Development and external validation of a clinical prediction model for new-onset atrial fibrillation in intensive care: a multicentre, retrospective cohort study

Background: New-onset atrial fibrillation, a condition associated with adverse outcomes in the short and long term, is common in patients admitted to intensive care units (ICUs). Identifying patients at high risk could inform trials of preventive interventions and help to target such interventions. We aimed to develop and externally validate a prediction model for new-onset atrial fibrillation in patients admitted to ICUs.

Methods: We conducted a multicentre, retrospective cohort study in three ICUs across the UK and four ICUs across the USA. Patients aged 16 years and older admitted to an ICU for more than 3 h without a history or presentation of clinically significant arrhythmia were eligible for inclusion. We analysed clinical variables to investigate the associations between predetermined candidate variables and risk of new-onset atrial fibrillation and to develop a model to estimate these risks. We developed the METRIC-AF model, a machine learning model incorporating dynamic variables. Model performance was assessed through internal–external cross-validation during model development and externally validated by use of multicentre data from ICUs across the UK. We then developed a simple graphical prediction tool using three important predictors.

Findings: Among 39 084 eligible patients admitted to an ICU between 2008 and 2019, 2797 (7·2%) developed new-onset atrial fibrillation during the first 7 days of their ICU stay. We identified multiple non-linear associations between candidate variables and risk of new-onset atrial fibrillation, including hypomagnesaemia at serum concentrations below 0·70 mmol/L. The final METRIC-AF model contained ten routinely collected clinical variables. Compared with a published logistic regression model, the METRIC-AF model showed superior calibration, net benefit across clinically relevant risk thresholds, and discriminative performance (C statistic 0·812 [95% CI 0·805–0·822] vs 0·786 [0·778–0·801]; p=0·0003). The simple graphical tool performed well in attributing the risk of new-onset atrial fibrillation in the external validation dataset (C statistic 0·727 [95% CI 0·716–0·739]).

Interpretation: The METRIC-AF model and its companion graphical tool could support the identification of patients at increased risk of developing new-onset atrial fibrillation during ICU admission, informing targeted prophylactic strategies and trial enrichment by use of routinely available clinical data. An online app also developed as part of the study allows for the exploration of prediction generation among individuals and external validation in prospective studies.

Funding: National Institute for Health and Care Research (NIHR) and NIHR Oxford Biomedical Research Centre.