the key gene screening step, and orange boxes are used for the biological characterization of the key genes Assessment of OSA The assessment of OSA was based on self-reported questionnaire data related to sleep habits and sleep-related disorders. In accordance with previously published studies [12, 13], OSA was defined by the presence of one or more commonly reported symptoms: (1) frequent snoring occurring on three or more nights per week (SLQ030); (2) frequent snorting, wheezing, shortness of breath, or observed breathing pauses during sleep on more than three nights per week (SLQ040); and (3) excessive daytime sleepiness, defined as self-reported extreme sleepiness during the daytime occurring 16–30 times per month (SLQ120). Participants who reported at least one of these symptoms were classified as having symptoms suggestive of OSA. This symptom-based definition was originally proposed by Scinicariello et al. [14] and has been widely applied in subsequent population-based studies [15–17]. Measurement of blood lipid levels Blood lipid parameters, including TC, TG, LDL-C, and HDL-C, were obtained from NHANES laboratory data. Dyslipidemia was defined as an abnormal lipid profile characterized by elevated TC, TG, or LDL-C levels and/or reduced HDL-C levels, according to widely accepted clinical definitions. Measurement of covariates Several covariates were included in the analysis, including age, sex, race, education level, marital status, family poverty income ratio (PIR), body mass index (BMI), waist circumference (WC), smoking status, hypertension, and diabetes. Age was categorized into two groups (20–40 years and ≥ 40 years) based on previous studies [2, 6, 12, 18]. Race was classified into five categories. Education level and marital status were each categorized into three groups. Socioeconomic status was assessed using PIR and classified into three levels. BMI was calculated as weight in kilograms divided by the square of height in meters (kg/m²) and categorized into three levels. Detailed classifications of all covariates are presented in Table 1. Smoking status was defined based on item “SMQ020” from the “SMQ_I” questionnaire. Hypertension was defined using item “BPQ020” from the “BPQ_I” questionnaire. Diabetes was defined according to item “DIQ010” from the “DIQ_I” questionnaire. Table 1Basic characteristics of participants with or without OSACharacteristicsNon-OSAOSAp-valueN39624044TC (mmol/L)4.95 ± 1.094.99 ± 1.080.047TG (mmol/L)1.26 ± 0.741.42 ± 0.77< 0.001LDL-C (mmol/L)2.89 ± 0.932.98 ± 0.95< 0.001HDL-C (mmol/L)1.48 ± 0.441.36 ± 0.40< 0.001Age (years)< 0.001 < 401550 (39.12%)1147 (28.36%) >=402412 (60.88%)2897 (71.64%)Sex< 0.001 Male1695 (42.78%)2188 (54.10%) Female2267 (57.22%)1856 (45.90%)Race0.023 Mexican American604 (15.24%)700 (17.31%) Other Hispanic342 (8.63%)390 (9.64%) Non-Hispanic White1740 (43.92%)1709 (42.26%) Non-Hispanic Black835 (21.08%)841 (20.80%) Other Race441 (11.13%)404 (9.99%)Education level< 0.001 Less than high school901 (22.74%)1034 (25.57%) High school904 (22.82%)997 (24.65%) Higher than high school2157 (54.44%)2013 (49.78%)Marital status< 0.001 Married/living with partner2257 (56.97%)2693 (66.59%) Widowed/divorced/separated935 (23.60%)822 (20.33%) Never married770 (19.43%)529 (13.08%)PIR0.656 < 1.31133 (28.60%)1132 (27.99%) 1.3–3.51585 (40.01%)1658 (41.00%) >=3.51244 (31.40%)1254 (31.01%)BMI (kg/m²)< 0.001 < 251471 (37.13%)853 (21.09%) 25–301345 (33.95%)1317 (32.57%) >=301146 (28.92%)1874 (46.34%)WC (cm)95.82 ± 15.64103.20 ± 16.15< 0.001Hypertension< 0.001 Yes1218 (30.74%)1627 (40.23%) No2744 (69.26%)2417 (59.77%)Diabetes< 0.001 Yes411 (10.37%)636 (15.73%) No3551 (89.63%)3408 (84.27%)Smoking status< 0.001 Yes1647 (41.57%)2034 (50.30%) No2315 (58.43%)2010 (49.70%) Basic characteristics of participants with or without OSA Dataset acquisition and DELMRG identification Gene expression datasets related to OSA were obtained from the GEO database. The training dataset, GSE135917, included subcutaneous adipose tissue samples from 8 control subjects and 34 patients with OSA, while the testing dataset, GSE38792, consisted of visceral adipose tissue samples from 8 controls and 10 patients with OSA. Both datasets were generated using the GPL6244 platform. DEGs between OSA and control samples in the GSE135917 dataset were identified using the GEO2R tool, with an adjusted p value < 0.05 and |log₂ fold change| > 0.5 considered statistically significant. A total of 348 LMRGs were obtained from the Molecular Signatures Database (MSigDB). Genes overlapping between DEGs and LMRGs were defined as DELMRGs. The overall workflow of the study is summarized in Fig. 1. Analysis of functional enrichment in DELMRGs Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed using the Database for Annotation, Visualization, and Integrated Discovery (DAVID) to explore the potential biological functions and signaling pathways associated with the DELMRGs. GO analysis encompasses three categories: biological process (BP), cellular component (CC), and molecular function (MF). A false discovery rate (FDR) < 0.05 was considered statistically significant. Feature selection from DELMRGs via three machine learning (ML) methods Machine learning (ML) methods have been widely applied for feature selection in high-dimensional transcriptomic data [19]. To improve the robustness and accuracy of hub feature identification, three complementary ML algorithms were employed: least absolute shrinkage and selection operator (LASSO) regression, random forest (RF), and the Boruta algorithm. LASSO regression was used to identify key marker genes by applying an L1 penalty to the regression coefficients, with the optimal penalty parameter (lambda) determined through 10-fold cross-validation. Genes with non-zero coefficients were retained for further analysis, allowing effective feature selection and dimensionality reduction while mitigating multicollinearity [20]. RF, an ensemble learning method based on multiple decision trees, was applied