Association between time of discharge from ICU and hospital mortality: a systematic review and meta-analysis

Two independent investigators (SY and JW) performed a systematic search, without language or publication type restrictions, of the PubMed, Embase, and Scopus databases from their inception to 1 August 2016. The searches used a combination of the following search terms with the appropriate wildcards and spelling variations: “intensive care unit”, “night-shift”, “night”, “nighttime”, “out-of-hours”, “evening”, “off-hour”, “after-hours”, “time”, “discharge”, “transfer”, “mortality”, and “death”.

The search was limited to studies of human adults (Additional file 1: Table S1). Publications in non-English languages (e.g., French, Japanese, or German) were translated by an independent translation service. Additional searches were performed using two clinical trial registries (http://​clinicaltrials.​gov/​ and http://​www.​isrctn.​com/​), and abstracts from major international conferences were manually searched at their official journal websites (Society of Critical Care Medicine: Critical Care Medicine (1998–2015); American Thoracic Society: American Journal of Respiratory and Critical Care Medicine (2009–2016); European Society of Intensive Care Medicine: Intensive Care Medicine (1988–2014); International Symposium on Intensive Care and Emergency Medicine: Critical Care (1997–2016); American College of Chest Physicians: Chest (2003–2015); Australian and New Zealand Intensive Care Society Annual Meeting: Anaesthesia and Intensive Care (1990–2015)).

Articles that were published online ahead of print in major intensive care journals were searched manually. The two investigators also reviewed the reference lists (Additional file 1: Table S1) of the retrieved studies and relevant reviews to identify additional articles [19, 20]. In instances where further clarification was required, a third investigator (ZW) emailed the corresponding author of the relevant article.

The meta-analysis was pre-specified and performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) criteria (Additional file 2) [21]. The meta-analysis and systematic review protocol has not been published, and is not registered with the PROSPERO database or the Cochrane Library.

Studies were considered eligible if they fulfilled the following criteria: (1) a cohort study design; (2) a study population of mainly adult patients who were discharged alive from an ICU (general surgical, medical, or mixed) and were grouped into nighttime/daytime discharges and/or weekday/weekend discharges, and the study assessed outcomes for nighttime versus daytime discharges or outcomes for weekend versus weekdays discharges; (3) the primary outcome was hospital mortality among the patients who were discharged from the ICU (according to their nighttime, daytime, weekday, and/or weekend grouping); and (4) the study reported the effect size and 95% confident interval (CI) with adjustment for disease severity (or data to calculate these results).

Studies were excluded if they fulfilled any of the following criteria: (1) the study population comprised mainly pediatric patients; (2) the study population comprised patients discharged from a high-dependency or step-down unit; (3) there was no control population; (4) the study was not original research; or (5) the study did not provide sufficient information for data extraction and quality assessment (even after contacting the relevant authors). In cases of duplicate publication, we only included the most informative and complete study (typically the most recent publication).

Two investigators (SY and JW) independently screened the titles and abstracts of all citations. The full-text articles were retrieved for full-text review if either investigator thought that the citation might fulfill our eligibility criteria. The same two investigators independently evaluated the eligibility of all full-text articles that were selected during the screening process, and the κ value (i.e., chance-independent agreement) was found to be 0.82. Disagreements were resolved through a consensus process in which investigators discussed the reasoning behind their decisions. In all disagreements, one of the investigators realized that they had made an error.

Data extraction was independently performed by two investigators (SY and ZL); discrepancies were resolved using discussion and consensus. A predefined standardized data extraction form was used to collect data. The following data were collected from each study: the study name, the first author’s name, publication year, the study design, the study location, the patients’ ages, the patients’ sex distribution, the definition of night or weekend, disease severity, adjustments, outcomes, odds ratios (ORs) and 95% CIs, numbers of patients discharged during the nighttime and the daytime, crude hospital mortality among patients discharged during nighttime and daytime, numbers of patients discharged during the weekend and weekdays, crude hospital mortality among patients discharged during the weekend and weekdays, and total number of patients discharged. We also checked the supplementary files and contacted the study authors in cases where more detailed information was needed.

Because all of the included studies were cohort studies, the Newcastle-Ottawa Scale (NOS) was used to assess study quality (available at: http://​www.​ohri.​ca/​programs/​clinical_​epidemiology/​oxford.​asp) [22]. This scale uses a star system to evaluate study quality in three domains: cohort selection (maximum of four stars), comparability (maximum of two stars), and outcome (maximum of three stars). A score of nine stars indicates the highest possible quality. For the present study, we defined high-quality studies as having >5 stars (a low risk of bias) and low-quality studies as having ≤5 stars (a high risk of bias) [23, 24]. Two investigators (SY and ZL) independently performed the quality assessment; discrepancies were resolved using discussion and consensus.

