Predicting atrial fibrillation from intensive care unit numeric data

Atrial fibrillation is a common occurrence in intensive care units (ICUs) and is associated with a significant increase in patient mortality and morbidity, healthcare costs, and length of hospital stay. This burden can be significantly reduced through clinical tools to identify patients at increased risk of developing atrial fibrillation during ICU admission and to match these patients to appropriate prophylaxis (e.g., amiodarone). Unfortunately, despite its prevalence, predicting atrial fibrillation remains a challenge. In this paper, we address the goal of developing an accurate approach to stratify patients for atrial fibrillation using information available in numerics data (e.g., vital signs, arterial blood pressures) commonly collected during ICU admission. We explore the use of a support vector machine (SVM) classifier optimized for multivariate non-linear performance using an area under the receiver operating characteristic curve (AUROC) loss function with summary features derived from ICU numerics collected during the first 8 hours of admission. When evaluated on a cohort of 1,531 ICU patients, this approach achieved an AUROC of 0.73 and a sensitivity of 71% in identifying patients who experienced atrial fibrillation during admission using data only from the start of ICU hospitalization.