Database analysis part 2: intensive care unit databases

Parts of this chapter are adapted with permission from Bedford et al.81 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The text below includes minor additions and formatting changes to the original text.

The analysis of the RISK-II database shows that patients who develop NOAF during an ICU admission are at a significantly increased risk of mortality and hospital readmission with AF, heart failure and stroke. This highlights the importance of optimal management of this common problem.

Our scoping review has shown that the existing evidence for the best management of NOAF acquired in an ICU is limited. Common concerns included small sample sizes and inadequate adjustment for differences between treatment groups.

We, therefore, aimed to compare the clinical effectiveness of different NOAF treatments by analysing two large ICU databases after performing comprehensive adjustments for measured confounding.

Database analysis part 2: methods

Study design

We carried out a retrospective observational study of two large ICU databases from the UK (PICRAM82) and the USA [Medical Information Mart for Intensive Care III (MIMIC-III) v1.483].

The PICRAM database comprises data relating to > 12,000 patients who were treated in three general ICUs in the UK from 2008 to 2015. MIMIC-III comprises data relating to > 40,000 patients who were admitted to critical care units at a tertiary care hospital in the USA between 2001 and 2012.

All analyses were performed on each database individually given the potential for differences in case mix and interventions. Combined analyses were performed for each outcome to confirm findings. We adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.84

Study population

We included all adult (aged ≥ 16 years) patients. For patients who were admitted more than once to an ICU, we used their first admission. We excluded patients:

• cared for by a coronary care or cardiac surgery team
• with missing hospital outcome data
• with an ICU length of stay of < 24 hours85
• with significant arrhythmia in the first 3 hours of arrival to an ICU
• with pre-existing arrhythmias.

We defined patients as having a pre-existing arrhythmia if an arrhythmia or a medication prescribed with an indication of heart rhythm management were identified in the patient’s medical history.

Exposure and outcomes

For all ICUs that were included in the databases, routine practice was to display three lead electrocardiograms continuously, with heart rhythm recorded by the bedside nurse at regular intervals. We defined NOAF as the documentation of AF or atrial flutter lasting for ≥ 30 seconds.30 A documented heart rhythm was assumed to persist until the next identifiable rhythm was recorded.

The availability of data relating to the interventions of interest was assessed in the MIMIC-III and PICRAM databases. The MIMIC-III database allowed four interventions to be compared: i.v. amiodarone, i.v. beta-blockers, i.v. calcium channel blockers and electrical cardioversion. The PICRAM database allowed three interventions to be compared: i.v. amiodarone, i.v. beta-blockers and i.v. digoxin. We analysed each intervention in an intention-to-treat fashion, in which treatment groups were determined by first treatment after NOAF onset.86

Primary outcomes

The primary end points of this study were ICU mortality, hospital mortality, rate control and rhythm control, which were analysed as time-to-event outcomes. We censored rate and rhythm control at 24 hours and censored mortality at 30 days. In the absence of a consensus definition of treatment success for ICU-acquired NOAF, we defined time to cardioversion with our expert panel as the time to first reversion of sinus rhythm42 and we defined time to rate control as the time to a heart rate of < 110 b.p.m. in the subset of patients with a heart rate of ≥ 110 b.p.m.87

Secondary outcomes

We analysed the association of NOAF with hospital mortality. We calculated the CHR to estimate the aetiological association between NOAF and hospital outcomes, considering hospital discharge as a competing risk to mortality. We adjusted for the Oxford Acute Severity of Illness Score (OASIS),88 which was limited to the first 3 hours of ICU admission to avoid confounding post-NOAF onset. The use of early scores has been shown to remain predictive of outcome.89

We explored the haemodynamic changes (heart rate, blood pressure and vasoactive medication dose) that are associated with NOAF. We calculated the proportion of patients receiving vasoactive medications in our cohort before and after NOAF onset. We calculated the vasoactive-inotropic score90 to quantify the change in composite dose of vasoactive medications for patients already receiving vasoactive medications prior to AF onset.

Focusing on the period 6 hours pre and post NOAF, we used smooth additive quantile regression models91 to fit the 75%, 50% and 25% quantiles of the distribution of each haemodynamic variable. We excluded the haemodynamic data recorded after each patient’s first treatment for NOAF to establish a natural history. All models included a binary covariate to indicate the onset of NOAF and allowed for changes in smoothing spline post-AF onset.

