We conducted a retrospective study based on the MIMIC-IV database (version 1.0) (15). This database is an updated version of MIMIC-III with preexisting Institutional Review Board approval (Massachusetts Institute of Technology, no. 0403000206; Beth Israel Deaconess Medical Center, 2001P001699) (16). Several improvements have been made, including simplifying the structure, adding new data elements, and improving the usability of previous data elements. Currently, the MIMIC-IV contains comprehensive and high-quality and de-identified data of patients admitted to the ICU or emergency department of the Beth Israel Deaconess Medical Center between 2008 and 2019. One author who has finished the Collaborative Institutional Training Initiative examination (certification number: 39050603 for author Jhou) can access the database and was responsible for data extraction.

The patients were identified in the MIMIC-IV database from 2008 to 2019. The inclusion criteria were as follows: adult patients (age, 18–89 years) with ischemic stroke, defined as ICD-9 codes of 433, 434, 436, 437.0, and 437.1 or ICD-10 codes of I63, I65, and I66. Patients with ischemic stroke who received acute reperfusion therapy, such as intravenous tissue plasminogen activator or endovascular mechanical thrombectomy were also enrolled in our analysis (17). The exclusion criteria were as follows: patients who were admitted to the ICU more than once, only data on the first ICU admission were recorded and used. The patients enrolled in this study were subsequently divided into the AF and non-AF groups.

The following baseline characteristics were identified: hypertension, hyperlipidemia, diabetes mellitus, coronary artery disease, congestive heart failure, peripheral vascular disease, chronic obstructive pulmonary disease, liver disease, peptic ulcer disease, chronic kidney disease, rheumatoid arthritis, dementia, malignancy, Acute Physiology Score III (APS III), CHA2DS2-VASc score, and HAS-BLED score (Supplementary Tables 1, 2) (18). The overall Charlson Comorbidity Index (CCI) encompassed 18 categories of medical conditions, which were identifiable in medical records (Supplementary Table 3) (19–21). Secondary prevention agents for ischemic stroke were also recorded and used, including antiplatelet agents (e.g., aspirin, clopidogrel, cilostazol, ticlopidine, ticagrelor, prasugrel, and dipyridamole) and anticoagulation agents (e.g., warfarin, dabigatran, rivaroxaban, apixaban, and edoxaban). Other clinical indicators included mean arterial pressure, heart rate, body temperature, saturation of peripheral oxygen, leukocyte count, hemoglobin, platelet, creatinine, blood urea nitrogen, sodium, potassium, and bilirubin within 24 h of ICU admission. When the aforementioned indicators had multiple results within 24 h, the worst value was recorded and used for the analysis.

The major outcomes were ICU mortality and in-hospital mortality. The minor outcomes included the use of percutaneous endoscopic gastrostomy or jejunostomy tube placement and the complications of intracerebral hemorrhage.

Categorical variables were represented as number (percentage) and were compared using the chi-square test and Fisher's exact test. Continuous variables were described as means (standard deviation) and were compared using the independent samples t-tests or Wilcoxon rank-sum test (Mann–Whitney U-test). Missing values of each subjects were not defaulted to negative, and denominators were only reported cases.

Propensity scores were calculated involving the following preoperative variables: age, sex, CCI, APS III, CHA2DS2-VASc score, and HAS-BLED score. A logistic regression model was developed to estimate the patients' propensity scores for the AF group (22). PSM was performed using the Greedy 5-to-1 Digit-Matching algorithm between the AF and non-AF groups (23). In propensity-matched patients, univariate analyzes were conducted using the paired t-test or Wilcoxon signed-rank test for continuous variables and McNemar's test for categorical variables. Statistical testing was performed to evaluate the effectiveness of PSM. Both primary and secondary outcomes were compared based on the matched data. Sensitivity analysis was performed using the removal of new-onset AF (Supplementary Table 4).

For time-to-event outcomes, the major survival outcomes, the times that elapsed until the first event between the two groups were compared using the log-rank test, whereas the Kaplan–Meier method was used to estimate the absolute risk of each event for each group. Univariate and multivariate Cox hazards model analyzes were performed to identify the association between AF and major outcomes, and results were expressed as hazard ratio (HR) with 95% confidence interval (CI). The relationship between AF and minor outcomes was assessed using univariate and multivariate logistic regression analyzes, and results were expressed as odds ratio (OR) with 95% CI.

All comparisons were planned, the tests were 2-sided, and p < 0.05 were used to denote statistical significance. Statistical analyzes were performed using Statistical Package for the Social Sciences (version 25.0; IBM Corp., Armonk, NY, USA) and R (version 4.0.3; R Foundation for Statistical Computing, Vienna, Austria).