to rank gene importance and identify informative features in the high-dimensional dataset. Owing to its robustness and resistance to overfitting, RF is well suited for gene selection in microarray-based studies [20, 21]. The Boruta algorithm, an RF-based wrapper feature selection method, was used to further assess the relevance of candidate genes by comparing their importance with that of randomly permuted shadow features. Features were classified as “Confirmed,” “Tentative,” or “Rejected” based on iterative importance testing [22, 23]. In this study, the Boruta algorithm was executed with 1,000 permutations, and a Bonferroni-type correction was applied to the shadow feature threshold. Only features classified as “Confirmed” were retained. To reduce method-specific bias and enhance the reliability of feature selection, a conservative consensus strategy was adopted, whereby only genes identified by all three algorithms were considered hub genes. This approach minimizes noise and effectively addresses overfitting concerns in our small-sample study. All analyses were performed in R using the “glmnet” and “randomForest” packages, and visualization was conducted using the “ggplot2” package. Protein‒protein interaction (PPI) network construction and identification of core genes To further explore the functional interactions among DELMRGs and gain insight into the molecular mechanisms underlying disease onset and progression, a protein–protein interaction (PPI) network was constructed using the STRING database. The resulting interaction network was imported into Cytoscape software (version 3.9.1) for visualization and subsequent analysis. Core genes were identified using two Cytoscape plug-ins, MCODE and CytoHubba, to complement the results obtained from the ML algorithms. MCODE, which identifies densely connected regions within a network based on topological features, was applied with default parameters to detect key functional modules. In parallel, CytoHubba was used to rank genes according to four algorithms: Maximal Clique Centrality (MCC), Maximum Neighborhood Component (MNC), Edge Percolated Component (EPC), and Degree. The top 10 genes identified by each method were considered high-confidence candidates. Finally, genes that overlapped between the highest-scoring CytoHubba results and the most significant MCODE module were integrated to identify candidate core genes for further analysis. Expression levels and diagnostic performance of the candidate genes The GSE135917 dataset was used for initial screening, while GSE38792 served as an internal testing dataset. The two datasets were analyzed separately in a sequential manner to evaluate the diagnostic performance of candidate genes and to identify genes with consistent discriminatory ability across datasets. No independent external validation cohort was included in the present study. Receiver operating characteristic (ROC) curve analysis was performed to assess the discriminatory ability of each candidate gene. Genes demonstrating robust and reproducible diagnostic performance in both datasets were selected for further analyses. Logistic regression analysis was subsequently applied to construct a multigene diagnostic model by integrating the selected genes, and the diagnostic performance of the combined model was evaluated using ROC analysis. All statistical analyses were conducted using SPSS software (version 22.0). For single-gene analyses, genes with an area under the ROC curve (AUC) greater than 0.7 were considered to exhibit acceptable discriminatory ability. Box plots were generated to visualize gene expression levels and to compare differences between the OSA and control groups. Gene set enrichment analysis (GSEA) To investigate the biological significance of DELMRGs in OSA, samples in the GSE135917 dataset were stratified into high- and low-expression groups for each key gene based on the median expression value. A ranked gene list was generated according to log₂ fold change values derived from differential expression analysis. Gene set enrichment analysis (GSEA) was subsequently performed using KEGG pathway gene sets obtained from the MSigDB. Single-gene GSEA was conducted using the R package clusterProfiler to identify biological pathways significantly enriched in association with the expression levels of key genes. Regulatory network construction NetworkAnalyst [24] (https://www.networkanalyst.ca/) is a web-based platform for integrative gene expression and network analysis. All identified DELMRGs were uploaded into NetworkAnalyst to construct transcription factor (TF)–gene and gene-microRNA (miRNA) interaction networks, aiming to explore the regulatory mechanisms of key DELMRGs at both transcriptional and post-transcriptional levels. For TF–gene network construction, TF–target interaction data were obtained from the ChEA3 database [25], which integrates literature-curated ChIP-X experimental evidence. The gene-miRNA interaction network was generated using the TarBase database [26], which contains experimentally validated miRNA–target interactions. All interaction networks were subsequently exported to Cytoscape for visualization and further analysis. Statistical analysis Demographic and clinical characteristics of the study participants are presented as means ± standard deviations (SDs) for continuous variables and as numbers (percentages) for categorical variables. Group differences were assessed using chi-square tests for categorical variables and the Kruskal–Wallis rank-sum test for continuous variables, as appropriate. Fisher’s exact test was applied when the expected cell count was less than 10. Multicollinearity among covariates in the fully adjusted model was assessed using variance inflation factors (VIFs). A VIF value greater than 10 was considered indicative of significant multicollinearity, and such variables were considered for exclusion from the model to ensure the stability of parameter estimates [6]. The primary analysis was conducted using a