This retrospective observational study used data from the MIMIC-IV (V.2.2) database, which contains high-quality information on ICU admissions at Beth Israel Deaconess Medical Center from 2008 to 2019. Access to MIMIC-IV is limited to certified individuals; one author (JQ) with access permission (record ID 60311526) extracted the data. As all patient records in MIMIC-IV were fully de-identified, the institutional review board at Beth Israel Deaconess Medical Center waived the need for individual patient consent.

We initially identified patients with recorded ICU admissions in the MIMIC-IV 2.2 database. From this population, we selected those admitted to the ICU via ED. Patients were excluded if they lacked recorded ED registration time (edregtime) or if there was a time difference exceeding 6 hours between hospital admission (admittime) and ICU admission (intime), as these cases were not directly transferred from the ED to the ICUs and may have been admitted to a general ward before ICU transfer. For patients with multiple hospital admissions, only the first hospitalisation with an ICU stay was analysed. After applying these criteria, the final study cohort was divided into four quartiles based on the ED to ICUs time.

All baseline characteristics were extracted within the first 24 hours of ICU admission from the MIMIC-IV database, including age, gender, ethnicity, admission type, first care unit, chronic conditions, severity scores (Charlson Comorbidity Index (CCI) score, Acute Physiology Score III (APSIII)) and intervention during ICU. The ED to ICUs time was defined as the interval between edregtime (the time of registration in the ED) and intime (the time of transfer into the ICU) recorded in the MIMIC-IV database. Admittime provides the date and time the patient was admitted to the hospital. Data extraction was performed using pgAdmin4 PostgreSQL 16.1.

The primary outcome of this study was in-hospital all-cause mortality. Of these, hospital death included ICU death and general ward inpatient death. The secondary outcomes included 28-day all-cause mortality, LOS in ICU and LOS in hospital. Patient mortality information for discharged patients was accessed from the US Social Security Death Index.

Patients and/or the public were not involved in this study.

Our study divided patients into four quartiles based on their ED to ICUs admission times. We assessed the distribution of continuous data using the Kolmogorov-Smirnov test and presented them as means with SD or medians with IQRs. Categorical data were shown in counts and percentages. The Mann-Whitney U test and the Pearson χ² or Fisher’s exact tests were used for comparisons.

We used logistic regression models to estimate ORs and 95% CIs for hospital and ICU mortality. Analyses were conducted using three models: an unadjusted model, a model adjusted for illness severity alone and a fully adjusted model addressing potential multicollinearity. After fitting the logistic models, pairwise comparisons of the four quartiles were performed, and we applied Tukey’s method to adjust for multiple comparisons. The fully adjusted model included a comprehensive set of covariates to assess the relationship between ED to ICUs time and in-hospital mortality, while controlling for key confounders. Age and sex were included to account for demographic differences, while the APSIII and CCI scores provided a detailed assessment of acute illness severity and chronic comorbidity burden, respectively. Additional covariates included diabetes, congestive heart failure, renal disease and malignancy, given their known impact on critically ill patients’ outcomes. Adjustments were also made for invasive mechanical ventilation, dialysis and vasopressor use, as these reflect the severity of respiratory failure, renal impairment and haemodynamic instability. This selection of covariates aimed to minimise residual confounding and precise effect estimation of ED to ICUs time on in-hospital mortality. Secondary outcomes, including 28-day mortality, were analysed using similar models and Kaplan-Meier survival analysis, with additional tests for non-linear relationships. Restricted cubic spline (RCS) models were applied to further explore the non-linear association between ED to ICUs time and both in-hospital and ICU mortality.

To examine potential effect modification, we conducted subgroup analyses based on age, gender, APSIII score, CCI score and the presence or absence of key comorbidities. Interaction terms between ED to ICUs time and each subgroup variable were incorporated into the logistic regression model to assess whether the association between ED to ICUs time and mortality varied across subgroups. Additionally, we performed stratified analyses, fitting separate logistic regression models within each subgroup to estimate the adjusted ORs and 95% CIs. The significance of effect modification was determined using the p value for interaction from the fully adjusted model.

Generalised Boosted Models (GBMs) were used to estimate generalised propensity scores to achieve covariate balance among the four quartiles of ED to ICUs admission times (Q1, Q2, Q3, Q4). The GBM was implemented using the twang R package with the following hyperparameters: 10 000 boosting iterations, an interaction depth of 3 and a shrinkage parameter of 0.01. The GBM estimated the average treatment effect for all four quartiles, ensuring comparability between groups. Covariates included age, gender, race, first care unit, CCI scores, APSIII scores, diabetes, congestive heart failure, renal disease, malignancy, invasive mechanical ventilation, dialysis and vasopressor use. The balancing process was guided by minimising the maximum standardised effect size across all covariates. Balance diagnostics were conducted using standardised mean differences (SMDs) and Kolmogorov-Smirnov statistics to evaluate the effectiveness of the model in achieving covariate balance before and after adjustment. Sample weights from the GBM were used to estimate effective sample sizes (ESSs) for quartiles.

