Leading organ-failure phenotype and subsequent new-onset atrial fibrillation in prolonged ICU stays: derivation in MIMIC-IV and external validation in eICU-CRD

Abstract Background New-onset atrial fibrillation (NOAF) is common during critical illness, but early risk stratification remains imprecise. Organ-dysfunction trajectories have been studied in critical care, yet the association between the first dominant organ-failure phenotype and subsequent NOAF has not been defined. Methods We performed a retrospective derivation-validation cohort study of adult ICU stays lasting at least 72 h. The derivation cohort used MIMIC-IV and the validation cohort used eICU-CRD. A five-system leading-organ phenotype was defined by the first organ-specific SOFA subscore to reach ≥ 3 within the first 72 h (cardiovascular, respiratory, renal, hepatic, coagulation, or no failure). Neurological SOFA was excluded from the primary phenotype because of limited cross-database transportability. The primary outcome was NOAF identified after 72 h of ICU stay. Multivariable logistic regression was used in the derivation cohort, and a reduced transportable model was externally validated in eICU. Results The derivation cohort included 11,735 ICU stays and 367 NOAF events after 72 h (3.1%). The validation cohort included 42,942 stays and 2,216 pragmatic AF events after 72 h (5.2%). In MIMIC-IV, crude event rates were highest in the cardiovascular-leading phenotype (5.5%) and respiratory-leading phenotype (3.3%), compared with no failure (1.4%). In the primary adjusted derivation model, cardiovascular-leading phenotype was associated with higher odds of NOAF versus no failure (OR 1.81, 95% CI 1.01–3.23; p = 0.047), whereas the respiratory-leading phenotype was not independently significant (OR 1.32, 95% CI 0.86–2.03; p = 0.206). In eICU, crude event rates were again highest in cardiovascular-leading (7.5%) and respiratory-leading phenotypes (5.7%), compared with no failure (4.2%). When the reduced transportable derivation model was applied to eICU, discrimination was modest (AUROC 0.68), calibration slope was 0.64, and calibration intercept was 1.75. In the eICU refit model, cardiovascular-leading phenotype remained associated with higher odds of pragmatic AF after 72 h (OR 1.20, 95% CI 1.04–1.40; p = 0.015). Conclusions A clinically transparent five-system leading-organ phenotype identified differential NOAF risk across prolonged ICU stays. Cardiovascular-leading phenotype showed the most consistent signal across derivation and validation cohorts. This approach may complement existing critical-care NOAF risk models by adding a temporally ordered organ-dysfunction framework.