This study was conducted as a systematic review and meta-analysis of RCTs, following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (25). The study protocol was registered in PROSPERO on 23 March 2025 (registration number: CRD420251053389) and was updated on 23 January 2026 to ensure methodological transparency.

The inclusion and exclusion criteria were strictly defined using the PICO framework to ensure clarity and rigor. Eligible studies met the following criteria: Population: Adults aged ≥ 18 years with confirmed OSA (AHI ≥ 5), regardless of severity or comorbidities. Intervention: Experimental group receiving CPAP combined with lifestyle interventions (multi-component, e.g., diet + exercise; single-component, e.g., diet alone). Comparison: Control group receiving non-CPAP treatments, usual care, or CPAP alone; subgroup stratification and sensitivity analysis by comparator type were conducted to address heterogeneity. Outcome: Primary outcome was absolute AHI change (events/h) for clinical interpretability, with MD also reported for cross-study comparison. Studies were excluded if they involved surgical, pharmacological, or unspecified primary interventions; non-RCT designs (e.g., case reports, preprints); uncontrolled confounding co-interventions; or lacked extractable AHI data.

A comprehensive literature search was conducted across multiple databases, including PubMed, Web of Science, Cochrane Library, and Embase. The search protocol was registered on 6 March 2025. To identify relevant studies, both Medical Subject Headings (MeSH) terms and free-text terms were used to capture key concepts related to the research topic. The search strategy included the following terms: (“obstructive sleep apnea” OR “sleep apnea, obstructive” OR “apneas obstructive”) AND (“continuous positive airway pressure” OR “CPAP”) AND (“lifestyle intervention” OR “lifestyle” OR “diet” OR “exercise” OR “physical activity”) AND (“apnea hypopnea index” OR “AHI”) AND (“Controlled Trial” OR “Randomised Controlled Trial” OR “RCT” OR “Clinical Trial”). See the Supplementary file. Additionally, the reference lists of eligible articles were manually reviewed to identify further studies.

All records were imported into EndNote for deduplication. Two independent reviewers screened titles and abstracts, followed by full-text reviews to determine eligibility. Discrepancies were resolved by consulting a third reviewer. Data were independently collected using a standardized form, and any disagreements were resolved through discussion to ensure accuracy and reliability. When data were missing or presented only graphically, authors were contacted to request the necessary information. If this contact attempt was unsuccessful and the data were available only in graphical form, relevant data were extracted using WebPlotDigitizer 4.1 (https://automeris.io/WebPlotDigitizer) (26). This software has been shown to have high reliability and validity (27). If we failed to obtain the missing information, the specific study was excluded from the analysis.

All meta-analyses were conducted in R using the meta, metafor, dmetar, and ggplot2 packages (28). Because AHI was consistently reported in identical units (events per hour) across the included studies, MD was used as the effect size to retain the original clinical units (29). The MD was calculated as the difference in mean AHI between the intervention and control groups. The magnitude of the MD was interpreted using the absolute reduction in AHI, in conjunction with established OSA severity classifications and their clinical implications for disease severity and symptom burden (12, 30).

The primary analyses were conducted using multilevel random-effects models via the rma.mv function, which accounted for correlations among multiple effect sizes from the same study (levels: sampling variance, within-study variance, and between-study variance) (31). Restricted maximum-likelihood (REML) estimation was used, and when heterogeneity was minimal (I² < 50%), results from fixed-effects models were also included. Statistical heterogeneity was evaluated using Cochran’s Q test (p < 0.10 considered significant) and the I² statistic, with values exceeding 50% indicating considerable heterogeneity (32). Publication bias was assessed using a funnel plot and Egger’s linear regression, and data with an asymmetric distribution were adjusted using the trim-and-fill technique (33). Outlier studies with strong influence on the model were identified using standardized residuals (|Z| > 2.5) or Cook’s distance (threshold > 3 times the mean) (34). To further verify the robustness of the results, the following sensitivity analyses were conducted: (1) leave-one-out analysis (Metainf function); (2) subgroup stratification tests to assess the sources of heterogeneity using categorical variables; (3) meta-regression analysis using REML to quantify the association between potential moderators and effect size (32).

The risk of bias in the included studies was assessed using the Cochrane Risk of Bias tool (RoB 2.0) for RCTs (35). This assessment emphasized several key factors: randomization, allocation concealment, blinding, incomplete data, and selective reporting. Two reviewers conducted the assessments independently, with a third reviewer available to resolve disagreements. Overall confidence in the evidence for primary outcomes (AHI) and all subgroup analyses was assessed using the GRADE framework, with specific criteria for downgrading. Findings were summarized in tables created with the GRADEpro GDT online tool (36).

A total of 3,181 studies were initially identified from four databases (Figure 1). Before formal screening, duplicate records were removed, some records were automatically marked as ineligible, and additional records were excluded for other reasons. Subsequently, 1,426 titles and abstracts were screened (at this stage, Cohen’s κ = 0.998), and 1,211 records were excluded. A total of 215 reports were sought for retrieval; of these, 106 could not be retrieved, and the remaining 109 underwent full-text eligibility assessment (at this stage, Cohen’s κ = 0.997) (see upplementary fil). Of these, 95 full-text articles were excluded for an irrelevant intervention, an irrelevant comparator, a lack of a control group, or irrelevant outcome indicators. Ultimately, 14 studies were included in this meta-analysis.

