Systematic review and meta-analysis of the relationship between sleep disorders and suicidal behaviour in patients with depression

A systematic search strategy with no language, geographical, or publication-type restrictions was developed in January 2019. Our search consisted of two steps. First, we searched the electronic databases PubMed, EMBASE and the Cochrane Library from inception to January 1, 2019. We used the following Medical Subject Headings (MeSH) terms and their combinations of free words to search: sleep disorders/sleep arousal disorders/sleep deprivation//sleep wake disorders/sleep initiation and maintenance disorders, depressive disorder/depression/mental disorders, and suicidal behaviour/suicide ideation/suicide attempt /suicide. Second, we implemented a manual search based on the reference lists of the articles included in the first step and contacted the authors via email to obtain unpublished data. We supplemented the results of electronic searches with manually searched comments, reviews, and conference abstracts. The detailed search strategy can be found in Additional file 1: Table S2. The entire search was independently performed by two reviewers (XFW & SXC).

We included all observational studies that reported an association between sleep disorders and suicidal behaviour in patients with depression. “Depression” in this study referred to the definition of the diagnosis of depression (including major depression) according to internationally recognized diagnostic tools, excluding patients with depressive symptoms but diagnosed with mental illnesses other than depression, such as bipolar disorder. If the original study clarified that the patient’s diagnosis was depression, our default assumption was that they had excluded a disorder such as bipolar disorder that needed to be excluded in the diagnosis of depression. According to the inclusion criteria of the original studies, if the author reported major depressive disorder (MDD) or indicated that the subject’s depression was severe, we assigned this study to the “major depression” group. If the author only mentioned depression and did not report the severity of the patient’s depression, we assigned the study to the “depression” group. The classification of sleep disorders was broad and included insomnia, hypersomnia, nightmares, and various sleep problems defined by the authors in their original studies, such as low sleep quality, short sleep duration, and low habitual sleep efficiency. Suicidal behaviour in this meta-analysis included suicidal ideation, suicide attempt and completed suicide. Studies were excluded if they (a) studied an irrelevant topic, used an unavailable population or had incomplete data, (b) were reviews, conference abstracts, editorial letters, qualitative research or animal model papers, or (c) were duplicated reports. If there were duplicate studies, the most recent or most complete version was selected. If multiple languages were used to describe and publish the same data, the English version was selected. We extracted data on the first author, year of publication, country of origin, study design, participant numbers and characteristics, exposure and outcomes and their measurement tools. The data from all included studies were extracted and reviewed independently by XFW and SXC. Any dispute or disagreement between the two reviewers that could not be agreed upon was arbitrated by a third reviewer (HLX).

The methodological quality of all included studies was independently assessed using the modified version of the Newcastle-Ottawa Scale (NOS) [25]. Potential risk of bias was identified by five items, including representativeness of the subjects, diagnosis of depression, adjustment for confounders, ascertainment of outcomes and reporting loss to follow-up. The risk of each item was identified as “low risk”, “high risk” or “unclear risk”. Low risk was recorded as “1 point”, and the other two types of risk were recorded as “0 points”. A study with a total score of 3 points or above indicated a lower risk of bias.

Confidence in the estimate for the outcome was determined by the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach [26]. This assessment was performed by GRADEpro software (version 3.2, Evidence Prime, Hamilton, ON, Canada, 2015) [27]. The rating criteria were based on risk of bias, inconsistency, indirectness, imprecision, publication bias and other considerations. Because all included studies in this meta-analysis were observational studies, the quality of the evidence was initially defined as low. However, the rating could be upgraded to high (a) if the magnitude of the treatment effect was large (relative risk (RR) > 2 or RR <  0.5) or very large (RR > 5 or RR <  0.2), (b) if there was evidence of a dose-response relation or (c) if all plausible biases could decrease the magnitude of an apparent treatment effect. In addition, according to the handbook of GRADE, because inconsistency in this study could be explained by differences in populations (depression vs. major depression), we analysed, presented and graded the quality of the evidence by stratifying according to this variable.

