Predicting Catheter Ablation Outcomes with Pre-ablation Heart Rhythm Data: Less Is More

Abstract

While numerous studies have shown that catheter ablation is superior to anti-arrhythmic drug (AAD) in treating atrial fibrillation (AF), its long-term outcomes have been limited by arrhythmia recurrence, which is considered a negative outcome per current clinical standard. This gives rise to difficulty in choosing between AAD and catheter ablation, which pose risks of complications but may achieve higher efficacy when compared to the former. As an effort to overcome this dilemma, we evaluate in this work the joint utility of machine learning methods and cardiac data measured prior to ablation for outcome prediction. We advanced research along two fronts. On the clinical front, we evaluated the plausibility of developing models that take as input pre-ablation heart rhythm time-series data to predict future outcome of ablation. On the technical front, we conducted extensive experiments to address the following questions: 1) Could the use of recurrent neural networks achieve the best predictive performance for this application? 2) How would multi-layer perceptron (MLP) compare to recurrent networks? 3) How might the design of bottleneck in MLPs affect performance? 4) How would traditional classification algorithms compare to (deep) neural networks? As an initial attempt to answer these questions, we conducted over 100 sets of cross-validation experiments and found that the top-performing predictive model achieved 71.0 ± 2.1 in area under receiver operating characteristic curve (AUC), with sensitivity of 63.0 ± 4.3 and specificity of 64.2 ± 4.5, as evaluated on a cohort of 343 samples. We also found that all models evaluated in this work achieved greater predictive performance than two risk scores commonly cited in the clinical research literature.