An umbrella review of systematic reviews and meta-analyses of observational investigations of obstructive sleep apnea and health outcomes

From initiation until November 23, 2020, literature searches were performed using online databases such as Embase, PubMed, the Cochrane Database of Systematic Reviews, and the Web of Science. Literature searches were independently conducted by two researchers (CZ and LG). The search terms applied were (“obstructive sleep apnea” OR “obstructive sleep apnea–hypopnea” OR “OSA” OR “OSAH”) AND (Meta-Analysis[ptyp] OR metaanaly*[tiab] OR meta-analy*[tiab] OR Systematic review [ptyp] OR “systematic review”[tiab]). The references were manually screened to identify eligible articles to be included in the study. The article titles, abstracts, and the complete manuscripts of the identified paper were then further assessed. A discussion was used to resolve potential discrepancies; ST acted as an arbiter to deal with discrepancies that could not be resolved by discussion among the investigators.

The eligibility of articles was based on a systematic search by the authors to identify the most pertinent studies. Only systematic reviews or meta-analyses on the basis of the epidemiological studies performed in humans were considered in the analysis. Diagnostic trials and meta-analyses of interventional trials were not performed as part of the current study. Furthermore, the abstracts of the conference on review questions were not included in the final analysis. The final systematic reviews and meta-analyses that were analyzed had to include the data of pooled summary effects(i.e., relative risks (RRs); odds ratios (ORs); hazard ratios (HRs); mean difference (MD); weighted mean difference (WMD); standard mean difference (SMD); and their 95% confidence intervals (CIs)), number of included researches, number of participants and cases, heterogeneity, and publication bias. Whenever more than one meta-analysis was executed using on the basis of the same outcome, the agreement with the main conclusions reported in the study were verified. When the reported conclusions were conflicting, the meta-analysis with the greatest number of investigations was considered.

For investigations to be eligible for inclusion in the meta-analysis, two researchers (WC and YL) independently extracted data from the articles. This included the first author, the number of included investigations, the year of publication, the study design, the whole numbers of cases, and participants. The reported relative summary risk evaluates (ORs, RRs, HRs, SMD, WMD, or MD) and the corresponding 95% CIs were extracted, for each eligible systematic review and meta-analysis. The values of p for the total pooled effects, Cochran Q measurement, Egger’s measurement, and I² were extracted. Discrepancies in the analyses were resolved by discussion among the investigators.

Two investigators (WC and YL) independently assessed the quality of the methods reported in the studies. This was performed using a 16-criteria checklist included in AMSTAR 2 [24]. AMSTAR 2 is a fundamental revision of the original instrument of AMSTAR which was devised to evaluate systematic reviews that included randomized controlled experiments. The AMSTAR 2 score is categorized as high in studies that have no or one noncritical weakness, moderate in surveys with more than one noncritical weakness, low when the study has only one serious flaw without or with noncritical weaknesses, and seriously low when a study has more than one serious flaw without or with nonserious weaknesses. Discrepancies between the AMSTARS 2 scores for the articles were resolved by discussion between the investigators.

Two investigators (WC and YL) independently evaluated the quality of the evidence conforming to the parameters that have previously been applied in various fields [25‐28]. Discrepancies were resolved by discussion. First, p value for the estimate < 0.001 [29, 30] and more than 1000 cases of the disease, which indicated fewer false-positive results. Second, I² < 50% and p value for Cochran Q test > 0.10, which indicated consistency of results. Third, p value for Egger’s test > 0.10, which exhibited no evidence of small-study impacts. When all of the above criteria were satisfied, the strength of the epidemiologic evidence was rated as high. When 1 of the criterion was not satisfied and the p value for the estimate was < 0.001, the strength of the epidemiologic evidence was rated as moderate. Then, the rest was defined as weak (p < 0.05). The value of p for the evaluation can be assessed from the 95% confidence interval of the pooled impact estimate utilizing an established method [31] if it was not directly reported in the article.

From each of the published studies, the outcome data of the available meta-analyses was extracted along with the estimated summary effect at the corresponding 95% CI. The total impacts of the pooled meta-analysis were considered significant when the p-value was < 0.05. Heterogeneity was appraised by the I² test and Q test, publication bias was estimated by utilizing Egger’s test, and both were considered significant at p < 0.1. Studies that did not have the heterogeneity or publication bias results were reanalyzed if raw data were available.

The outcomes of the systematic investigation and the selection of eligible investigations are summarized in Fig. 1. Overall, 1972 articles were searched from which 66 meta-analyses of observational investigations were identified that had 136 unique outcomes [21, 22, 32‐95]. The 66 eligible non-overlapping meta-analyses had publication dates ranging from 2009 to 2020 and are summarized in Table 1. The median number of primary investigations per evidence synthesis was 7 (range 2–64). Furthermore, 1 meta-analysis [54] lacked the data of both participants and cases, and 2 meta-analyses [52, 95] lacked the data of cases. Among the meta-analyses identified in this study, the median number of cases was 900 (88–3,117,496) and the median number of participants was 2962 (170–56,746,100). An extensive range of data were reported such as cardiovascular disorders (n = 31), cerebral and cerebrovascular disease (n = 7), mortality (n = 5), postoperative complications (n = 20), pregnancy-related disorders (n = 13), ophthalmic disorders (n = 8), digestive disorders (n = 13), endocrine and metabolic system disorders(n = 17), urological disorders (n = 7), and other data (n = 15) (Fig. 2).

A brief explanation of the effects of the included meta-analysis is given in Table 1. Overall, 111 (82%) of the 136 data reported significant summary outcomes (p < 0.05). These associations relate to the outcomes of the following different systems: 29 meta-analyses in cardiovascular disorders, 5 in cerebral and cerebrovascular disease, 4 in mortality, 14 in postoperative complications, 8 in pregnancy-related disorders, 7 in ophthalmic disorders, 11 in digestive disorders, 14 in endocrine and metabolic system, 7 in urological disorders, and 12 in other outcomes. Therefore, it can be concluded that OSA can enhance the risk of disease and have adverse effects on human health.

For heterogeneity, 5 results in 5 articles were reanalyzed owing to that they did not exhibit the outcomes of heterogeneity [22, 36, 46, 59, 64]. Among the 136 outcomes including the reanalyzed articles, 47 outcomes showed no heterogeneity between researches (p ≥ 0.1 of Q test), whereas 69 indicated significant heterogeneity (p < 0.1 of Q test). However, there were still 20 results in 2 articles that could not be reanalyzed due to the lack of raw data [52, 95], so we could not evaluate their heterogeneity. For publication bias, 76 outcomes demonstrated no statistical evidence on publication bias (p ≥ 0.1 of Egger’s test), whereas 25 outcomes presented publication bias (p < 0.1 of Egger’s test). There were still 35 results in 9 articles that could not be reanalyzed due to the lack of raw data [45, 52, 54, 55, 87, 92‐95], so we could not evaluate their publication bias.

The results for the evaluation of the methodological qualities of the 66 included articles are shown in Table 2. Only 3 (5%) studies were determined to be low; the remaining 63 (95%) studies were determined to be critically low (Fig. 3). Based on the AMSTAR 2 criteria, none of the investigations were graded as moderate or high quality.

The outcomes of the evidence measurement are shown in Table 3. When a study did not present the result of heterogeneity and publication bias, the corresponding criteria were considered to be not satisfied. Among the 111 statistically significant outcomes, 7 (6%) showed high epidemiologic evidence, 24 (22%) showed moderate epidemiologic evidence, and the remaining 80 (72%) were rated as weak (Fig. 4).