The association between sleep disturbances and suicidal behaviors in patients with psychiatric diagnoses: a systematic review and meta-analysis

Inclusion and exclusion criteria

Eligible studies were comparative observational studies that reported sleep disturbances in psychiatric patients and the outcome of interest (any type of suicidal behaviors). We used the original study’s definition of sleep disturbances, which ranged from snoring, nightmares, hypersomnia insomnia, sleep panic attacks, and sleep deprivation. With respect to suicidal behaviors, we include suicide ideations, suicide attempts, completed suicides, and any other suicidal behaviors defined by the original studies. Studies were excluded if they did not report outcomes for psychiatric patients or did not provide sufficient data to quantitatively estimate the association between suicidal behaviors and sleep disturbances. We also excluded publications without original data (clinical reviews, editorials, letters, or erratum). No language or country restrictions were used.

Data sources and search strategy

An expert reference librarian and study authors with expertise in conducting systematic reviews developed the search strategy. A comprehensive search of databases from each database’s earliest inclusive dates to 26 June 2013 was conducted. The databases included Ovid Medline In-Process & Other Non-Indexed Citations, Ovid MEDLINE, Ovid EMBASE, Ovid PsycInfo, Ovid Cochrane Database of Systematic Reviews, Ovid Cochrane Central Register of Controlled Trials, and Scopus. Controlled vocabulary, supplemented with keywords, was used to search for the concept areas: suicide, sleep disorders, and psychiatric conditions. The database search strategy is available in Additional file 1. We also searched additional references from a recent systematic review [6].

Study selection and data extraction

Reviewers working independently and in duplicate assessed each abstract for eligibility. Disagreements yielded an automatic inclusion into the following level of screening. Included studies were retrieved and full text screening commenced in duplicate as well. Disagreements at this level were resolved by discussion and consensus. Two reviewers working independently and in duplicate extracted baseline and outcome data and assessed the quality of each included study. A third reviewer compared the reviewers’ data and resolved inconsistencies by referring to the full text article.

Reviewers independently extracted study details from the full text articles, using a pilot-tested form. The following data were abstracted: study design, country, patient characteristics (sex, age), psychiatric diagnoses, sleep disturbances, and suicidal behaviors. We extracted numbers of patients with outcomes (suicidal behaviors) at the longest duration of complete follow-up.

Assessment of study quality and publication bias

Using the Newcastle-Ottawa scale, we assessed the quality of included observational studies by determining outcome ascertainment, adjustment for confounders, and proportion of patients lost to follow-up as well as sample selection. We assessed potential publication bias by visual inspection of funnel plots and the Egger’s regression asymmetry tests.

Statistical analysis

The outcome of interest was suicidal behaviors, including suicidal ideations, suicide attempts, and completed suicides. Outcomes were either dichotomized by the individual study or when presented as a scale, were converted to log transformed OR (logOR) based on the commonly used formula suggested by Borenstein et al. [10].

We pooled the logOR from all of the included studies using the DerSimonian & Laird random effects method with the estimate of heterogeneity from the Mantel-Haenszel model [11]. We conducted a priori planned subgroup analysis based on the type of suicidal behaviors (ideations, attempts, and completed suicides). We subsequently evaluated separately the different types of sleep disturbances (insomnia, hypersomnia, parasomnia, sleep-related breathing disorders, and mixed/unclear types) to further explore heterogeneity across studies. When one study reported outcomes for multiple subgroups, these outcomes were pooled separately in each subgroup. When multiple outcomes were reported in one subgroup, we chose the one that clearly defined and/or objectively measured.

We used the I² statistic to measure heterogeneity across the studies. Statistical analyses were conducted using STATA version 12 (StataCorp, College Station, TX, USA) and R software version 3.0.1 (R Foundation for Statistical Computing, Vienna, Austria).

Description and quality of included studies

Our search identified 372 candidate references (Figure 1). After deleting duplicated studies and excluding irrelevant studies, 19 studies met the inclusion criteria and were included in this analysis. The characteristics of these studies are summarized in Table 1. Overall, 104,436 patients were included in this review. The average age was 49.4 years (range, 17-79 years). Fifty-eight percent of the patients were female. The average length of follow-up in the prospective studies was 9.9 years (range, 3 days - 27 years). Of 19 studies, 13 (68.4%) reported outcomes in depressive patients. The other psychiatric diagnoses included were post-traumatic stress disorder (PTSD), panic disorder, schizophrenia, and anxiety. Eleven studies (57.9%) were prospective studies, four (21.1%) were cross-sectional, and three (15.87%) were retrospective.