complete-case analysis, excluding observations with missing data for any variable included in the corresponding model. To examine the associations between lipid levels and OSA, multiple logistic regression analyses were performed. Model 1 was an unadjusted model. Model 2 was adjusted for age, sex, and race. Model 3 was fully adjusted for potential confounding variables, including education level, marital status, PIR, BMI, WC, smoking status, hypertension status, and diabetes status. The results are reported as odds ratios (ORs) with corresponding 95% confidence intervals (CIs). To explore potential nonlinear associations between continuous lipid concentrations and OSA risk, smooth curve fitting was performed using generalized additive models (GAMs), with smoothing parameters selected by restricted maximum likelihood. To further identify potential inflection points (threshold effects), a two-piecewise linear regression model was fitted within the logistic regression framework. The threshold value was determined using a recursive algorithm that maximized the model likelihood, and statistical significance was assessed using a likelihood ratio test comparing the piecewise model with a linear model. All statistical analyses were performed using RStudio and EmpowerStats software. Figures were generated using GraphPad Prism (version 8.0) and the online tool SRplot (https://www.bioinformatics.com.cn/srplot). A two-sided p value < 0.05 was considered statistically significant. Results Baseline characteristics A total of 8006 participants were included in this study, all of whom had complete demographic and follow-up data and met the predefined inclusion and exclusion criteria (Fig. 1). Participants were stratified into two groups based on their self-reported OSA status: 4044 with OSA and 3962 without OSA. Participants with OSA had significantly higher TC (4.99 ± 1.08 vs. 4.95 ± 1.09, p < 0.05), TG (1.42 ± 0.77 vs. 1.26 ± 0.74, p < 0.001), and LDL-C (2.98 ± 0.95 vs. 2.89 ± 0.93, p < 0.001), and lower HDL-C (1.36 ± 0.40 vs. 1.48 ± 0.44, p < 0.001) levels compared to those without OSA. OSA patients were more likely to be middle-aged or older adults (age ≥ 40), with a higher percentage of males and non-Hispanic Whites. The two groups also differed significantly in terms of educational level, marital status, BMI, and WC. Additionally, a history of smoking, diabetes, and high blood pressure was more common in those with OSA. However, no significant difference in PIR was observed, despite other significant differences between the two groups (Table 1). Continuous variables were presented as mean ± standard deviation, and categorical variables were expressed as n (%). For statistical analysis, Kruskal-Wallis tests were used for continuous variables, and chi-square tests for categorical variables. TC: total cholesterol; TG: triglycerides; LDL-C: low-density lipoprotein cholesterol; HDL-C: high-density lipoprotein cholesterol; PIR: family poverty income ratio; BMI: body mass index; WC: waist circumference. Connections between blood lipid levels and OSA To ensure model stability, multicollinearity among the variables was assessed. Except for TC, the VIF values for the other variables were below 10 (Supplementary Table S1). Variables with VIF values greater than 10 were excluded from the analysis. To further investigate the relationship between OSA and blood lipid levels (TG, LDL-C, and HDL-C), three multiple regression models were constructed (Fig. 2 and Supplementary Table S2). Analysis of continuous variables revealed that higher TG and LDL-C levels were positively correlated with OSA incidence (OR = 1.32, p < 0.0001; OR = 1.10, p < 0.0001, respectively). Conversely, higher HDL-C levels were significantly negatively correlated with OSA (OR = 0.50, p < 0.0001) in Model 1 (unadjusted). These findings indicate that for every one-unit increase in TG and LDL-C levels, the risk of OSA increased by 32% and 10%, respectively, while a one-unit increase in HDL-C levels was associated with a 50% reduction in OSA risk. These significant relationships persisted after adjusting for covariates in Models 2 and 3. Forest plot showing the results of association between blood lipid levels and OSA. The orange frame represents Model 1 of the logistic regression analysis, the blue frame represents Model 2, and the pink frame represents Model 3. The green diamonds indicate the association between TG and OSA, the red squares indicate the association between LDL-C and OSA, and the purple circles indicate the association between HDL-C and OSA. Model 1: Non-adjusted; Model 2: Adjusted for age, sex, race; Model 3: Adjusted for age, sex, race, education level, marital status, PIR, BMI, WC, smoking status, diabetes and hypertension. TG: triglycerides; LDL-C: low-density lipoprotein cholesterol; HDL-C: high-density lipoprotein cholesterol; OR: odds ratios; CI: confidence intervals In the sensitivity analysis, TG, LDL-C, and HDL-C were categorized into tertiles. Individuals in the highest tertile (T3) of TG levels had a significantly higher risk of OSA compared to those in the lowest tertile (T1) in Model 1 (OR = 1.75, p < 0.0001). The risk of OSA increased progressively with rising TG levels (p for trend < 0.0001), and similar trends were observed in Models 2 and 3 (p for trend < 0.0001 and p = 0.0097, respectively). Higher LDL-C levels were also associated with a greater risk of OSA, with significant trends in all models (p for trend < 0.0001 in Model 1, p = 0.0010 in Model 2, and p = 0.0063 in Model 3). In contrast, individuals in the highest tertile (T3) of HDL-C levels were 49% less likely to develop OSA compared to those in the lowest tertile (T1) in Model 1 (OR = 0.49, p < 0.0001). This phenomenon remains significant in Models 2 and 3 (p for trend < 0.0001 for both models), and the trend was consistent across tertiles (p for trend < 0.0001). To visualize the relationship between TG, LDL-C, HDL-C, and OSA, smooth curve fitting based on Model 3 was performed. As