We identified 1961 papers, of which 329 were deemed potentially eligible and reviewed at abstract. A total of 154 full text papers were reviewed following abstract screening (Fig. 1). We identified 34 articles eligible for inclusion (4 conference abstracts and 30 papers), of which 16 were subsequently excluded. Nine studies were excluded because they included data also reported in a larger study (i.e. large national database studies) [6, 13, 19, 29,30,31,32,33,34]. Four studies were deemed ineligible on further review. We contacted the authors of three papers: two were excluded as we could not obtain patient numbers, and one paper contained reporting errors which the authors were unable to resolve.
We included 18 studies (14 papers and 4 conference abstracts) in the meta-analysis [1, 2, 5, 8,9,10,11, 15,16,17,18, 35,36,37,38,39,40,41]. The 18 studies (9 multicentre and 9 single centre) included 1,191,178 patients. The characteristics of the included studies are shown in Table 1. Study size ranged from 296 to 263,082 patients. Study duration varied between 5 months and 9 years. ICU admission periods spanned 1994–2014. Nine papers reported both mortality and readmission, seven reported mortality only and two readmission only.
Definitions of out-of-hours varied, starting between 16:00 and 22:00 and ending between 05:59 and 09:00. Two studies [2, 37] performed more than one analysis using different definitions of out-of-hours. As 13 of the other 16 studies defined out-of-hours as commencing between 18:00 and 22:00, we selected the definition starting between these times for inclusion in this analysis (Table 1). All studies presented data for the same time periods at the weekend as in the week.
Five of seven studies that compared illness severity between in-hours and out-of-hours discharges found significantly higher severity of illness at admission in the out-of-hours group (Table 2) [1, 2, 8, 10, 35]. Different measures of illness severity were used, preventing pooling of data. Two of eight studies that compared age between in-hours and out-of-hours discharge found significant differences (both Australasian studies finding patients discharged at night to be slightly younger) [1, 8]. None of the five studies that compared gender between in-hours and out-of-hours found significant differences. The absence of data in many of the included studies, combined with the different measures of illness severity used prevented post hoc analysis to investigate whether differences between in-hours and out-of-hours populations accounted for differences in outcome.
We included 16 studies containing data on 927,046 patients in the mortality analysis. Figure 2 shows the association between out-of-hours ICU discharge and mortality. The pooled relative risk estimate for discharge at night (95% CI) was 1.39 (1.24, 1.57), p < 0·0001. Out-of-hours discharge was associated with significant increases in in-hospital mortality for all definitions of out-of-hours (supplementary material, Fig. 1). Overall heterogeneity was high (I2 statistic 90.1%), mainly arising from differences in the size (rather than the presence and direction) of the effect in studies defining out-of-hours commencing 18:00–21:59.
The effect of out-of-hours discharge on mortality remained for four of the five geographical areas: UK [relative risk (RR) 1.41 95% CI 1.27, 1.57]; Australasia (RR 1.65, 95% CI 1.40, 1.94); Europe (RR 1.38, 95% CI 1.08, 1.76); and United States of America with South America and Canada (RR 1.31, 95% CI 1.23, 1.40). Asia included only one small study and found no effect (RR 0.41, 95% CI 0.10, 1.63) (supplementary material, Fig. 2). Discharge out-of-hours remained significantly associated with subsequent in-hospital mortality in six of eight included studies that undertook multivariate analysis (Table 3) [1, 2, 8, 10, 15, 35].
We included 11 studies, including 1,156,904 patients in the ICU readmission analysis. Figure 3 shows the association between out-of-hours discharge and readmission to an ICU. The pooled risk estimate for discharge out-of-hours (95% CI) was 1.30 (1.19, 1.42). Heterogeneity was high (I2 statistic 90.2%). Heterogeneity arose from differences in effect size rather than the presence or direction of effect [42].
The effect of out-of-hours discharge on readmission remained when analysed for studies in Australasia (RR 1.18, 95% CI 1.09, 1.28), Europe (RR 3.02, 95% CI 2.41, 3.79) and United States of America with South America and Canada (RR 1.14, 95% CI 1.07, 1.21). The effect in the UK was borderline (RR 1.42, 95% CI 1.00, 2.02) (supplementary material, Fig. 3).
Table 3 shows studies that adjusted for potential confounders. We show the confounders used (for which there was no consensus). The summary adjusted odds ratio (95% CI) for mortality was 1.33, (1.30, 1.36), p < 0.001. For comparison, the unadjusted odds ratio was 1.33, (1.28, 1.62), p < 0.001. Analysing only studies that adjusted for potential confounders reduced heterogeneity (the eight studies reporting adjustment tended to be larger studies). One study undertook multivariate adjustment for readmission (out-of-hours discharge remained significant) [38]. We were unable to perform planned sub-group analyses of discharge destination and palliation status due to inconsistent reporting of these data. Too few studies in each group meant we were unable to perform sub-group analysis of out-of-hours definition.
Funnel plots and Egger’s regressions for the effect of out-of-hours discharge on mortality and readmission are shown (supplementary material, Figs. 4, 5, 6, 7). Both funnel plots and Egger’s regression suggest there may be some publication bias whereby studies showing a strong association between mortality and out-of-hours discharge, particularly smaller studies, are not published (p = 0.014). This was not as obvious for studies of readmission (p = 0.057), but this may have been due to a smaller sample of studies.
Quality assessment findings are shown in supplementary material, Table 2. Most studies scored well, between seven and nine out of nine. However, only five studies defined whether the two patient groups were discharged from ICU to a ward or higher dependency area. To assess the influence of each study on bias, we omitted each study in turn (supplementary material, Figs. 8, 9). Removal of any individual study did not remove the effect for either mortality or readmission. The largest effect for mortality occurred when removing a study including the majority of Australasian ICUs [1] reducing the RR (95% CI) to 1.36 (1.26, 1.47). For ICU readmission, both funnel and regression plots supported removal of two major outliers [9, 38]. Removing the outliers reduced the heterogeneity but did not significantly change the RR (supplementary material, Figs. 6, 7, 8, 9).