The primary meta-analysis evaluated the association between nighttime ICU discharge and hospital mortality. The secondary meta-analysis evaluated the association between weekend ICU discharge and hospital mortality. The overall estimates were presented as OR and 95% CI values, which were determined using a random-effects model that accounted for any differences between the studies, even if there was no statistically significant heterogeneity [25].

The individual estimates were used in the sub-analyses for one study [12] that separately reported ORs for weekday night discharge, weekend daytime discharge, weekend night discharge, and weekday daytime discharge. However, to compare nighttime and daytime discharge, we combined the weekend night discharges and weekday night discharges into a single group. An overall estimate for this group was calculated from the available ORs for weekend night discharge and weekday night discharge using a fixed-effects model and the inverse-variance method. To compare weekend and weekday discharge, we combined weekend night discharges and weekend day discharges into another group. An overall estimate for this group was calculated from the available ORs for weekend night discharge and weekend day discharge using a fixed-effects model and the inverse-variance method [26].

Heterogeneity was evaluated using the Cochran Q statistic and I ² statistic, which are quantitative measures of inconsistency across studies [27], and heterogeneity was considered statistically significant at P values <0.1 or I ² values >50%. Subgroup analyses were performed to examine the potential sources of heterogeneity using pre-specified subgroups that included geographical region, study design, and study characteristics (definition of nighttime and the total number of patients discharged).

We also performed post-hoc subgroup analyses according to adjustments for certain confounding factors (adjustment for illness severity at the ICU discharge, adjustment for treatment limitation orders, and adjustment for premature discharge), as these confounding factors might affect the results of our analyses.

Sensitivity analysis was performed to explore the possible causes of any heterogeneity and to estimate the influence of missing studies on the overall estimates by changing the pooling model (from a random-effects model to a fixed-effects model) and using the one-study-out method. Egger linear regression testing was performed to test for publication bias [28]. All statistical analyses were performed using STATA software (version 12.1; StataCorp, College Station, TX, USA), and differences were considered statistically significant at a two-tailed P value <0.05.

The initial search identified 5259 potentially relevant publications, although 2113 reports were excluded because of duplicate publication. We also excluded 3106 studies based on reviews of the titles and abstracts. Full-text reviews were performed for the remaining 40 studies, and we ultimately identified 14 cohort studies for inclusion in the meta-analysis [5‐15, 29‐31]. The justifications for the study exclusions are shown in Additional file 1: Table S2. The strategies for study identification and study selection are shown in Fig. 1.

The main characteristics of the 14 included studies are shown in Tables 1, 2 and 3. The 14 studies were all published in English between 2000 and 2015. Six studies were performed in Oceania [7‐10, 15, 30], four studies were performed in Europe [5, 6, 14, 29], and the other four studies were performed in North America [11‐13, 31]. The studies included six single-center studies and eight multicenter studies, and used a retrospective design (n = 12) or a prospective design (n = 2).

Disease severity was reported based on the Acute Physiology and Chronic Health Evaluation (APACHE) II score (n = 8 studies), the APACHE III score (n =4), or the Simplified Acute Physiology Score (SAPS) II (n = 2). In all 14 studies there was adjustment for a wide range of potential confounders, such as age, treatment limitation orders, premature discharge, diagnostic category, and the use of mechanical ventilation. In 11 studies there was adjustment for disease severity on ICU admission [5‐12, 14, 30, 31], and in 3 studies adjustment for disease severity on ICU discharge [13, 15, 29]; in 6 studies there was adjustment for treatment limitation orders [7, 13‐15, 29, 30], but no such adjustment in the other 8 studies [5, 6, 8‐12, 31]. In two studies there was adjustment for premature discharge [5, 7].

A total of 953,312 patients were included in the meta-analysis, with the study samples ranging from 1654 patients to 710,535 patients. Four studies included ≤10,000 patients [6, 7, 10, 29] and 10 studies included >10,000 patients [5, 8, 9, 11‐15, 30, 31]. The mean proportion of nighttime discharges in 13 studies was 15.3% (range 3.6–34.7%), and one study did not report this information [8]. The evaluated discharge times included nighttime in all 14 studies [5‐15, 29‐31] and weekends plus nighttime in 5 studies [8, 10, 12, 14, 31]. None of the 14 studies used consistent definitions of nighttime or weekend. The average NOS score of the included studies was 6.5 (range 6–8) (see Additional file 1: Table S3).

