Prediction of early death after atrial fibrillation diagnosis: a french nationwide study

Abstract Funding Acknowledgements Type of funding sources: None. Background Atrial fibrillation (AF) is associated with important mortality. Dedicated clinical scores to predict mortality have been developed but perform modestly and are not specific for this population. Machine learning models developing in the field AF may be able outperform existing tools prediction Purpose This study aimed train evaluate machine death occurrence within critical period year following atrial diagnosis compare predictive ability usual scores. Methods We used purpose a nationwide cohort 2,435,541 newly diagnosed patients seen all French hospitals from 2011 2019. Three ML (logistic regression, random forests, deep neural networks) were trained first on set (70% cohort). The best model was selected test (30% Discrimination evaluated using C index. finally compared our previously published Results Within diagnosis, 342,005 (14.4%) died after mean time 83 (SD 98) days whom 107,715 cardiovascular deaths (31.5%). Among 110 variables, 18 most variables identified Random Forest algorithm: age, metastasis, resuscitated cardiac arrest, cancer, congestive heart failure, decubitus ulcers, renal pneumonia, lung disease, difficulty walking, malnutrition, anemia, impaired mobility, liver blood transfusion urinary tract infection. After training 3 algorithms, network index 0.785 (95% CI, 0.781-0.789) set. incidence cause at one rises stepwise fashion 13.8 per 1000 quintile 352.4 fifth quintile. Compared traditional risk scores, significantly superior CHA2DS2-VASc HAS-BLED dedicated such as Charlson Comorbidity Index Hospital Frailty Risk Score (C indexes: 0.597; 0.562; 0.643; 0.626 respectively. P<0.0001) (Figure 1). improved shown by net reclassification integrated discriminatory improvement increase (P<0.0001, respectively) decision curve analysis 2). Conclusion algorithms early better than A approach help clinicians stratify high