Non-valvular atrial fibrillation (NVAF) is the most common sustained arrhythmia observed in critically ill patients, linked to a higher risk of embolic and haemorrhagic events. Conventional tools, such as CHADS2, CHA2DS2-VASc, and HAS-BLED scores, are ineffective for risk stratification and do not offer guidance for anticoagulation strategies in this population. Recently, we engineered new machine-learning (ML) models retrospective cohorts, with promising results; in this work, we aim to validate our ML models in a larger cohort. We performed a retrospective analysis of all consecutive critically ill patients admitted to our step-down unit over a 10-year period who had a history of NVAF. We calculated classical risk scores and trained our ML models on pre-specified outcomes: the main outcome (MO) which was a composite of in-hospital death or intensive care unit (ICU) transfer, stroke/TIA, and major bleeding (MB) during the admission. After eliminating trauma and non-critical patients, we obtained 2105 subjects, with 314 MO, 134 cardioembolic stroke/TIA and 227 MB. Classical risk scores (APACHE-II for MO, CHADS2 and CHA2DS2-VASc for stroke/TIA, HAS-BLED for MB) performed poorly, while ML confirmed its accuracy in predicting outcomes also in this extended cohort (AUC APACHE-II:0.6397; 95%CI:0.6064-0.6729; AUC MO-ML:0.96; 95%CI:94.6-97.2; p<0.0001; AUC CHADS2:0.5775; 95%CI:0.5332-0.6218; p<0.0001; AUC CHA2DS2-VASc:0.5793; 95%CI:0.5357-0.6228; AUC stroke/TIA-ML:0.95; 95%CI:94.3- 96.6; p<0.0001; AUC HAS-BLED:0.5089 95%CI:0.4786-0.5392; AUC MB-ML:0.973 95%CI 95.5–98.1; p<0.0001). ML models can be considered as potential candidates in this setting to guide anticoagulant therapy. Multicenter, prospective cohorts will be necessary to establish their applicability in clinical practice.
The Atrial Fibrillation In Critically Ill patients (AFICILL) studies: validation and implemetation of topological data analysis and machine learning techniques in the prediction of atrial-fibrillation related outcomes in patients admitted to medical sub-intensive care units / Guerrieri, Emanuele. - (2026 Mar 24).
The Atrial Fibrillation In Critically Ill patients (AFICILL) studies: validation and implemetation of topological data analysis and machine learning techniques in the prediction of atrial-fibrillation related outcomes in patients admitted to medical sub-intensive care units
GUERRIERI, EMANUELE
2026-03-24
Abstract
Non-valvular atrial fibrillation (NVAF) is the most common sustained arrhythmia observed in critically ill patients, linked to a higher risk of embolic and haemorrhagic events. Conventional tools, such as CHADS2, CHA2DS2-VASc, and HAS-BLED scores, are ineffective for risk stratification and do not offer guidance for anticoagulation strategies in this population. Recently, we engineered new machine-learning (ML) models retrospective cohorts, with promising results; in this work, we aim to validate our ML models in a larger cohort. We performed a retrospective analysis of all consecutive critically ill patients admitted to our step-down unit over a 10-year period who had a history of NVAF. We calculated classical risk scores and trained our ML models on pre-specified outcomes: the main outcome (MO) which was a composite of in-hospital death or intensive care unit (ICU) transfer, stroke/TIA, and major bleeding (MB) during the admission. After eliminating trauma and non-critical patients, we obtained 2105 subjects, with 314 MO, 134 cardioembolic stroke/TIA and 227 MB. Classical risk scores (APACHE-II for MO, CHADS2 and CHA2DS2-VASc for stroke/TIA, HAS-BLED for MB) performed poorly, while ML confirmed its accuracy in predicting outcomes also in this extended cohort (AUC APACHE-II:0.6397; 95%CI:0.6064-0.6729; AUC MO-ML:0.96; 95%CI:94.6-97.2; p<0.0001; AUC CHADS2:0.5775; 95%CI:0.5332-0.6218; p<0.0001; AUC CHA2DS2-VASc:0.5793; 95%CI:0.5357-0.6228; AUC stroke/TIA-ML:0.95; 95%CI:94.3- 96.6; p<0.0001; AUC HAS-BLED:0.5089 95%CI:0.4786-0.5392; AUC MB-ML:0.973 95%CI 95.5–98.1; p<0.0001). ML models can be considered as potential candidates in this setting to guide anticoagulant therapy. Multicenter, prospective cohorts will be necessary to establish their applicability in clinical practice.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