shown in Fig. 3, a nonlinear relationship was observed between TG and OSA, while linear positive and negative correlations were found for LDL-C and HDL-C, respectively. To further explore the connection between OSA and TG, a threshold effect analysis was conducted (Table 2). The results revealed an inverted U-shaped association, with the inflection point identified at 1.32 (log-likelihood ratio = 0.016). For TG levels below 1.32, every unit increase in TG was associated with a 32% increased risk of OSA. However, once TG levels exceeded this threshold, the association weakened and became statistically insignificant, suggesting that further increases in TG do not significantly increase the risk of OSA. Smooth curve fitting based on generalized additive models (GAMs) showing the nonlinear association between (A)TG, (B) LDL-C, (C) HDL-C and OSA. The red solid line represents the log (OR) of OSA, blue shaded region represents 95% CI. TG: triglyceride; LDL-C: low-density lipoprotein cholesterol; HDL-C: high-density lipoprotein cholesterol; OSA: obstructive sleep apnea; OR: odds ratios; CI: confidence intervals Table 2Threshold effect analysis of TG on OSA using a two-piecewise logistic regression modelThreshold effect analysis:OSA OR (95%CI) p-valueInflection point of TG (K)1.32 < K slope1.32 (1.10, 1.58) 0.0023 > K slope0.99 (0.90, 1.09) 0.7968Log-likelihood ratio test0.016 Threshold effect analysis of TG on OSA using a two-piecewise logistic regression model Identification and functional enrichment analysis of DELMRGs Data from the OSA database (GSE135917) revealed a total of 2,329 DEGs based on the specified criteria. Among these, 337 were upregulated, while 1,992 were downregulated (Fig. 4A). The DEGs were then intersected with the LMRGs using a Venn diagram tool, resulting in 34 overlapping DELMRGs, including 26 downregulated and 8 upregulated genes (Fig. 4B and C). Identification of DELMRGs. (A) Volcano plot of DEGs in GSE135917 dataset. (B) Venn diagram of DEGs and LMRGs. (C) The heat map of 34 DELMRGs between control and OSA samples. OSA: obstructive sleep apnea; DELMRGs: differentially expressed lipid metabolism-related genes; DEGs: differentially expressed genes; LMRGs: lipid metabolism-related genes Functional enrichment analysis To investigate the biological functions and associated signaling pathways of the 34 DELMRGs, GO and KEGG pathway enrichment analyses were performed. As summarized in Supplementary Table S3, 33 significantly enriched GO terms and six KEGG pathways were identified. In the BP category, the DELMRGs were predominantly associated with lipid-related processes, including low-density lipoprotein particle remodeling, lipid hydroxylation, and long-chain fatty acid metabolism. In the CC category, these genes were mainly linked to the endoplasmic reticulum membrane, mediator complex, and core mediator complex. In the MF category, enrichment was most prominent in heme binding, oxidoreductase activity, and iron ion binding. The top three enriched terms in each GO category (BP, CC, and MF) are depicted in Fig. 5A. KEGG pathway analysis (Fig. 5B) revealed significant enrichment of the DELMRGs in several metabolic pathways, most notably the PPAR signaling pathway, fat digestion and absorption, and ether lipid metabolism. Function enrichment analysis. (A) GO analysis results of the top three BP, CC and MF terms. (B) KEGG pathway analysis of the 34 DELMRGs between control and OSA samples. The size of the nodes reflects the number of genes enriched in that term, while the color indicates different pathways. GO: gene oncology; BP: biological process; CC: cell component; MF: molecular function; KEGG: Kyoto Encyclopedia of Genes and Genomes; DELMRGs: differentially expressed lipid metabolism-related genes; OSA: obstructive sleep apnea Feature selection via ML methods LASSO regression analysis with lambda.min = 0.039 identified eight feature-selected genes from the 34 DELMRGs: CYP1A1, CYP3A4, FABP7, FAR1, MED18, SLC27A5, SOS1, and UGT8 (Fig. 6A-B). Next, the variable importance in the RF model was evaluated using the mean decrease in Gini index, which led to the identification of the top ten characteristic genes among the 34 DELMRGs: CYP4A22, UGT8, FAR1, CYP24A1, MMP1, SLC27A5, MED18, APOB, HSD17B13, and CYP3A4 (Fig. 6C). In the Boruta analysis, the evolution of Z-scores is shown in Fig. 6D. A total of eight genes were selected by the Boruta algorithm from the 34 DELMRGs: CCNC, CYP24A1, CYP4A22, FAR1, MED18, MMP1, SLC27A5, and UGT8. Finally, Venn diagram analysis identified four overlapping genes (FAR1, MED18, SLC27A5, and UGT8) as the critical feature genes (Fig. 6E). Identification of feature genes for OSA based on ML algorithms. (A) Based on a logarithmic sequence of the regularization parameter (lambda), a coefficient profile was constructed. Non-zero coefficients obtained at the optimal lambda indicate key genes, which were defined as characteristic genes exerting a significant influence on the disease under investigation. (B) The optimal parameter (lambda) in the LASSO model was determined using 10-fold cross-validation, applying the minimum criterion plus one standard error (indicated by the right vertical line). (C) Variable importance ordered by the gini index of a mean decrease in RF. The bands on the X-axis indicate the significance of the gene in reducing the Gini metric according to the average value. The higher the score, the greater the importance. (D) The results of Boruta algorithm. In the figure, Tentative is yellow, Rejected is red, Accepted is green, and Shadow is blue. (E) The Venn diagram illustrating the critical feature genes that are common to LASSO, RF and Boruta. OSA: obstructive sleep apnea; ML: machine learning; LASSO: least absolute shrinkage and selection operator; RF: random forest Core DELMRGs identified from the PPI network Using the STRING database, we constructed the PPI network for the 34 DELMRGs, and the resulting interactions were visualized with Cytoscape (Fig. 7A). To