A sensitivity analysis was also performed by reintroducing patients with a time difference greater than 6 hours between hospital admission and ICU admission. This analysis aimed to assess the robustness of the study findings by evaluating whether the association between ED to ICUs time and in-hospital mortality persisted when including patients potentially admitted to non-ICU wards before ICU transfer.

We performed statistical analyses with R software (V.4.3.3) and considered two-sided p values below 0.05 as significant.

From 50 920 unique patients with recorded ICU admissions in the MIMIC-IV database, we identified 15 246 adults admitted directly from the ED to the ICU during their first hospitalisation. After excluding those without recorded ED registration times and those with a hospital-to-ICUs admission interval exceeding 6 hours (n=2432), the final analysis cohort comprised 12 703 patients. The selection process is detailed in figure 1. Online supplemental table 1 summarised the characteristics of 12 703 study participants categorised by ED to ICUs time quartiles. The median patient age was 64 years, and 56.51% were men. There were 1185 ICU deaths, 1988 in-hospital deaths, and 2490 deaths within 28 days. The median ED to ICUs time was 3.98 hours. Among the chronic conditions in the cohort, diabetes was present in 2938 patients (23.13%), congestive heart failure in 2401 (18.90%), renal disease in 1753 (13.80%) and malignant cancer in 1090 (8.58%). The median age increased from 61 years in Q1 to 66 years in Q3 and Q4, with an overall median of 64 years (p<0.01). Male patients comprised 56.51% of the study population, with their proportion decreasing from 61.86% in Q1 to 51.43% in Q4 (p<0.01). The CCI score exhibited a statistically significant difference across quartiles, with a median (IQR) ranging from 4 (2–6) in Q1 to 5 (3–7) in Q4 (p<0.01). Similarly, the APSIII score varied significantly among groups, with a median (IQR) decreasing from 40 (30–60) in Q1 to 39 (30–54) in Q4 (p<0.01).

In terms of outcomes, hospital mortality occurred in 1988 patients (15.65%), with rates ranging from 17.60% in Q1 to 12.21% in Q4 (p<0.01). ICU mortality was observed in 1185 patients (9.33%), with the highest rate in Q1 (11.64%) and the lowest in Q4 (6.37%) (p<0.01). The 28-day mortality rate was 19.60% overall, varying from 20.64% in Q1 to 16.85% in Q4 (p<0.01). The median hospital LOS was 5.0 days (IQR: 2.8–9.4) across all patients, with a slight difference among quartiles, from 5.0 days in Q1 to 5.3 days in Q4 (p=0.05). The median ICU LOS was 1.80 days, ranging from 1.96 days in Q1 to 1.70 days in Q4 (p<0.01), indicating significant differences between groups.

Following the application of a GBM and propensity score weighting, we achieved significant enhancement in baseline comparability across the ED to ICUs time quartiles, as shown in online supplemental table 2, where demographic characteristics, first care unit distribution, severity scores and other covariates demonstrated improved balance across groups. As demonstrated in online supplemental figure 1 and online supplemental table 3, the maximum SMD was reduced from 0.56 before weighting to 0.04 after weighting, and the maximum Kolmogorov-Smirnov statistic improved from 0.20 to 0.02, with the minimum Kolmogorov-Smirnov p value increasing from <0.01 to 0.22. These improvements in covariate balance were achieved through iterative optimisation, as illustrated in online supplemental figure 2. ESSs remained robust post-weighting: for instance, Q1 retained an ESS of 2651.1, and Q4 achieved 2475.5, ensuring sufficient analytical power for subsequent modelling (online supplemental table 4).

Under these more balanced conditions, three weighted logistic regression models were constructed to assess the association between ED to ICUs time and all-cause mortality (table 1). In Model 1, which incorporated weighting without additional adjustments, patients in the longest ED to ICUs time quartile (Q4) already exhibited significantly lower odds of in-hospital mortality compared with those in Q1 (OR 0.80, 95% CI: 0.74 to 0.86, p<0.01) (figure 2). Further adjustment for illness severity (Model 2) and comprehensive adjustment for severity scores, comorbidities, demographics, first care unit type and ICU-level interventions (Model 3) reinforced this inverse relationship. In the fully adjusted Model 3, the odds of in-hospital mortality in Q4 relative to Q1 were 0.75 (95% CI: 0.69 to 0.82, p<0.01), while ICU mortality odds were 0.63 (95% CI: 0.61 to 0.74, p<0.01). Variance inflation factor analyses demonstrated no significant collinearity issues within Model 3, supporting the stability and interpretability of the final model (online supplemental figure 3).

Building on these findings, RCS analyses revealed a gradual, non-linear decline in both in-hospital and ICU mortality risk as ED to ICUs time increased, with a more pronounced reduction observed after 5.65 hours (online supplemental figure 4).

