Application of quantitative methods to selection of rhythm management strategies in patients with newly diagnosed atrial fibrillation: a retrospective observational study (Preprint)

BACKGROUND Decisions regarding initial rhythm management strategy for patients with newly diagnosed atrial fibrillation (AF) are challenging to individualize. Prior work by our team has demonstrated feasibility of using machine learning algorithms at the level of diagnostic codes to predict rhythm management strategies; however, this work did not examine prediction models at the individual level. OBJECTIVE We aimed to examine decisions about rhythm management for AF and associated outcomes at the individual patient level, through mapping clinical trajectories within the electronic health record. We examined use of quantitative methods to predict the use of rate- or rhythm-control therapy in patients with a new diagnosis of AF. METHODS We performed a targeted chart review of patients with a new diagnosis of atrial fibrillation in the University of Colorado healthcare system from 2011 to 2020. We used clinical and demographic information to develop and test machine learning algorithms for predicting initial rhythm management strategy, and compared models based on accuracy as well as interpretability. Time-to-event and regression analyses were performed to predict the likelihood of change in rhythm management and the risk of subsequent hospitalization or death. RESULTS Of 419 patients with an EHR-diagnosis of new AF, we confirmed the diagnosis in 289 patients, 194 of whom we classified as paroxysmal and 95 persistent. For all patients with AF, rhythm-control therapy resulted in more total management changes. For paroxysmal AF, rhythm-control therapy resulted in fewer hospitalizations (incidence rate ratio 0.32, 95% CI 0.19-0.34) and no difference in mortality compared to rate-control therapy. For persistent AF, rhythm control resulted in more hospitalizations (incidence rate ratio 2.44, 95% CI 1.41-4.22) with no difference in mortality. Machine learning models could predict the initial strategy with limited accuracy, which was higher in patients with persistent than paroxysmal AF. CONCLUSIONS Quantitative decision models for rhythm management of AF can be developed through trajectory mapping using chart review. Discussion about patient-centered outcomes should be considered when selecting initial therapy for atrial fibrillation. CLINICALTRIAL n/a