identify significant gene clusters, we applied the MCODE plugin, which detected a notable module containing 11 nodes and 25 edges as a core cluster. This core cluster included genes such as CYP3A4, CCNC, MED18, PTGS2, CYP2E1, MED30, CYP1A1, MED27, CYP4A22, PPARGC1A, and PLA2G10 (Fig. 7B). The top 10 genes were then identified using the MCC, EPC, MNC, and Degree algorithms (Fig. 7C-F). By integrating these results, six core DELMRGs were ultimately selected: CYP3A4, PTGS2, CYP2E1, CYP1A1, CYP4A22, and PLA2G10 (Fig. 7G). These core genes, along with the hub features identified in the previous ML analysis, were further subjected to ROC analysis. Screening core genes by PPI Network. (A) PPI network of DELMRGs. In the figure, the nodes of the central circle represent the core genes. (B) A critical cluster with 11 genes was further chosen as candidate genes by MCODE. (C-F) The top 10 core genes were sequentially calculated by algorithms through the CytoHubba. (G) Venn diagram of interaction. DELMRGs: differentially expressed lipid metabolism-related genes; PPI: protein-protein interaction ROC analysis In the training dataset, ROC curve analysis was conducted to evaluate the diagnostic performance of each selected gene, and the corresponding AUC values are summarized in Supplementary Table S4. All ten DELMRGs exhibited strong discriminatory ability, with AUC values exceeding 0.80. When further evaluated in the testing dataset, only three genes—CYP3A4, CYP4A22, and MED18—consistently demonstrated statistically significant discriminatory ability (p < 0.05) (Supplementary Table S4). Therefore, subsequent analyses focused on these three genes. Comparative expression analysis revealed that the transcript levels of CYP3A4, CYP4A22, and MED18 were significantly lower in the OSA group compared to the control group in the training dataset, and a similar expression pattern was observed in the testing dataset (Fig. 8A). The expression levels and ROC analysis of the key DELMRGs. (A) CYP3A4, CYP4A22, and MED18 expression levels in two databases, with yellow boxplots representing OSA and green boxplots representing controls. Mann-Whitney test with p < 0.05 was used to determine statistical significance. *p < 0.05; ** p < 0.01; **** p < 0.0001. (B) The ROC curves showing AUC values of CYP3A4, CYP4A22, and MED in two databases. ROC: receiver operating characteristic; DELMRGs: differentially expressed lipid metabolism-related genes Among these three genes, CYP4A22 exhibited the strongest discriminative ability between OSA patients and controls, with an AUC of 0.945, sensitivity of 87.5%, specificity of 97.1%, and a Youden index of 0.846 in the training dataset. In the testing dataset, CYP4A22 maintained strong discriminatory power, yielding an AUC of 0.838, sensitivity of 87.5%, specificity of 80.0%, and a Youden index of 0.675 (Fig. 8B). To further assess whether a combination of genes could improve diagnostic performance, a multigene diagnostic model incorporating CYP3A4, CYP4A22, and MED18 was constructed using logistic regression. As shown in Supplementary Table S4, in the GSE135917 dataset, the three-gene model achieved the highest diagnostic accuracy, with an AUC of 0.978, sensitivity of 88.2%, specificity of 100%, and a Youden index of 0.882. Similarly, in the GSE38792 dataset, the combined model demonstrated robust discriminatory ability, yielding an AUC of 0.925, sensitivity of 100.0%, specificity of 87.5%, and a Youden index of 0.875. Overall, these results suggest that the combined three-gene signature provides improved diagnostic performance compared to individual genes. Potential mechanisms of key genes in OSA To investigate the potential involvement of CYP3A4, CYP4A22, and MED18 in the pathophysiology of OSA, single-gene GSEA-KEGG analysis was conducted for each gene. The six most significantly enriched pathways, based on their p-values, were presented. CYP3A4 was negatively correlated with pathways related to lysosomes and glutathione metabolism, among others (Fig. 9A). CYP4A22 was negatively correlated with lysosomes and glutathione metabolism, but positively correlated with the MAPK signaling pathway (Fig. 9B). MED18 was negatively correlated with pathways associated with type I diabetes mellitus, allograft rejection, and asthma. Conversely, it was positively linked with pathways related to cancer, ribosomes, and the MAPK signaling pathway (Fig. 9C). Single-gene GSEA of key DELMRGs. (A) Correlation of CYP3A4 and with common significantly enriched pathways. (B) Correlation of CYP4A22 with common significantly enriched pathways. (C) Correlation of MED18 with common significantly enriched pathways. GSEA: gene set enrichment analysis; DELMRGs: differentially expressed lipid metabolism-related genes; OSA: obstructive sleep apnea Construction of the related regulatory network Regulatory networks of TFs and miRNAs associated with three DELMRGs (CYP3A4, CYP4A22, and MED18) were constructed using the NetworkAnalyst platform. To identify reliable candidates, a stringent network topology-based filtering strategy was applied to all nodes in the network. First, nodes with a degree greater than 2 were retained to eliminate peripheral nodes and potentially spurious interactions [27]. Second, the retained nodes were ranked according to betweenness centrality, which quantifies the importance of a node as a “shortest-path bridge” between different parts of the network, thereby identifying key hubs that control global information flow [28]. In the TF–DELMRG regulatory network, 65 TFs were identified, with TP53 emerging as the most pivotal regulator (Fig. 10A). Similarly, analysis of the miRNA–DELMRG regulatory network identified 54 miRNAs, with hsa-miR-26a-5p recognized as the most pivotal miRNA (Fig. 10B). The regulatory networks of key DELMRGs in OSA. (A) The network presents the TF-gene coregulatory network. (B) The network presents the gene-miRNA coregulatory network. Red diamonds represent key DELMRGs, green circles denote TFs, yellow circles