A machine learning-based risk prediction model for atrial fibrillation in critically ill patients.

Background: Atrial fibrillation (AF) in critically ill patients increases morbidity, hospital stays, and costs. Existing prediction tools are limited in intensive care unit (ICU) settings. Objective: This study developed a machine learning-based model to enable early AF risk identification and prevention. Methods: In this retrospective cohort study, adult patients admitted to the ICU were identified from the MIMIC-IV (Medical Information Mart for Intensive Care-IV) database, including 47 clinical and laboratory variables. The primary outcome was AF within the first 48 hours of admission. Multiple machine learning models were trained to predict AF, with the top-performing model undergoing hyperparameter tuning. A compact model was developed using 15 variables and 2 novel features-one identifying patients 70 years of age or older with sepsis and another representing a composite score of pre-existing cardiac risk factors. Model performance was evaluated using accuracy, area under the receiver-operating characteristic curve (AUROC), and predictive values. SHAP (Shapley Additive exPlanations) analysis interpreted individual feature contributions to the model's predictions. Results: The cohort comprised 46,266 ICU patients, with 4.6% developing AF within 48 hours. The CatBoost classifier model achieved an AUROC of 0.850 on the test set, while the compact model with new features yielded an AUROC of 0.820. SHAP analysis highlighted total serum magnesium, age, and the newly created features as key predictors of AF development. Conclusion: This study demonstrates the potential of machine learning models in predicting AF development in ICU patients. The compact model, with a satisfactory AUROC, can be a valuable tool for identifying high-risk patients and facilitating timely interventions.