We used multilevel linear models to test whether or not there were significant changes in heart rate, blood pressure and vasoactive medication dose associated with NOAF. Each model included fixed linear segmented regression terms, with a random effect per patient to account for repeated measurements.

Adjustment for confounding

We carried out a propensity score-weighted time-to-event analysis to adjust for measured confounding in the selection of patients between treatment groups. All statistical analyses were performed using R Core v4.0.2. We generated propensity score weights that were optimised to balance the covariate distributions of the treatment groups92 using the WeightIt package.93 The confounding variables included admission variables, laboratory variables and physiological variables adjacent to NOAF onset. The list of confounding variables was generated based on the studies identified in our scoping review. This list was then reviewed and supplemented by members of the study oversight panel. The admission variables included age; sex; the OASIS88 within the first 3 hours of ICU admission; use of preadmission beta-blocker, antipsychotic or thyroid medication; severe congestive cardiac failure; chronic obstructive pulmonary disease (COPD); liver disease; and thyroid disease. The laboratory variables at NOAF onset included the most recent (to NOAF onset) plasma sodium, potassium, magnesium, creatinine and urea concentrations; white cell count; platelet count; haemoglobin concentration; and prothrombin time. The physiological/intervention variables at NOAF onset included systolic and mean blood pressure, heart rate, body temperature, presence and dose of vasoactive agent, presence of bronchodilator therapy, mechanical ventilation, renal replacement therapy and presence of central venous access.

We assessed the balance of covariates across weighted groups by tabulating group means pre and post weighting. We calculated standardised mean differences (SMDs)94 and the maximum SMD of all pairwise treatment group comparisons for each covariate.

We carried out a weighted Cox survival analysis to determine the average treatment effect of NOAF treatments on our outcomes of interest. Missing laboratory values were handled by using multiple imputation. We generated 20 imputed data sets. To account for the uncertainty in the generated propensity scores and to allow for the estimation of 95% CIs around effect estimates, we performed resampling with replacement (bootstrapping) with recalculation of propensity score weights and effect estimates with each bootstrap sample. We obtained 1000 bootstrap samples from each imputed data set.95 The effect estimates and CIs from each imputed data set were combined using Rubin’s rules.96

Critical Care Health Informatics Collaborative database analysis

We also analysed the Critical Care Health Informatics Collaborative (CCHIC) database.97 This database was created with retrospectively collected detailed data from the ICU clinical information systems from four general ICUs in London and Cambridge, in the UK, from 2014 to 2018.

Of our drugs of interest, the CCHIC database contains beta-blocker data only. We, therefore, decided to use this database to analyse only the epidemiology and characteristics of NOAF to compare with our main analyses.

We used the eligibility criteria stated in Study population. However, we were unable to exclude patients with documented pre-existing arrhythmias because these data were not available in the CCHIC database. Pre-existing arrhythmia was, therefore, determined only by the presence of arrhythmia during the first 3 hours of ICU admission. Full methods are outlined in Appendix 7.

Database analysis part 2: results

Study population

The MIMIC-III database contains data from 22,684 adult index ICU admissions. Of these patients, 220 had an ICU length of stay of < 24 hours. We identified 3905 of the remaining 22,464 patients as having pre-existing AF or AF documented within the first 3 hours of their ICU admission. Of the 18,559 patients who fulfilled our inclusion criteria, 1065 (5.7%) developed NOAF during their ICU stay. Of these patients, 742 went on to receive one of the interventions of interest. Only two patients received digoxin as their initial treatment and were, therefore, excluded, leaving 740 patients for the comparative analysis. This process is displayed in Figure 6.

FIGURE 6 Study CONSORT flow diagrams. (a) MIMIC-III database; (b) PICRAM database. Reproduced with permission from Bedford et al.81 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure. (Figure not reproduced here; see the original article.)