The 14 studies had 953,312 patients who were evaluated for daytime/nighttime discharge (Table 2). The adjusted OR for hospital mortality was significantly higher among patients discharged during the nighttime, compared to patients discharged during the daytime (OR 1.31, 95% CI 1.25–1.38, P < 0.0001) (Fig. 2), and the individual studies had low heterogeneity (I ² = 33.8%, P = 0.105). We also performed subgroup analyses and sensitivity analyses (Table 4), which revealed that the significant association between nighttime discharge and hospital mortality was not substantially modified by geographical region, the total number of discharges, or study design (Additional file 3: Figures S1–S3).

Among the 11 studies in which there was adjustment for disease severity on ICU admission, the summary OR for hospital mortality was 1.31 (95% CI 1.25–1.38). A summary OR was provided for hospital mortality with a broader range (OR 1.26, 95% CI 1.02–1.57) in the three studies in which there was adjustment for disease severity at ICU discharge (Fig. 3). In our meta-analysis, the risk of hospital mortality did not depend on whether or not the study analyses were adjusted for disease severity at ICU discharge.

Among the six studies in which there was adjustment for treatment limitation orders, the summary OR for hospital mortality was 1.28 (95% CI 1.15–1.43); among the remaining eight studies without this adjustment, the summary OR for hospital mortality was 1.32 (95% CI 1.23–1.42) (Fig. 4). In our meta-analysis, the risk of hospital mortality did not depend on whether or not there was adjustment for treatment limitation orders. We did not observe a significant relationship between nighttime discharge and hospital mortality in the subgroup analysis that was adjusted for premature discharge (Fig. 5). Our sensitivity analyses suggested that the overall estimates were not materially altered by changing the pooling models (random-effects model, OR 1.31, 95% CI 1.25–1.38; fixed-effects model, OR 1.33, 95% CI 1.30–1.37) and were not materially altered when an individual study was omitted from the sequence, with a range of 1.29 (95% CI 1.22–1.36) to 1.33 (95% CI 1.27–1.39) (Table 4).

Five studies with 70,883 patients evaluated weekday/weekend discharges (Table 3). There was no difference in the adjusted ORs for hospital mortality when we compared patients who were discharged during the weekend or on weekdays (OR 1.03, 95% CI 0.88–1.21, P = 0.68) (Fig. 6). However, the studies exhibited significant heterogeneity (I ² = 72.5%, P = 0.006), and we performed sensitivity analyses to explore the possible explanations for the heterogeneity. Our sensitivity analyses suggested that the overall estimates were not materially altered by changing the pooling models (random-effects model, OR 1.03, 95% CI 0.88–1.21; fixed-effects model, OR 1.00, 95% CI 0.93–1.08) and were not materially altered when an individual study was omitted from the sequence, with a range of 0.95 (95% CI 0.87–1.03) to 1.08 (95% CI 0.89–1.31). We were unable to identify the specific study that caused the heterogeneity, which might have been explained by the different patient populations, weekend definitions (including a different number of weekend days and weekend nights), and adjustments for confounding factors. However, when the Tobin study [8] was omitted, the heterogeneity among the remaining four studies was markedly reduced (I ² = 0.0%, P = 0.792), and we did not observe an association between weekend discharge and hospital mortality.

Egger’s test did not reveal any significant evidence of publication bias (P = 0.662 for nighttime studies, P = 0.507 for weekend studies) (Additional file 3: Figures S4 and S5).

This study has several strengths. First, we identified studies using a comprehensive systematic literature search. Second, we evaluated 953,312 patients with daytime/nighttime discharge, and the large sample size significantly increased the statistical power of the analysis. Third, the pooled estimates were stable after the comprehensive sensitivity analyses. Fourth, we did not detect publication bias, which indicates that the pooled estimates may be unbiased.

This study also has several limitations. First, the cohort studies were observational and descriptive, and we cannot comment on the causality of the relationships between hospital mortality and the factors that we evaluated. Second, the studies used different definitions for nighttime and weekend, which inevitably introduces heterogeneity. Third, although both disease severity at discharge and admission can predict post-ICU mortality, only three of the included studies measured disease severity at ICU discharge. Premature ICU discharge was also more common at night, although only two of the included studies adjusted for premature discharge. Thus, the limited numbers of studies that adjusted for premature discharge or disease severity at ICU discharge might cause residual confounding. Fourth, none of the studies reported the exact demand of care and actual level of care in the ward, which obscures whether or how treatment and nursing care insufficiencies might influence post-ICU mortality. Last, the included studies did not report long-term prognosis or its possible relationship with nighttime discharge, which may also be an important issue. Moreover, the studies failed to address novel trends in critical care, including tele-ICU, liaison nurses, and rapid response teams, which may profoundly alter intensive care and post-ICU care and further alter the rate and effect of nighttime discharge. Therefore, the generalizability of our results in the real world must be tested in future studies [51].