Abstract
Atrial fibrillation (AF) represents a condition of irregularities of heartbeats. Timely diagnosis and treatment of AF are crucial to avoid serious consequences such as stroke, heart failure and death. AF is detected from patient’s rhythm data as a binary longitudinal sequence. In this article, we introduce an approach for analyzing the binary longitudinal sequence of AF. Our approach is based on gradient boosting, a machine learning approach. We use multivariate tree as a base learner to model complex relationships between multiple covariates and response. In order to model covariate-time interactions, we use B-spline. Application of our approach to a randomized trial data provides the importance of surgical ablation in the treatment of AF among patients with persistent or long-standing persistent AF. Comparison of prediction performance of our approach with other methods using randomized trial data and simulated data shows that our approach has a better prediction performance. Variable selection using variable importance has identified duration of AF as an important covariate that has strong interaction with post-surgery time. This helps to identify patients’ variability and groups of patients who derived the most benefits from the treatment. Our method can be implemented using the R package boostmtree.