The PICRAM database contains data from 12,270 adult index ICU admissions. Of these patients, 3004 had an ICU length of stay of < 24 hours. We identified 899 of the remaining 9266 patients as having pre-existing AF or AF documented within the first 3 hours of ICU admission. Of the 8367 patients who fulfilled our inclusion criteria, 952 (11.4%) developed NOAF during their ICU stay. Of these patients, 471 went on to receive one of the interventions of interest. Five patients had missing outcome data and were, therefore, excluded from the analysis. Of those 466 patients with outcome data, only six patients received DCC or calcium channel blockers and were, therefore, excluded, leaving 460 patients for the comparative analysis. This process is displayed in Figure 6. In both databases, patients who developed NOAF were older, with a similar age difference as was identified in the RISK-II database analysis. Patients who developed NOAF also had longer ICU and hospital length of stay, and higher ICU and hospital mortality (see Appendix 6, Tables 25 and 26), than those who did not develop NOAF. The characteristics of included patients are displayed in Table 14.

The CCHIC database included data that were related to 33,451 adult first admissions to an ICU. Of these patients, 7889 had an ICU length of stay of < 24 hours. We identified 2713 patients being paced or with another significant arrhythmia during the first 3 hours of ICU admission. Of the remaining 22,849 patients, 1003 had missing hospital mortality data. Of the remaining 21,846 eligible patients, 2618 (12%) developed NOAF (see Appendix 7, Figure 23). The characteristics of patients with and without NOAF are shown in Appendix 7, Table 34.

Characteristics of new-onset atrial fibrillation in treated patients

The time from ICU admission to the onset of NOAF in treated patients was similar between the MIMIC-III and the PICRAM databases [median 40.5 hours (IQR 21–79 hours) vs. 40.3 hours (IQR 41–75 hours), respectively]. Patients with data reported in the MIMIC-III database had, on average, shorter total durations of AF [median 11.6 hours (IQR 4–37 hours) vs. 18.1 hours (IQR 6–44 hours), respectively]. The timing of onset and AF duration data are displayed in Figures 7 and 8, respectively.

FIGURE 7 Time from ICU admission to AF onset in treated patients. (a) MIMIC-III database; (b) PICRAM database. Data from 93 patients with time to AF onset > 168 hours (7 days) not shown. (Figure not reproduced here; see the original article.)

FIGURE 8 Total duration of AF per treated patient. (a) MIMIC-III database; (b) PICRAM database. Data from 89 patients with AF duration > 120 hours (5 days) not shown. (Figure not reproduced here; see the original article.)

Association of new-onset atrial fibrillation with hospital mortality

In the unadjusted analysis, NOAF was associated with an almost identical increased risk of hospital mortality in the MIMIC-III and PICRAM databases [CHR 1.89 (95% CI 1.69 to 2.13) and CHR 1.89 (95% CI 1.68 to 2.13) respectively]. After adjustment for illness severity at ICU admission, this association remained evident [CHR 1.47 (95% CI 1.31 to 1.65) and CHR 1.73 (95% CI 1.54 to 1.96) respectively].

New-onset atrial fibrillation treatments

Of the patients who were identified in the MIMIC-III database, 94 received amiodarone, 473 received beta-blockers, 144 received calcium channel blockers and 29 received electrical cardioversion as their initial NOAF treatment. In the PICRAM database, 344 patients received amiodarone, 47 received beta-blockers and 69 received digoxin as their initial NOAF treatment. The characteristics of patients by treatment group are displayed in Appendix 6, Tables 27 and 28, for the MIMIC-III and PICRAM databases, respectively.

Adjustment for confounding

After propensity score weighting, covariates were well matched across all treatment groups in each database. Across 30 variables, only the mean urea concentration in the MIMIC-III database had the maximum pairwise SMD of > 0.2. Unweighted and weighted means for each treatment group can be found in Appendix 6, Tables 29 and 30.

Rate control

The MIMIC-III database

The time at which 50% of patients had achieved rate control was 60 minutes, 60 minutes, 80 minutes and 10 minutes in the amiodarone, beta-blocker, calcium channel blocker and electrical cardioversion groups, respectively. The cumulative incidence curves of rate control for each treatment group are shown in Appendix 6, Figure 13.

In the unadjusted analysis, no differences were observed between any intervention and amiodarone in the time to achieving rate control (see Appendix 6, Table 31).

