Prediction of Incident Atrial Fibrillation and Association With Outcomes Using Routine Electronic Health Records in a Western Pacific Population

Background and Aim

Atrial fibrillation (AF) affects over 37 million people internationally and confers increased risk of cardiovascular conditions. Prediction algorithms have attempted to predict incident AF, but other cardio–renal diseases could also provide targets for earlier intervention.

Method

We derived a random forest classified for incident AF within 5 years (Future Innovations in Novel Detection of Atrial Fibrillation [FIND-AF] Taiwan) using routinely collected data from the National Taiwan University Hospital database. We compared this to congestive heart failure, hypertension, age >75 years (two points), diabetes mellitus, stroke/transient ischaemic attack/thromboembolism (two points), vascular disease, age 65–74 years, sex category (CHA2DS2-VASc) and coronary artery disease/chronic obstructive pulmonary disease (one point each), hypertension, elderly (age ≥75 years, two points), systolic heart failure, thyroid disease (hyperthyroidism) (C2HEST). Youden’s Index was calculated to determine optimal threshold for higher versus lower predicted AF risk. We calculated cumulative incident curves and hazard ratios for incident AF, heart failure hospitalisation, or development of moderate-to-severe renal impairment, stroke or transient ischaemic attack and cardiovascular and all-cause mortality.

Results

Overall, 103,321 patients were included, with an average age of 64.6 years and 52.8% women. Overall, 4.4% had incident AF over the 5-year follow-up period. FIND-AF Taiwan had better discrimination (area under received operating characteristic 0.792, 95% confidence interval [CI] 0.777–0.807) than CHA2DS2-VASc (0.737; 0.721–0.754) and C2HEST (0.750; 0.733–0.766). After adjustment, individuals at higher risk were at increased hazard for heart failure hospitalisation (hazard ratio 15.32; 95% CI 9.19–25.54), transient ischaemic attack or ischaemic stroke (28.4; 21.01–38.47), progression to moderate or severe chronic kidney disease (1.18; 1.09–1.27), cardiovascular mortality (1.32; 0.88–1.98), and all-cause mortality (1.23; 1.11–1.36).

Conclusions

We derived a supervised machine learning algorithm for AF in a Western Pacific population and demonstrated that higher risk was associated with hospitalisation for other cardio–renal diseases and death. This tool could be used to target interventions to reduce hospitalisation.