The result of interest was the relationship between sleep disorders and suicidal behaviour. Sleep disorders included insomnia, nightmares, and hypersomnia; no specific sleep disorder was classified as “other”. Suicidal behaviour included suicidal ideation, suicide attempt and completed suicide. Studies that did not specifically address suicidal behaviour were also classified as “other”. The association between sleep disorders and suicidal behaviour was expressed as the odds ratio (OR). We obtained the ORs and corresponding 95% confidence intervals (CI) from most of the included studies. If those values were not directly available, we calculated the OR and 95% CI with the coefficient estimates and standard errors mentioned in the text.

The statistical analysis was conducted using Stata 12.0 (College Station, TX: StataCorp LP) and RevMan 5.3 (Cochrane Collaboration, Copenhagen, Denmark). Outcomes were assessed by pooled logOR among all included studies, and the corresponding 95% CIs were calculated by a fixed effects model or a random effects model. Statistical heterogeneity was defined by Cochran’s Q test with a P value < 0.1 or an I² value > 50%, and a random effects model was used for subsequent analysis. Conversely, when the P value was > 0.1 or the I² value was < 50%, a fixed effects model was used. To explore potential sources of heterogeneity, meta-regression and subgroup analyses were implemented through several categories: year of publication (1997 ~ 2010 vs. 2010 ~ 2019), study design (retrospective vs. prospective vs. cross-sectional), sample size (< 1000 vs. > 1000), age of subjects (mean < 40 vs. mean > 40), diagnostic outcome (major depression vs. depression), diagnostic criteria (ICD-10 vs. DSM-III-R vs. DSM-IV vs. CES-D), sleep disorder type (insomnia vs. nightmares vs. hypersomnia vs. other), and suicidal behaviour type (suicidal ideation vs. suicide attempt vs. completed suicide vs. other). Potential publication bias was detected by testing the asymmetry of the funnel plot and performing an Egger’s linear regression test. Sensitivity analysis was determined by excluding studies with a relatively high risk of bias and comparing the results to previous effect values. P <  0.05 was considered statistically significant in all analyses.

A total of 1634 studies were identified through a preliminary database search, and 293 duplicates were excluded from Endnote X8 (Thomson Reuters, MI). By browsing the titles and abstracts, 132 studies were available and entered full-text screening. After reading these documents carefully, 114 studies were excluded (68 had data that were not extractable, 19 were reviews or conference abstracts, 15 were editorial letters, 9 were qualitive research studies, and 3 full texts could not be obtained). Finally, a total of 18 eligible studies were included in this meta-analysis. The detailed research identification and selection are shown in Fig. 1.

The 18 studies were published between 1997 and 2019, covering a total of 198,893 participants whose mean ages in each study ranged from 12.5 (SD =3.1) to 48.1 years (SD =15.6). There were 6 retrospective studies, 6 prospective studies and 6 cross-sectional studies. The sample sizes in each study varied from 41 to 163,512, and all patients were formally diagnosed with depression or major depression. Regarding the diagnostic criteria, the Diagnostic and Statistical Manual of Mental Disorders (DSM) was used in most studies (15), including the DSM-III-R (4) and the DSM-IV (11); the International Classification of Diseases, Tenth Revision (ICD-10) was adopted in two studies; and the Center for Epidemiology Scale for Depression (CES-D) was used in one study. The sleep disorders we included were hypersomnia, insomnia, sleep quality, etc. The sleep measures in these studies varied considerably, with most studies using stable scales (11). A few of the studies used questionnaires produced by the authors themselves (4), and two studies used a combination of a maturity scale and a homemade questionnaire. Of these studies, 13 provided ORs for sleep disorders and suicidal ideation, 7 reported ORs for sleep disorders and suicide attempt, and only 2 studies showed the results of sleep disorders and completed suicide. Moreover, 17 studies reported patient characteristics and the risk between sleep disorders and suicidal behaviours for both male and female subjects, whereas one study only reported female subjects’ relevant outcomes. The characteristics of the included studies are shown in Table 1.