After propensity score weighting, beta-blockers, calcium channel blockers and cardioversion were not associated with any significant difference in rate of achieving rate control when compared with amiodarone (HR 1.09, 95% CI 0.78 to 1.51; HR 0.81, 95% CI 0.55 to 1.19; and HR 1.59, 95% CI 0.44 to 5.75; respectively) (Figure 9; see Appendix 6, Table 31).

FIGURE 9 Unadjusted and adjusted HRs for rate and rhythm control. (a) MIMIC-III unadjusted; (b) PICRAM unadjusted; (c) MIMIC-III adjusted; (d) PICRAM adjusted. Adapted with permission from Bedford et al.81 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure. (Figure not reproduced here; see the original article.)

After initial rate control, reversion to a heart rate of ≥ 110 b.p.m. was common. Of those patients achieving rate control, 65%, 62%, 80% and 68% of patients in the amiodarone, beta-blocker, calcium channel blocker and electrical cardioversion groups, respectively, had at least one episode of a heart rate of ≥ 110 b.p.m. within the 24 hours after initial rate control (see Appendix 6, Figure 14).

The PICRAM database

The time at which 50% of patients had achieved rate control was 96 minutes, 115 minutes and 241 minutes in the amiodarone, beta-blocker and digoxin groups, respectively. The unadjusted cumulative incidence curves of rate control for each treatment group are displayed in Appendix 6, Figure 15.

In the unadjusted analysis, there was no evidence of a difference in achieving rate control between beta-blocker therapy and amiodarone (HR 0.85, 95% CI 0.57 to 1.27); however, digoxin appeared inferior to amiodarone in achieving rate control (HR 0.64, 95% CI 0.45 to 0.92).

After propensity score weighting, beta-blocker therapy was not associated with any significant difference in the rate of achieving rate control [adjusted hazard ratio (aHR) 0.82, 95% CI 0.48 to 1.42] compared with amiodarone. The reduced rate of achieving rate control with digoxin therapy remained evident (aHR 0.56, 95% CI 0.34 to 0.92) (see Figure 9 and Appendix 6, Table 32).

After initial rate control, reversion to a heart rate of > 110 b.p.m. was common. Of those achieving rate control, 66%, 59% and 76% of patients in the amiodarone, beta-blocker and digoxin groups, respectively, had at least one episode of a heart rate of ≥ 110 b.p.m. within the 24 hours after initial rate control. These findings are displayed in Appendix 6, Figure 16. These differences were not significant in the unadjusted or adjusted analysis.

Combined database analysis

In the unadjusted combined analysis, beta-blocker therapy was associated with a higher rate of achieving rate control (HR 1.26, 95% CI 1.10 to 1.43) and digoxin was associated with a lower rate of achieving rate control (HR 0.69, 95% CI 0.52 to 0.91) than amiodarone. After adjustment, we found no evidence of a difference between beta-blocker therapy and amiodarone in the rate of achieving rate control (aHR 1.14, 95% CI 0.91 to 1.44). Consistent with the PICRAM database analysis, we found that digoxin therapy was associated with a lower rate of achieving rate control than amiodarone (aHR 0.52, 95% CI 0.32 to 0.86) (see Appendix 6, Table 33).

In the unadjusted combined analysis, calcium channel blockers were associated with an increased rate of reversion to a heart rate of ≥ 110 b.p.m. in those patients who initially achieved rate control (HR 1.62, 95% CI 1.28 to 2.06). This finding bordered on significance after adjustment (HR 1.54, 95% CI 1.00 to 2.37).

Rhythm control

The MIMIC-III database

The time at which 50% of patients had achieved rhythm control was 159 minutes, 144 minutes, 285 minutes and 40 minutes in the amiodarone, beta-blocker, calcium channel blocker and electrical cardioversion groups, respectively (see Appendix 6, Figure 17).

In the unadjusted analysis, there was no evidence of a difference in achieving rhythm control between beta-blockers or cardioversion and amiodarone. Calcium channel blocker therapy was associated with a reduced rate of achieving rhythm control (HR 0.65, 95% CI 0.50 to 0.84).

