Prediction of atrial fibrillation recurrence before catheter ablation using P-wave ECG features: simplicity of random forest versus CNN in small datasets

Abstract

Atrial fibrillation (AF) is an arrhythmia affecting 3% of the general population. Transcatheter ablation (CA) is the most effective treatment option for patients with AF; however, recurrence rates remain high. Early classification of patients based on pre-ablation data could help personalize decisions regarding CA by identifying candidates with a higher likelihood of clinical success. We examined 123 patients for whom we collected 6 minute, 3-lead orthogonal ECG recordings in sinus rhythm prior to the procedure, alongside clinical data. We benchmarked a state-of-the-art convolutional neural network (CNN) classifier widely used in AF prediction and found that it performed poorly, with 62% accuracy on our dataset. To improve performance, we conducted beat-to-beat P-wave analysis and formed a feature vector combining continuous wavelet transform features, 3-dimensional spatiotemporal variables, and time-domain parameters. Through recursive feature elimination, a Random Forest (RF) classifier achieved an accuracy of 76% using flat 10-fold cross-validation (CV) and maintained a generalization accuracy of 68% under leak-free 10-fold nested CV. The RF model outperformed the CNN on unseen data, offering a more effective and explainable framework. Extracting clinically interpretable features may pave the way for decision-support tools for physicians. Our study shows that when the task is challenging, such as distinguishing AF-free patients from those who relapse, and the dataset is small, simpler and computationally lighter approaches may be preferable.