Figure 2 shows the quality indicators of the included studies. We found variable quality across studies. Most (68.4%) of the studies did not adjust for risk factors. However, 15 studies reported low risk of bias for the ascertainment of outcomes. Using the Egger regression asymmetry test and visual inspection of the funnel plot, we found potential publication bias in this body of evidence. Across the included studies, substantial heterogeneity was observed in most of the pooled outcomes (>50%) [28]. In summary, the risk of bias is moderate to high due to potential publication bias and substantial heterogeneity.

Figure 3 shows the pooled odds ratios (ORs) from all of the included studies. Compared to those without sleep disturbances, patients with psychiatric diagnoses and co-morbid sleep disturbances were significantly more likely to report suicidal behaviors (OR = 1.99, 95% CI 1.72, 2.30, P <0.001). Regarding to specific psychiatric conditions, we found strong association between suicidal behaviors and sleep disturbances in depression (OR = 3.05, 95% CI 2.07, 4.48, P <0.001), PTSD (OR = 2.56, 95% CI 1.91, 3.43, P <0.001), panic disorders (OR = 3.22, 95% CI 1.09, 9.45, P = 0.03), and schizophrenia (OR = 12.66, 95% CI 1.40, 114.44, P = 0.02). Table 2 shows the pooled OR for different suicidal behaviors. Sleep disturbances were also significantly associated with suicide ideations, suicide attempts, and completed suicides. Subgroup analysis based on the type of sleep disturbance showed that insomnia, parasomnia, and sleep-related breathing disorders were significantly associated with increased risk of suicidal behaviors (Table 3). The only exception was hypersomnias (OR = 1.91, 95% CI: 0.60, 6.06, P = 0.27).

Clinical implications

Given the findings of this study, the presence of sleep disturbances in individuals with psychiatric illness may trigger the need for further evaluation of increased risk for suicide. In particular, clinicians evaluating psychiatric patients who report acute changes in sleep should be aware that this may represent a specific vulnerability for suicide, and a careful suicide risk assessment may be warranted. Correspondingly, a comprehensive suicide risk assessment may include an evaluation of sleep quality and maintenance. Screening measures such as the Pittsburgh Sleep Quality Index (PSQI) [30], which assesses sleep quality, efficiency, duration, disturbances, and daytime dysfunction, and has good reliability and validity in detecting sleep problems, may assist in this effort. A suicide risk assessment that includes evaluation of sleep disturbances may not only add to the estimation of risk, but it may also provide a potential target for intervention. This is particularly promising given that many of the previously established risk factors are person variables such as age, type of psychiatric disorder, and family history of suicide attempts, and thereby not amenable to intervention.

In terms of intervention, there is considerable evidence for the efficacy of Cognitive Behavioral Therapy for Insomnia (CBT-I) with large effect sizes found in meta-analytic studies [31]. Not only is CBT-I effective for treating primary sleep disorders, research examining CBT-I in patients with major depression and co-morbid insomnia found that the treatment led to improvements in both disorders [32]. Although CBT-I is considered to be a first-line treatment for insomnia [33], pharmacological treatments may also be given consideration for insomnia and other sleep problems. However, caution should be taken as studies found sedative-hypnotics were associated with increased suicidal behaviors [34–37]. A large cohort study demonstrated that receiving hypnotic prescriptions was associated with greater than three-fold increased hazards of death even when prescribed <18 pills/year. This association held in separate analyses for several commonly used hypnotics and for newer shorter-acting drugs [38].

Limitations and strengths

Strengths of the study include the comprehensive literature search, application of bias protection measures in the study selection, and careful evaluation of methodological quality. There are important limitations. Observational studies are subject to high risk of bias due to potential outcome confounding. Substantial heterogeneity was observed across the studies. Publication bias may have also affected our results. At last, we observed substantial variations of how the included studies define sleep disturbances and suicidal behaviors, and of how these studies measure them. All of the above may affect our findings. Based on the Grading of Recommendations Assessment, Development and Evaluation (GRADE) methodology, the strength of the current evidence is low due to potential publication bias and heterogeneity [39].