indicate pivotal TF, purple arrows represent miRNAs, and blue arrows indicate pivotal miRNA. DELMRGs: differentially expressed lipid metabolism-related genes; OSA: obstructive sleep apnea; miRNAs: micro-RNA; TFs: transcription factors Discussion OSA is a common breathing disorder associated with dysfunctions in various bodily systems. The increasing global prevalence of OSA, coupled with persistent challenges in its timely diagnosis and effective clinical management, has made it an increasingly important focus of medical research. The precise fundamental pathophysiological mechanisms of OSA remain incompletely understood, although numerous investigations have demonstrated a significant association between lipid metabolism and OSA. This study aimed to clarify the intricate link between OSA and lipid metabolism and explore how lipid-related genes may contribute to this disorder. Integrating clinical data with transcriptomic analysis revealed evidence of lipid metabolic disturbances in OSA patients. Moreover, CYP3A4, CYP4A22, and MED18 were identified as key DELMRGs associated with OSA. Numerous studies have demonstrated a clear association between OSA and dyslipidemia, characterized by reduced HDL-C levels and elevated TG, and LDL-C concentrations [29]. These findings are consistent with the results of the present study. Our analysis revealed a nonlinear association between TG levels and OSA risk, following an inverted U-shaped pattern. Specifically, TG levels were positively associated with the OR of OSA at low to moderate concentrations. However, this association reached a plateau once TG levels exceeded a threshold value of 1.32, indicating the presence of a significant threshold effect. This observation aligns with emerging evidence supporting nonlinear relationships between newly proposed lipid metircs and OSA [12]. Obesity is a well-established risk factor for both the onset and progression of OSA. The strong positive association observed below the TG threshold may reflect the cumulative lipotoxic effects of elevated TG levels, which can exacerbate systemic inflammation, oxidative stress, and insulin resistance. These pathophysiological processes may impair upper airway neuromuscular function, thereby increasing susceptibility to OSA [4, 12]. In contrast, the plateau in risk at higher TG concentrations may indicate the saturation of inflammatory responses associated with extreme dyslipidemia [30], or the activation of compensatory mechanisms that mitigate further damage caused by prolonged hyperlipidemia. These proposed explanations remain hypothetical and require further experimental investigation. To obtain a better understanding of the underlying processes of OSA, the transcriptome data were further investigated. And the expression profiles of 34 DELMRGs across all samples were visualized using a heatmap. The heatmap revealed that a subset of OSA samples clustered more closely with controls than with other OSA samples, indicating either notable heterogeneity within the OSA group or possibly a limited disease-related signal in these particular individuals. An increasing body of evidence supports significant heterogeneity even among patients with comparable disease severity [31]. Our findings further suggest that lipid metabolism–related transcriptomic signatures vary substantially among OSA patients, which may be attributable to differences in clinical phenotypes, disease severity, or underlying metabolic subtypes. Further exploration of these distinct expression patterns could facilitate the stratification of OSA into molecular subgroups with specific biological characteristics and clinical relevance. KEGG pathway analysis indicated that the identified genes are closely associated with the PPAR signaling pathway. Dysregulation of this pathway has been implicated in multiple diseases, including lung adenocarcinoma [32], hepatocellular carcinoma [33], breast cancer [34], and OSA [35]. PPARs are nuclear receptor transcription factors and comprise three subtypes: PPARα, PPARβ, and PPARγ. Among them, PPARγ plays a key role in metabolic regulation by promoting glucose uptake in adipocytes, facilitating triglyceride storage in subcutaneous adipose tissue, and improving insulin sensitivity under physiological conditions. OSA is frequently accompanied by metabolic dysfunction, particularly insulin resistance and related abnormalities such as reduced HDL-C levels and hypertriglyceridemia. Consistent with this, previous studies have reported decreased PPARγ expression in patients with OSA [36]. Our findings suggest that the identified DELMRGs may be associated with metabolic disturbances in OSA, possibly through involvement in the PPAR signaling pathway. Nevertheless, this association remains exploratory, and further mechanistic studies are required to clarify the precise role of these genes in OSA-related insulin resistance. By integrating PPI network analysis, LM approaches, and ROC analysis, key genes were systematically identified. An important finding was that CYP3A4, CYP4A22, and MED18 consistently exhibited high diagnostic performance across both the training and testing datasets. To further improve classification accuracy beyond single-gene ROC analyses, these three genes were combined into multigene panels, and their diagnostic performance was evaluated using logistic regression–based ROC analysis. Ultimately, the three-gene panel comprising CYP3A4, CYP4A22, and MED18 demonstrated superior discriminatory ability compared with any individual gene alone. However, the limited sample size of both the training and testing datasets warrants caution in interpreting the high AUC values, which should be validated in larger, independent cohorts. Building upon these findings, the potential biological relevance of the three key DELMRGs in OSA was further explored using GSEA and regulatory network analyses. GSEA was conducted on gene lists ranked by their correlation strength with the key DELMRGs to investigate potential