Machine learning models demonstrated moderate to excellent discrimination for predicting mortality in ICU patients with atrial fibrillation, achieving AUCs ranging from 0.768 to 0.978.
Systematic Review (n=40,900)
Do supervised machine learning models accurately predict mortality and length of stay in adult ICU patients with atrial fibrillation?
Machine learning models show strong predictive performance for mortality in ICU patients with atrial fibrillation, but significant gaps remain in calibration reporting, fairness assessment, and clinical implementation.
Effect estimate: AUC 0.768-0.978
Atrial fibrillation (AF) in intensive care unit (ICU) patients is associated with higher mortality, longer stays, and greater resource use than in patients without AF. Machine learning may improve risk stratification in this high-risk population, but existing models have not been systematically synthesized. This rapid review summarizes how machine learning methods have been used to predict outcomes in ICU patients with AF, with primary emphasis on mortality and current gaps in length of stay (LOS) modeling. Searches of PubMed, Embase, and Scopus (2015-2025) identified studies applying machine learning to intensive care outcomes in AF. Screening and data extraction were conducted in a web-based system using a single-reviewer approach with verification. Extracted items included study design, cohort characteristics, modeling approach, and performance metrics, and risk of bias and applicability were appraised using tools appropriate for prediction-modeling studies. Of 597 records screened, three studies met the inclusion criteria. All were US-based and used large electronic health record (EHR) datasets (sample sizes: 5,998-10,144). Algorithms evaluated included adaptive boosting, random forest, and stacking ensembles, with discrimination ranging from moderate to excellent for mortality prediction (area under the curve (AUC): 0.768-0.978). Frequently selected predictors included age, ICU severity indices (Acute Physiology Score III, Simplified Acute Physiology Score II, Sequential Organ Failure Assessment), vital signs, renal and metabolic laboratory values (e.g., blood urea nitrogen, estimated glomerular filtration rate, glucose), blood indices (such as white blood cell count and red cell distribution width), treatment indicators (mechanical ventilation, vasopressors, anticoagulation), and glycemic variability (GV). Steps toward clinical use were limited to prototype or web-based tool development, and routine deployment was not reported. Notably, none of the included studies developed or validated an LOS regression model. Overall, machine learning shows clear promise for mortality prediction in ICU patients with AF, but implementation remains limited, and key operational outcomes remain understudied. Priorities for future work include external validation across diverse settings, prospective evaluation of clinical impact, development of models for additional resource and utilization outcomes alongside mortality prediction, and assessment of fairness across patient groups to support safe, equitable, and scalable clinical use.
Nguyen et al. (Sat,) conducted a systematic review in Atrial fibrillation in intensive care unit patients (n=40,900). Machine learning prediction models vs. Traditional clinical risk scores (e.g., SOFA, SAPS II) was evaluated on Mortality prediction (AUC) (AUC 0.768-0.978). Machine learning models demonstrated moderate to excellent discrimination for predicting mortality in ICU patients with atrial fibrillation, achieving AUCs ranging from 0.768 to 0.978.