AFICILL: a single-cohort, retrospective study on Atrial Fibrillation In Critically ILL patients admitted to a medical sub-intensive care unit: implications for clinical management, outcomes and elaboration of new data-driven models

Introduction: atrial fibrillation (AF) is common among critically-ill patients, who are considered at increased cardioembolic and haemorragic risk. Consequently, anticoagulant therapy might be ineffective or harmful for an excess of haemorragic events which could not be counterbalanced by an adequate reduction of cardioembolic occurrences. Aims: main outcome (MO) was the composite of death or intensive care unit (ICU) transfer in a population of critically-ill subjects admitted to a medical subintensive care unit (sICU); we assessed (i) thromboembolic events (TEE) and major haemorrhages (MH); (ii) current guidelines (GL) adherence and related outcomes; (iii) performance of validated risk scores for TEE and MH; we engineered (iv) new scores adopting machine learning (ML) predicting MO, TEE, MH. Patients and Methods: single-center, retrospective study enrolling all the consecutive AF-affected patients admitted to a sICU for critical illness. Demographic, clinical, therapeutic and laboratoristic data were collected. Performance of CHA2DS2-VASc and HAS-BLED scores was evaluated. GL-adherence and its relationship with outcomes was studied. ML was used to engineer new predictive models. Results: we enrolled 1430 subjects; CHA2DS2-VASc (AUC:0.516;95%CI:0.472-0.560) and HAS-BLED (AUC:0.493;95%CI:0.443-0.543) did not predict TEE or MH; in-hospital warfarin use was associated to increased MO risk (OR:1.73;95%CI:1.06-2.83; p<0.05); low-molecular-weight-heparin use was not associated to an increased MO risk; antiplatelet drugs use was associated to MO risk reduction (OR:0.51;95%CI:0.34-0.78;p<0.002). GL-adherent treatment was associated to TEE risk reduction and MH and MO risk increase; ML identified specific features for MO, TEE, MH: ML-based classifiers outperformed CHA2DS2-VASc (AUC: from 0.516 to 0.90, p<0.0001) and HAS-BLED (AUC: from 0.493 to 0.82, p<0.0001). Discussion: AF-related outcomes cannot be predicted in critically-ill patients with currently validated methods. GL-adherence is associated to a significant TEE reduction, but also to MH and MO increase. ML algorithms can identify the most important features and shape specific scores able to outperform the classical models.