Prediction of incident hospitalisation with atrial fibrillation and association with cardio-renal outcomes in taiwan

Background Prediction models have been developed and/or validated to identify individuals at risk of AF in the community,1 including through the use of machine learning.2 AF hospitalisation may be a more important outcome to predict, as it is associated with adverse events and health costs.3 4 It is unknown how prediction models developed and/or validated for incident AF prediction predict hospitalisation with AF, and how AF hospitalisation risk is associated with other cardio-renal hospitalisations and death. Purpose To test prediction models developed and/or validated for AF for the outcomes of AF hospitalisation in a Taiwanese population, develop a new model, and investigate association of higher predicted AF hospitalisation risk with cardio-renal hospitalisations and death. Methods The study population was derived from the National Taiwan University Hospital Integrated Medical Database. We included patients aged ≥50 years who were seen at NTUH between 01/01/2014-31/12/2019 without known AF or atrial flutter. Comorbidities were identified using ICD-10 codes. We divided the dataset 8:2 into training and testing cohorts We developed a new random forest model using the original FIND-AF variables (FIND-AF Taiwan) and evaluated performance in the testing dataset with internal bootstrap validation with 200 samples and conducted internal validation. We compared prediction performance compared with CHA2DS2-VASc, C2HEST, and FIND-AF. We then calculated the cumulative incidence rate for heart failure, stroke or transient ischaemic attack (TIA), chronic kidney disease (CKD), and both all-cause and cardiovascular mortality by FIND-AF Taiwan risk stratification. Results A final cohort of 103,321 patients (mean age 64.6 (10.6) years, 52.8% women) was included in the analysis, with 4544 new cases of AF at 5 years. Prediction performance for AF hospitalisation was only moderate for FIND-AF (AUC 0.687, 95% CI 0.669-0.706), CHA2DS2VASc (AUC 0. 679, 95% CI 0.661-0.697) and C2HEST (AUC 0. 679, 95% CI 0.662-0.697). The derived FIND-AF Taiwan model had superior prediction performance (AUC 0.792, 95% CI 0.777-0.807) (Table 1). When stratified by FIND-AF Taiwan, patients with higher-predicted AF risk had a higher hazard for AF hospitalisation (unadjusted HR 7.5, 95% 4.7-11.9) but also for heart failure hospitalisation (HR 49.0, 95% CI 31.1-77.8), TIA or ischaemic stroke (HR 17, 13.7-21.8), moderate to severe CKD (HR 4.0, 95% CI 3.8-4.2). Patients at higher predicted AF risk also had higher hazard for cardiovascular mortality (HR 1.32, 95% CI 0.88-1.98) and all-cause mortality (HR 1.23, 95% CI 1.11-1.36), in comparison with lower risk patients. (Figure 1) Conclusion A machine learning algorithm developed for the Taiwanese population has superior prediction performance for incident AF hospitalisation, and higher AF hospitalisation risk is associated with adverse cardio-renal outcomes.