functional associations. These results demonstrated that CYP3A4 and CYP4A22 are associated with the lysosomal pathway and glutathione metabolism. Similarly, Zhang et al. reported that the lysosome pathway, the most enriched pathway, participated in OSA [37]. Emerging evidence indicates that OSA induces multiorgan damage, affecting the myocardium [38], liver [39], kidneys [40] and so on, potentially through dysregulation of lysosome-dependent autophagic processes [41]. Autophagy has organ-specific effects on OSA, exhibiting dual protective and detrimental effects. The latest research indicates that in OSA, acetyl-CoA carboxylase -driven lipogenesis disrupts lysosomal iron or lipid homeostasis, triggering ferroptosis through accumulated lipid peroxides and restoring lysosomal function via the NRF2-TFEB axis, reversing both iron and lipid dysregulation. These findings suggest that lysosomes are essential for the onset and progression of OSA. Glutathione (GSH), a critical nonenzymatic antioxidant in mammalian cells, directly neutralizes free radicals and pro-oxidants to protect against oxidative damage [42]. A previous study reported that glutathione (GSH) levels were reduced in patients with OSA compared with controls, indicating the presence of oxidative stress in OSA [43]. Furthermore, studies have reported that marked elevation in serum and hepatic lipid peroxidation following CIH strongly suggests increased generation of reactive oxygen species within the liver [44]. Our findings of associations with glutathione metabolism pathways are consistent with the literature, but causative roles have not been proven. In addition, CYP4A22 and MED18 were found to be associated with the MAPK signaling pathway. One study revealed significant enrichment of the MAPK signaling pathway in a chronic IH rat model [45], proposing an association between MAPK signaling and OSA progression. The MAPK pathway coordinates oxidative stress and inflammatory responses [46], and its inhibition can alleviate IH-triggered inflammation [47]. Taken together, these enriched pathways suggest that the development of OSA may be correlated with CYP3A4, CYP4A22, and MED18 expression patterns, but mechanistic conclusions require experimental validation. Recent studies have highlighted the importance of miRNA–mRNA–TF regulatory networks in disease mechanisms [48]. The miRNAs primarily regulate gene expression at the post-transcriptional level, whereas TFs control transcriptional activation. Both regulatory layers are essential for normal biological function and are closely involved in disease processes. Accordingly, we constructed TF–gene and miRNA–gene regulatory networks to better characterize the molecular basis of the disease. TP53 is a well-established tumor suppressor that is frequently inactivated in human cancers and exerts protective effects against tumorigenesis by regulating genes involved in cell cycle arrest, cellular senescence, apoptosis, DNA repair, and cell growth [49]. In the present study, strong correlations were observed between TP53 and CYP3A4 as well as MED18, suggesting potential regulatory interactions among these genes in OSA. Previous studies have demonstrated that miR-26a-5p plays a critical role in endothelial dysfunction and induces cell cycle arrest under hypoxic conditions, thereby impairing angiogenesis and cell proliferation [50]. Notably, miR-26a-5p expression was significantly reduced in a mouse model of obstructive sleep apnea–hypopnea syndrome (OSAHS) with arterial hypertension compared with OSAHS alone [50]. Our gene–miRNA network analysis revealed that hsa-miR-26a-5p regulates both CYP3A4 and CYP4A22. Collectively, these results suggest a potential mechanism through which miR-26a-5p may contribute to vascular dysregulation in OSA. However, further experimental studies are required to determine the role of miR-26a-5p. The cytochrome P450 (CYP) superfamily comprises a large group of enzymes that play essential roles in oxidation–reduction reactions and xenobiotic metabolism [51]. Among these enzymes, CYP3A4 is highly expressed in the liver and gastrointestinal tract and is responsible for the metabolism of several lipid-lowering agents, including simvastatin [52]. In addition to its metabolic function, CYP3A4 has been reported to be responsive to hypoxic regulation. Previous studies have shown that hypoxia stabilizes hypoxia-inducible factor-1α (HIF-1α), which can transcriptionally repress CYP3A4 expression in hepatocyte and tumor cell model [53, 54]. In the context of OSA, chronic IH may activate HIF-1α–dependent signaling, which could be associated with the observed downregulation of CYP3A4 and other lipid-metabolizing enzymes. From this perspective, altered CYP3A4 expression is more plausibly interpreted as a downstream molecular response to hypoxic stress rather than a primary driver of OSA. However, the conclusions drawn from our study remain hypothetical, and further experimental work is required to validate them. In contrast, the roles of CYP4A22 and MED18 in OSA remain less clear. Previous studies have reported that members of the cytochrome P450 4A (CYP4A) family are involved in fatty acid metabolism and have been associated with lipid peroxidation, oxidative stress, mitochondrial dysfunction, and hepatic fibrosis [55]. In addition, enzymes from the CYP4A and CYP4F families catalyze the ω-hydroxylation of arachidonic acid to generate 20-hydroxyeicosatetraenoic acid, a bioactive lipid that has been linked to metabolic disorders such as hypertension, obesity, and metabolic syndrome [56]. Although we observed reduced CYP4A22 expression in OSA patients in our cohort, direct evidence linking it to hypoxia signaling or OSA pathogenesis is still lacking. Similarly, while MED18 participates in RNA processing in yeast [57], its function in mammalian systems, especially in the context of hypoxia or metabolic disease, remains uncharacterized. Their expression changes are correlative observations, and further experimental