After propensity score weighting, there remained no evidence of a difference in achieving rhythm control between beta-blockers (HR 0.91, 95% CI 0.61 to 1.35) or cardioversion (HR 2.00, 95% CI 0.86 to 4.65) and amiodarone. Calcium channel blockers remained associated with a lower rate of achieving rhythm control (HR 0.59, 95% CI 0.37 to 0.92) than amiodarone (see Figure 9 and Appendix 6, Table 31).

After initial rhythm control, reversion to AF was common. Of those patients achieving rhythm control, 36%, 39%, 47% and 46% of patients in the amiodarone, beta-blocker, calcium channel blocker and electrical cardioversion groups, respectively, had at least one episode of AF within the 24 hours after initial cardioversion (see Appendix 6, Figure 18). The differences in reversion rates were not significant in the unadjusted or adjusted analysis (see Appendix 6, Table 31).

The PICRAM database

The time at which 50% of patients had achieved rhythm control was 80 minutes, 37 minutes and 255 minutes in the amiodarone, beta-blocker and digoxin groups, respectively. Cumulative incidence curves of rhythm control for each treatment group are shown in Appendix 6, Figure 19.

In the unadjusted analysis, there was no evidence of a difference in achieving rhythm control between beta-blocker therapy and amiodarone (HR 0.95, 95% CI 0.64 to 1.40); however, digoxin appeared to be inferior to amiodarone (HR 0.57, 95% CI 0.41 to 0.81).

After propensity score weighting, beta-blocker therapy was not associated with any significant difference in the rate of achieving rhythm control (HR 0.99, 95% CI 0.57 to 1.72). Digoxin was no longer significantly associated with a lower rate of achieving rhythm control (HR 0.67, 95% CI 0.41 to 1.09) (see Figure 9 and Appendix 6, Table 32).

After initial rhythm control, reversion to AF was common. Of those patients achieving rhythm control, 54%, 51% and 70% of patients in the amiodarone, beta-blocker and digoxin groups, respectively, had at least one episode of AF within the 24 hours after initial cardioversion (see Appendix 6, Figure 20). Differences in reversion rates were not significant in the unadjusted or adjusted analysis (see Appendix 6, Table 32).

Combined analysis

Consistent with the individual database analyses, the unadjusted combined analysis suggested that calcium channel blockers (HR 0.58, 95% CI 0.47 to 0.71) and digoxin (HR 0.58, 95% CI 0.41 to 0.83) were associated with a reduced rate of achieving rhythm control compared with amiodarone. Furthermore, beta-blocker therapy was also associated with a lower rate of achieving rhythm control than amiodarone (HR 0.81, 95% CI 0.71 to 0.93).

After adjustment, we found no evidence of differences between beta-blockers (aHR 0.86, 95% CI 0.67 to 1.11), digoxin (aHR 0.64, 95% CI 0.35 to 1.17) or electrical cardioversion (aHR 1.58, 95% CI 0.71 to 3.51) and amiodarone in the rate of achieving rhythm control. We found that calcium channel blocker therapy was associated with a lower rate of achieving rhythm control than amiodarone (aHR 0.56, 95% CI 0.39 to 0.79), which was consistent with our MIMIC-III database analysis (see Appendix 6, Table 33).

Consistent with the individual database analyses, there was no evidence of a difference in the rates of reversion to AF between treatments in the unadjusted or the adjusted analyses.

Hospital survival

The MIMIC-III database

In the unadjusted analysis, beta-blocker therapy was associated with a reduced hospital mortality rate (HR 0.64, 95% CI 0.44 to 0.93). Unadjusted survival curves for each treatment group in the MIMIC-III database are displayed in Appendix 6, Figure 21.

After propensity score weighting, we found no evidence of a difference between beta-blockers (aHR 1.03, 95% CI 0.53 to 2.03), calcium channel blockers (aHR 1.30, 95% CI 0.61 to 2.76) or electrical cardioversion (HR 0.96, 95% CI 0.31 to 3.01) and amiodarone in hospital mortality (see Appendix 6, Table 31).

The PICRAM database

We found no differences in hospital survival in the unadjusted analyses. Unadjusted survival curves for each treatment group in the PICRAM database are displayed in Appendix 6, Figure 22.