The 28-day survival was further evaluated as a secondary outcome using Kaplan-Meier estimates (figure 3). Throughout the follow-up period, Q4 consistently exhibited a higher cumulative survival rate compared with Q1, as well as Q2 and Q3. While the survival differences between Q1 and the intermediate quartiles were not statistically significant (p=0.60 and p=0.95, respectively), Q4 demonstrated a significantly improved 28-day cumulative survival compared with Q1 (p=0.03).

Online supplemental table 5 showed the relationship between ED to ICUs time and both hospital and ICU LOS. For hospital LOS, the unadjusted model indicated a significant negative association with ED to ICUs time (Coef.=−0.099, p=0.002); however, this association lost significance after adjustment in Models 2 and 3 (p>0.05). In contrast, for ICU LOS, a longer ED to ICUs time was consistently associated with a shorter duration of stay across all models, with coefficients of −0.157 (p<0.01) in Model 1, −0.097 (p<0.01) in Model 2 and −0.069 (p<0.01) in Model 3. These results indicated a sustained inverse relationship between ED to ICUs time and ICU LOS, even after adjustments.

The protective association between longer ED to ICUs time and reduced in-hospital mortality risk was broadly consistent across most examined subgroups in the subgroup analyses of hospital mortality, with ORs remaining below 1 (figure 4A). However, the magnitude of this protective effect differed by specific patient characteristics. Significant interactions were observed with gender (p=0.03), CCI score (p<0.01) and malignant cancer (p=0.01), indicating that the strength of the inverse association increased among patients with higher comorbidity burdens or malignancies and varied by gender. Similar trends were noted for ICU mortality, with ORs generally showing a reduction in risk with prolonged ED stays across most subgroups (figure 4B). Although the interaction with gender did not reach significance (p=0.06), CCI score (p<0.01) and congestive heart failure (p=0.03) emerged as significant effect modifiers, suggesting potential variability in the association depending on the severity of comorbidities. These findings underscored the robustness of the protective association between longer ED to ICUs time and improved survival, while highlighting certain patient characteristics that may influence this relationship.

In the sensitivity analysis, patients initially excluded due to a time gap of more than 6 hours between hospital admission (admittime) and ICU admission (intime) were reintroduced to evaluate the impact of this criterion on the study results. online supplemental table 6 presented the baseline characteristics of this expanded cohort. After applying GBM and propensity score weighting, baseline balance remained markedly improved in this expanded cohort (online supplemental table 7). As shown in online supplemental table 8, the maximum SMD decreased from 0.64 before weighting to 0.06 after weighting, and the maximum Kolmogorov-Smirnov statistic declined from 0.22 to 0.03, with the minimum Kolmogorov-Smirnov p value increasing from <0.01 to 0.07. Correspondingly, the ESSs remained robust after adjustment, as detailed in online supplemental table 9. Online supplemental figure 5 further illustrated the substantial reduction in SMDs following weighting. The ORs from the logistic regression models (online supplemental table 10 and online supplemental figure 6) continued to indicate lower odds of in-hospital and ICU mortality for patients in Q4 compared with those in the other groups. Moreover, Kaplan-Meier survival curves (online supplemental figure 7) showed the highest 28-day survival rates for Q4, with Q1 versus Q4 comparisons remaining significant (p<0.01). These findings suggested that the exclusion of patients with longer pre-ICU stays did not materially affect the relationship between extended ED stays and better survival outcomes, thereby reinforcing the robustness of the results.

This study has several limitations. First, as a retrospective analysis drawn from the MIMIC database, these findings cannot establish causality and may not be generalisable to other countries. Second, we excluded patients who died in the ED to focus on those successfully transferred to the ICUs. Although necessary, this may introduce selection bias by underrepresenting the most critically ill and potentially underestimating mortality associated with shorter ED to ICUs time. Third, the MIMIC database does not record critical interventions performed in the ED and records surgeries with daily resolution rather than hourly granularity. This limited temporal detail prevents determining whether interventions preceded ICU admission, complicating the interpretation of ED to ICUs time while possibly leading to an underestimation of pre-admission mortality. Lastly, the retrospective design prevented a detailed examination of the components influencing ED to ICUs time, including orientation decisions, clinical assessments, diagnostic workups and boarding durations. Future research should employ prospective, multicentre designs with more comprehensive and temporally precise data collection to better elucidate the impact of ED to ICUs time on patient outcomes.

Patients transferred from the ED to the ICUs within 5.65–34.55 hours experience lower in-hospital, ICU, and 28-day mortality rates compared with those transferred earlier. These findings suggest that appropriately extended ED stays may facilitate critical stabilisation, improving outcomes and emphasising the need for further research to optimise transfer protocols and ED management strategies.

Not applicable.

Given that this study analysed an anonymous public database with prior Institutional Review Board approval, no ethical review was necessary.