work is needed to determine whether these genes contribute to OSA progression. This study has several noteworthy strengths. First, this study is based on NHANES data from 2005 to 2008 and 2015–2018, yielding a substantial sample size that strengthens the robustness of the observed association between OSA and dyslipidemia. Second, comprehensive bioinformatics analyses of transcriptomic datasets identified three key genes (CYP3A4, CYP4A22, and MED18) closely associated with lipid metabolism and OSA. Moreover, this study provides mechanistic insights into the complex relationship between lipid metabolic pathways and OSA pathogenesis, supported by multiple analytical approaches, including GSEA and regulatory network analysis. The identified genes offer biologically plausible targets for future functional and experimental studies. Importantly, integrating NHANES and GEO datasets links population-level epidemiological associations with molecular-level transcriptomic evidence, enhancing both the clinical relevance and biological interpretability of the findings. Despite these strengths, this study has several limitations. First, because polysomnographic data were not available in the NHANES dataset, OSA was identified through questionnaire-based measures, which may have introduced misclassification bias. Specifically, OSA was defined according to three core symptoms: snoring, witnessed apneas, and daytime sleepiness. Although this symptom-based definition is widely applied in population-based studies and captures a substantial proportion of individuals with OSA, it reflects a simplified screening approach rather than a clinical diagnosis. Moreover, the absence of laboratory-confirmed sleep study data may have resulted in the misclassification or exclusion of individuals with asymptomatic OSA, potentially attenuating the observed exposure–outcome associations and affecting the robustness of the epidemiological findings. Second, the study population was predominantly composed of American participants, which may limit the generalizability of the findings to other ethnic or geographic populations. Given known racial differences in lipid profiles and OSA prevalence, further validation in diverse populations is warranted to enhance the external validity of the results. Additionally, NHANES sampling weights were not applied and the findings should be interpreted as associations within the study sample rather than nationally representative estimates. Third, the relatively small sample size of the GSE135917 dataset may have impacted the robustness of feature selection. Future studies incorporating larger transcriptomic cohorts and more comprehensive clinical phenotyping of OSA are needed to strengthen the reliability and reproducibility of these findings. Finally, due to practical constraints in clinical sample collection, experimental validation of key lipid-related genes was not performed, which precludes definitive causal inference. Therefore, the current findings should be interpreted as exploratory and require confirmation in future mechanistic and clinical studies. Conclusion This study investigated the association between blood lipid levels and OSA using data from the NHANES database. Through integrated bioinformatics analyses, CYP3A4, CYP4A22, and MED18 were identified as genes with potential diagnostic value in both the training and testing datasets. However, given the limited size and non-independence of the validation dataset, these findings should be interpreted with caution, and further validation in independent, larger cohorts is required to confirm their diagnostic utility. This study underscores the value of combining population-level data with bioinformatics approaches to gain insight into the complex relationship between metabolic abnormalities and OSA. Overall, our findings provide preliminary mechanistic insights into the molecular links between metabolic dysregulation and OSA pathophysiology, while highlighting the need for future studies using larger, independent cohorts to validate and extend these observations. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Acknowledgements We thank all the participants involved in this study. We thank NHANES database, GEO database, MSigDB database, DAVID database, STRING database and Networkanalyst database and their contributors for the valuable public datasets used in this study. We also thank the Beckman Coulter DxAI platform (https://www.xsmartanalysis.com/beckman/login/), the web of https://www.bioinformatics.com.cn/srplot and www.xiantao.love for their contribution to the analyzing and drawing in this study. Author contributions Study concept and design: CF and YZ. Acquisition of data: YL and SG. Analysis and interpretation of data: SC and JW. Results visualization: YL and SG. Drafting of the manuscript: SG and JW. Manuscript revising: CF. Funding This research was funded by the Startup Fund for scientific research, Fujian Medical University (Grant number: 2025QH1149). Data availability The data can be downloaded for free from the website: https://www.cdc.gov/ nchs/nhanes/index.htm and https://www.ncbi.nlm.nih.gov/geo/. Declarations Ethics approval and consent to participate The survey was approved by the National center for health statistics (NCHS) Research Ethics Review Board. All participants provided written informed consent. NCHS Research Ethics Review Board approval was waived for this analysis because of the publicly available and de-identified data. All methods were carried out in accordance with relevant guidelines and regulations (Declaration of Helsinki). Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yanhui Liu and Sisi Gong contributed equally to this work. Contributor Information Yiming Zeng, Email: zeng_yiming@fjmu.edu.cn. Chunmei Fan, Email: 13600718981@163.com. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The data can be downloaded for free from the website: https://www.cdc.gov/ nchs/nhanes/index.htm and https://www.ncbi.nlm.nih.gov/geo/.