Of the 257,366 medical records reviewed, 50,048 patients were admitted to the ICU. During the study period, we enrolled 1,662 patients with stroke, including 653 patients with AF and 1,009 patients without AF (Figure 1). There were 15 new-onset AF of those without history of AF (1.46%, 15/1,024). The basic demographic characteristics of the patients are shown in Table 1. The mean age of the AF and non-AF groups were 75.52 and 64.47 years, respectively. The prevalence of AF increased in a stepwise manner with advancing age, from 7.8% in those aged 50 years or younger to 63.5% in those older than 80 years (Figure 2). The AF group had higher CCI (7.66 vs. 6.35; p < 0.001), CHA2DS2-VASc score (5.49 vs. 4.60; p < 0.001), HAS-BLED score (4.02 vs. 3.53; p < 0.001), and APS III (44.27 vs. 35.77; p < 0.001) and had more comorbidities, including hyperlipidemia (25.4% vs. 19.8%; p = 0.007), congestive heart failure (30.9% vs. 11.6%; p < 0.001), chronic kidney disease (18.4% vs. 11.7%; p < 0.001), and dementia (5.7% vs. 3.4%; p = 0.024) (Table 1).

Propensity scores were calculated involving the following covariates: age, sex, CCI, APS III, CHA2DS2-VASc score, and HAS-BLED score. After 1:1 PSM, 576 pairs in each arm remained. The cohorts were well-balanced based on six covariates between the AF and non-AF groups (Figure 3).

The AF group had a higher risk of ICU mortality (11.5% vs. 6.4%; p < 0.001) and in-hospital mortality (19.1% vs. 10.1%; p < 0.001) than the non-AF group (Table 1). The Kaplan–Meier curves for ICU discharge and survival between the AF and non-AF groups are shown in Figures 4A,B, and both of these curves were significantly different. However, no difference in hospital discharge and survival was observed between patients with and without AF (Figures 4C,D). Survival was followed until hospital discharge, and the longest length of hospital stay was 88 days.

In the cohort before PSM, univariate Cox regression analysis revealed significant difference in ICU mortality (crude HR, 1.43; 95% CI, 1.03–2.00) and in-hospital mortality (crude HR, 1.61; 95% CI, 1.24–2.09); furthermore, patients with AF had a higher risk of intracerebral hemorrhage (crude OR, 2.17; 95% CI, 1.61–2.94) and percutaneous endoscopic gastrostomy (PEG) or percutaneous endoscopic jejunostomy (PEJ) placement before PSM (crude OR, 2.01; 95% CI, 1.50–2.70).

After PSM, the results of univariate Cox regression analysis showed that AF was not associated with higher ICU mortality (HR, 0.94; 95% CI, 0.64–1.37) and in-hospital mortality (crude HR, 0.95; 95% CI, 0.71–1.27). In minor outcomes, patients with AF remained a higher risk of intracranial hemorrhage and PEG/PEJ tube placement with crude ORs of 1.83 (95% CI, 1.27–2.62) and 1.89 (95% CI, 1.33–2.68), respectively.

After multivariate analysis, after adjusting for race, hypertension, congestive heart failure, liver disease, diabetes, anti-platelet agents, anti-coagulation agents, intravenous tissue plasminogen activator or endovascular mechanical thrombectomy, no differences in ICU mortality (adjusted HR, 0.95; 95% CI, 0.64–1.42), in-hospital mortality (adjusted HR, 1.08; 95% CI, 0.79–1.47). However, the risk of intracerebral hemorrhage (adjusted OR, 1.96; 95% CI, 1.31–2.92), and PEG or PEJ tube placement before PSM (adjusted OR, 1.76; 95% CI, 1.19–2.60) remained higher among patients with AF than those without AF. The results were shown in Table 2. In the sensitivity analyses, we noted the similar findings before and after PSM cohort, as well as adjusted PSM cohort (Supplementary Table 4).

The main strength of this study is its large-scale, diverse population study design using real-world data. However, the results should be interpreted in the context of the following limitations. First, the study design was retrospective, and the diagnosis of ischemic stroke relied solely on administrative diagnosis codes. We could not confirm the diagnostic accuracy by evaluating the patients directly. Therefore, misclassifications could lead to false associations. Second, although we adjusted as much bias as we could by using PSM and multivariable analysis to balance the baseline differences between the groups and eliminate residual confounding effects, biases related to unmeasured confounders remain a potential issue in this study. Third, given the nature of the MIMIC database, we lacked some potential factors on the National Institute of Health Stroke Scale, the subtypes of ischemic stroke (TOAST classification), onset time of stroke, the timing of AF diagnosis, the categorization of AF type, cardiac function parameters, and the mortality cause. No long-term follow-up events remained; therefore, the 3-month modified Rankin Scale score cannot be measured. Fourth, this was a single-institution study, and selection bias of patients might exit. Because of a narrow window time after symptom onset of ischemic stroke for reperfusion therapy and the evolution of stroke severity, some patients might not be referred from other hospitals. These high-risk patients might not have been included in our cohort. Finally, although novel OAC (NOAC) and warfarin were used for the secondary prevention of stroke in patients with AF, we could not obtain the actual and detailed information or reasons as to why physicians choose to prescribe warfarin or NOACs based on the administrative database, and thus, we could not add these potential factors into the propensity score models or adjust for these factors in our regression models. Furthermore, we enrolled the patients with AF and diagnosed as ischemic stroke in the intensive care unit, and the phase of the stroke diagnosis was not clear. Whether the patient was under an acute stroke phase might have an influence on the dissociation of poor clinical outcome and in-hospital mortality, which would be an another limitation to our study. Because this study focused on the ICU setting, further studies are warranted to examine the external generalizability of our results.