The risk of bias for the 14 included RCTs is shown in Figures 2, 3 (yellow “–” = Some concerns; green “+” = Low risk). Two reviewers independently assessed 5 RoB 2.0 domains (D1–D5); discrepancies were resolved through discussion or a third reviewer. Domain distribution: D1: some concerns for a subset; D2: some concerns; D3: 1 study with some concerns; D4: balanced low/some concerns; D5: all low risk. Overall, most studies had some concerns; only 5 (19, 37–40) had low risk.

Inter-rater agreement for each domain was as follows: For D1, the simple agreement rate was 71.4% (10/14), with Cohen’s κ = 0.43 and weighted κ = 0.60 (interpreted as “Moderate”); for D2, the simple agreement rate was 57.1% (8/14), with κ = 0.14 and weighted κ = 0.31 (interpreted as “Fair”); for D3, the simple agreement rate was 92.0% (13/14), with κ = 0.86 and weighted κ = 0.93 (interpreted as “Very Good”); for D4, the simple agreement rate was 71.4% (10/14), with Cohen’s κ = 0.43 and weighted κ = 0.60 (interpreted as “Moderate”); for D5, the simple agreement rate was 100% (14/14), with Cohen’s κ = 1.00 and weighted κ = 1.00 (interpreted as “Almost Perfect”) (see upplementary fil).

Overall, domains D3 and D5 had an extremely high proportion of “Low risk” ratings, indicating that the studies were relatively reliable regarding follow-up completeness and transparency in result reporting. In contrast, domains D1, D2, and D4 had more “Some concerns” ratings. Notably, D2 had the lowest inter-rater agreement (only “Fair”), reflecting relatively large discrepancies in judging this type of bias. These results indicate that some studies still had shortcomings in reporting randomization details, blinding participants and researchers, and ensuring intervention adherence. Based on this bias profile, we cautiously interpreted potential systematic bias related to D2 in the main and subgroup analyses and incorporated these biases into sensitivity analyses and GRADE evidence quality assessments.

Fourteen RCTs published between 2001 and 2022 were included (Table 1), covering countries such as Spain, Sweden, the United States, Greece, Brazil, China, and Italy. Participants were adults with OSA, aged 42 to 54 years, with baseline AHI of 20 to 45 events/h and baseline BMI of 29–36 kg/m². Most patients had moderate-to-severe OSA, with obesity as the main health condition and a few were overweight. Intervention durations ranged from 4 to 48 weeks and included CPAP combined with dietary, exercise, or comprehensive lifestyle interventions. Control groups received CPAP alone, routine care, or CPAP plus standard lifestyle advice. Outcomes primarily focused on total AHI, with some studies also assessing supine, REM, and NREM AHI (19, 20, 37–48).

This study systematically assessed the effects of CPAP combined with lifestyle interventions on patients with OSA. A total of 14 randomized controlled trials involving 1,623 participants were included to evaluate changes in AHI. Substantial heterogeneity was observed across studies (I² = 91.3%, p < 0.001). The pooled analysis showed that CPAP combined with lifestyle interventions was associated with a statistically significant overall reduction in AHI (MD = −9.99, 95% CI: −14.55 to −5.44, p < 0.001) (Figure 4). Because I² was greater than 50% (I² = 91.3%, p < 0.001), Egger’s regression was used to evaluate potential publication bias. Egger’s regression is a statistical method used to detect publication bias in meta-analysis, primarily to assess whether study results were affected by small-sample effects (49). The multilevel Egger’s test indicated significant asymmetry (intercept t = −2.348, p = 0.019), suggesting the possible influence of small-study effects or selective reporting (see Supplementary file). Sensitivity analysis in meta-analysis is a statistical method used to assess the robustness and reliability of the results by determining the extent of the impact on overall outcomes when key parameters are altered or specific studies are excluded (50). Sensitivity analysis showed that the top 5 studies with the most significant impact on the pooled effect size did not cause significant deviation, confirming the overall pooled effect size’s stability (see Supplementary file).

Subsequently, standardized residuals (|Z| > 2.5) and Cook’s distance (threshold > 3 times the mean) were used to identify influential outliers in the model (51). The results indicated that the studies by Servantes et al. (46) and Jurado-García et al. (48). Studies had Cook’s distances exceeding the threshold but did not exceed the standardized residual threshold and were therefore not excluded (see Supplementary file). Furthermore, the trim-and-fill analysis showed that no studies needed to be imputed (labeled “No studies imputed”), and the adjusted effect size was identical to that of the original multilevel model (MD = −9.99). This finding further corroborates the robustness of the meta-analytic conclusions (see supplementary file). In addition, the small-sample-size test confirmed the statistical significance of the pooled effect size (Estimate = −9.99, SE = 2.33; t-statistic = −4.29, df = 12.1, p = 0.00103) (see Supplementary file). According to the GRADE assessment, moderate-quality evidence indicates a potential benefit of CPAP combined with lifestyle interventions in reducing AHI among patients with OSA. However, the strength of this evidence is constrained by risk of bias and pronounced heterogeneity (see Supplementary file).

