Accurate detection of atrial fibrillation (AFib) from 12-lead electrocardiograms (ECGs) remains clinically important yet complex. This study, published in the Journal of Electrocardiology, developed a hybrid feature selection methodology combining Extremely Randomized Trees (Extra-Trees) with statistical association measures to identify physiologically meaningful ECG features for distinguishing AFib from normal sinus rhythm (NSR).

The analysis included 12-lead ECG recordings from 18 patients who underwent catheter ablation for AFib at a single center. Morphological, entropy-based, and spectral hand-crafted features were extracted. Two novel metrics—feature importance score (FIS) and overall feature importance score (OFIS)—were introduced to quantify feature relevance.

A total of 97 features were ranked. The framework identified the 10 most important features per ECG lead and 20 most relevant features overall, with high consistency across leads. The interquartile range of RR intervals achieved the highest normalized OFIS value (0.064), followed by rhythm-related and entropy-based measures. Feature dimensionality was reduced by nearly 80% while preserving interpretability and physiological meaning.

Limitations include small sample size (18 patients), single-center design, and absence of external validation.

This hybrid methodology provides a reproducible and interpretable framework for ECG feature selection in AFib detection. External validation in larger datasets is required before broader clinical application.