After propensity score weighting, there remained no differences between beta-blockers (aHR 0.75, 95% CI 0.30 to 1.84) or digoxin (aHR 1.37, 95% CI 0.75 to 2.50) and amiodarone in hospital mortality (see Appendix 6, Table 32).

Combined analysis

Consistent with our MIMIC-III database analysis, the combined unadjusted analysis suggested that beta-blockers were associated with a reduced hospital mortality rate (HR 0.78, 0.62 to 0.99). Cardioversion appeared to be associated with a significantly increased hospital mortality rate (HR 1.92, 95% CI 1.16 to 3.17). After propensity score weighting, we found no significant difference between beta-blockers (aHR 0.97, 95% CI 0.56 to 1.68), calcium channel blockers (aHR 1.21, 95% CI 0.62 to 2.39), digoxin (aHR 1.77, 95% CI 0.77 to 4.06) or cardioversion (aHR 0.87, 95% CI 0.25 to 3.00) and amiodarone in hospital mortality (see Appendix 6, Table 33).

Haemodynamic changes associated with atrial fibrillation onset

Multilevel linear modelling revealed that NOAF was associated with a significant heart rate increase of 22 b.p.m. (p < 0.001) and 19 b.p.m. (p < 0.001) in the MIMIC-III and PICRAM databases, respectively. The average heart rate after AF onset was 122 b.p.m. and 127 b.p.m., respectively.

New-onset atrial fibrillation was associated with a significant reduction in systolic blood pressure in the MIMIC-III and the PICRAM databases of 7 mmHg and 4 mmHg, respectively (p < 0.001). This was despite significant increases in the doses of vasoactive medication after NOAF onset in those receiving vasoactive medications prior to NOAF onset [vasoactive-inotropic score increase of 2.5 (p < 0.001) and 1.8 (p = 0.001), respectively]. New hypotension (systolic blood pressure of < 90 mmHg or mean blood pressure of < 65 mmHg) occurred after NOAF in 28% and 21% of patients with a systolic blood pressure of ≥ 90 mmHg or a mean blood pressure of ≥ 65 mmHg prior to AF onset, respectively. There was a non-significant increase in the proportion of patients receiving vasoactive medications after NOAF onset in the MIMIC-III database (17.6% to 20.2%; p = 0.29). This proportion was unchanged after NOAF onset in the PICRAM database (29%).

The change in heart rate, blood pressure and vasoactive medication use over time, before and after AF onset, is displayed in Figure 10.

FIGURE 10 Haemodynamic changes associated with AF onset. (a) Heart rate, MIMIC-III database; (b) heart rate PICRAM database; (c) systolic blood pressure, MIMIC-III database; (d) systolic blood pressure, PICRAM database; (e) vasoactive-inotropic score, MIMIC-III database; (f) vasoactive-inotropic score, PICRAM database; (g) proportion of patients on vasoactive medications, MIMIC-III database; (h) proportion of patients on vasoactive medications, PICRAM database. Vasoactive-inotropic score shown for those patients receiving vasoactive medications prior to AF onset. VIS = dopamine dose (µg/kg/minute) + dobutamine dose (µg/kg/minute) + 100 × adrenaline dose (µg/kg/minute) + 10 × milrinone dose (µg/kg/minute) + 10,000 ± vasopressin dose (units/kg/minute) + 100 × noradrenaline dose (µg/kg/minute). Reproduced with permission from Bedford et al.81 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. The figure includes minor additions and formatting changes to the original figure. (Figure not reproduced here; see the original article.)

Summary

This study demonstrated that NOAF during an ICU stay is common and is associated with significant increases in heart rate, reductions in blood pressure and increases in vasoactive medication requirements. NOAF was associated with an increased rate of hospital mortality despite adjusting for variables in a validated mortality prediction model.

Beta-blocker therapy was not associated with any difference in achieving rate or rhythm control, or with any difference in hospital survival when compared with amiodarone. The hospital mortality benefit with beta-blocker therapy that was identified in the unadjusted analysis of the MIMIC-III database was no longer apparent after adjustment. This suggests that differences in survival were only because of differences in patient characteristics between treatment groups.

Digoxin therapy was associated with a lower rate of achieving rate control than amiodarone. Calcium channel blocker therapy was associated with a lower rate of achieving rhythm control than amiodarone.