The risk of bias varied across the included studies. We found that most studies showed a low risk of bias in the representativeness of subjects (17) and the diagnosis of depression (18). Three studies did not control for confounding factors; thus, they were identified as “high risk” for this item. Fifteen studies were considered to have a low risk of bias in the ascertainment of outcomes due to their objective outcomes. However, only five studies reported follow-up and loss of follow-up, and similar information was not available in most studies (10). Overall, the included studies could be identified as having a low risk of bias because their NOS scores were 3 points or above. The details of the methodology for bias assessment in the overall study and for each study are shown in Fig. 2.

After combining the ORs reported in the 18 included studies, we found that sleep disorders were associated with a significant risk of suicidal behaviour across all depressed patients in the random effects model. The pooled OR was 2.45 (95% CI: 1.33 4.52), and substantial heterogeneity was detected (I² = 99.1%. P <  0.001) (Fig. 3). The meta-regression analysis showed that the R² of the depression diagnostic outcome and suicidal behaviour type was 39.99% (P = 0.011) and 48.39% (P = 0.003), respectively. These variables were the cause of partial heterogeneity among studies; that is, a depression diagnosis could explain 39.99% of the heterogeneous sources, and the type of suicidal behaviour could explain 48.39% of the heterogeneous sources. The year of publication, study design, sample size, age of subjects, diagnostic criteria and sleep disorder type failed to explain the heterogeneity among studies (P > 0.05) (Additional file 1: Table S3).

In the subgroup analysis, we found that patients diagnosed with major depression (OR = 2.92, 95% CI:1.42 5.98) had a higher risk of suicidal behaviour than those diagnosed with depression (OR = 1.41, 95% CI:1.14 1.73). Studies with the DSM-III-R as diagnostic criteria (OR = 4.80, 95% CI: 2.90 7.67) showed a higher risk of suicidal behaviour, followed by those with the ICD-10 (OR = 2.13, 95% CI: 1.10 4.12) and the DSM-IV (OR = 2.11, 95% CI: 0.96 4.65). For the variable “sleep disorder type”, we found that insomnia (OR = 2.29, 95% CI:1.69 3.10) and nightmares (OR = 4.47, 95% CI:2.00 9.97) were significantly associated with an increased risk of suicidal behaviour. This relationship was not found for hypersomnia or other types of sleep disorders. Analysis based on the “suicidal behaviour type” subgroup showed that sleep disorders had a significant risk association with suicidal ideation (OR = 2.32, 95% CI:1.11 4.88) and suicide attempt (OR = 2.41, 95% CI:1.45 4.02). This association also existed with completed suicide but at the margin of error (P = 0.05). In addition, cross-sectional studies published between 1997 and 2010 and with a sample size < 1000 as well as with a mean age of patients < 40 detected a higher risk of suicidal behaviour. The details of the subgroup analysis are shown in Table 2.

There was visible asymmetry in the funnel plots, and Egger’s test was − 2.17 (P = 0.045). Combining the results of both analyses, we inferred that there was evidence of publication bias in this meta-analysis (Additional file 1: Figure S1).

Sensitivity analysis was performed after excluding five studies with a relatively high risk of bias. The pooled OR was 2.31 (95% CI: 1.12 4.77), which was slightly but not significantly lower than the previous pooled OR (OR = 2.45 95% CI: 1.33 4.52), indicating the stable and trustworthy nature of our analysis.

The quality of the evidence was rated as very low for the association between sleep disorders and suicidal behaviour in the overall outcome and in the major depression subgroup. This evidence quality was increased by one point because the magnitude of the effect was large (OR > 2, based on consistent evidence from at least two studies with no plausible confounders). However, due to the substantial heterogeneity among studies and because publication bias was detected, which subtracted two points, the final score was one point, which represented a quality level of “very low”. Additionally, the evidence was assessed as low for the outcome of the depression subgroup. As with the status of evidence at the beginning of observational studies, no significant upgrade or downgrade factors were detected (Table 3).