Therefore, fourteen randomized controlled trials with 1,623 participants were synthesized using a random-effects model due to substantial heterogeneity across studies (I² = 91.3%, p < 0.001). The pooled estimate showed an absolute reduction in AHI with the combined intervention (MD = −9.99, 95% CI: −14.55 to −5.44) (Figure 4).

This study explored potential moderators through subgroup analyses, including (1) magnitude of BMI reduction, (2) intervention type (experimental/control), (3) intervention duration, (4) OSA severity, and (5) type of AHI measurement (Table 2).

In the experimental group, CPAP combined with comprehensive lifestyle intervention (MD = −11.99, 95% CI: −15.60 to −8.39, p < 0.001, GRADE: Moderate) and CPAP combined with exercise intervention (MD = −7.29, 95% CI: −10.65 to −3.94, p < 0.001, GRADE: Moderate) significantly improved AHI; CPAP combined with dietary intervention did not reach statistical significance (p = 0.071, GRADE: Very Low). In the control group, both CPAP alone (MD = −12.65, 95% CI: −19.07 to −6.23, p < 0.001, GRADE: Moderate) and CPAP plus routine lifestyle intervention (MD = −6.40, 95% CI: −10.87 to −1.93, p = 0.005, GRADE: Moderate) achieved significant AHI improvement. Regarding intervention duration, both < 12 weeks (MD = −19.29, 95% CI: −27.35 to −11.23, p < 0.001, GRADE: Low) and ≥ 12 weeks (MD = −7.97, 95% CI: −9.98 to −5.95, p < 0.001, GRADE: Moderate) significantly reduced AHI, with a more pronounced effect in the < 12 weeks subgroup. For OSA severity, moderate-to-severe patients showed significant improvement (MD = −11.55, 95% CI: −15.20 to −7.90, p < 0.001, GRADE: Moderate), whereas severe patients did not (p = 0.115, GRADE: Very Low). Regarding BMI reduction, the 0–3 kg/m² (MD = −12.02, 95% CI: −16.72 to −7.32, p < 0.001, GRADE: Moderate) and ≥ 5 kg/m² (MD = −23.39, 95% CI: −28.23 to −18.55, p < 0.001, GRADE: Low) subgroups showed significant AHI improvement, with no significant effect in the 3–5 kg/m² subgroup (p = 0.911, GRADE: Very Low). Among AHI measurement types, Total AHI (MD = −9.39, 95% CI: −13.84 to −4.95, p < 0.001, GRADE: Moderate), supine AHI (MD = −12.14, 95% CI: −20.73 to −3.55, p = 0.006, GRADE: Low), NREM AHI (MD = −11.60, 95% CI: −22.81 to −0.38, p = 0.043, GRADE: Low), and REM AHI (MD = −14.46, 95% CI: −21.91 to −7.02, p < 0.001, GRADE: Moderate) all showed significant reductions.

These results indicated that CPAP combined with exercise or comprehensive lifestyle intervention, shorter intervention durations (more effective in < 12 weeks), moderate-to-severe OSA, and BMI reductions of 0–3 kg/m² or ≥ 5 kg/m² were more conducive to optimizing AHI reduction in OSA patients; REM AHI was the most sensitive to intervention. The most effective subgroups had moderate GRADE evidence quality.

To explore potential sources of heterogeneity, random-effects meta-regression analyses using REML estimation were conducted to assess associations between eight pre-specified moderators and the pooled effect size (MD). The moderators included AHI outcome type, age, intervention cycle, baseline AHI, OSA severity, health condition, experimental intervention type, and control intervention type. As shown in Figure 5, most moderators did not significantly explain between-study heterogeneity (all p > 0.05), including AHI outcome type (β = −3.72, p = 0.478), intervention cycle (β = 0.12, p = 0.306), baseline AHI (β = −0.16, p = 0.562), OSA severity (β = −0.57, p = 0.938), health condition (β = 4.65, p = 0.208), experimental intervention type (β = 6.89, p = 0.117), and control intervention type (β = −2.54, p = 0.386).

Notably, age was significantly associated with the pooled effect size (β = −1.12, p = 0.024), indicating that greater reductions in AHI were observed with increasing age. This suggested that older patients tended to derive greater benefit from CPAP combined with lifestyle interventions. However, the explanatory power of the meta-regression analysis was limited, likely due to the relatively small number of included studies and substantial heterogeneity in intervention protocols.

Several limitations should be acknowledged. First, the number of included studies was relatively small, and follow-up durations were generally short, which may have limited the evaluation of long-term treatment effects. Second, variability in lifestyle intervention components and intensity contributed to substantial heterogeneity across studies. Future large-scale, long-term randomized controlled trials using standardized lifestyle intervention protocols are warranted to confirm these findings and further explore individualized treatment strategies across different patient populations.