Optimizing atrial fibrillation detection through ECG feature selection using Extra-Trees and statistical association...

Area of Science:

• Cardiology and Medical Informatics
• Signal Processing and Machine Learning

Background:

• Atrial fibrillation (AFib) is a common arrhythmia increasing stroke and heart failure risk.
• Accurate AFib detection from 12-lead ECGs is challenging due to complex interpretation.
• Machine learning (ML) and deep learning (DL) show promise but require optimal feature selection.

Purpose of the Study:

• To develop a hybrid feature selection methodology for identifying discriminative ECG features.
• To objectively distinguish AFib from normal sinus rhythm (NSR) using ECG data.
• To enhance the interpretability and efficiency of ML/DL models for AFib detection.

Main Methods:

• A hybrid framework combining Extremely Randomized Trees (Extra-Trees) with statistical association measures.
• Evaluation of morphological, entropy-based, and spectral hand-crafted features from 12-lead ECGs.
• Introduction of novel metrics: feature importance score (FIS) and overall feature importance score (OFIS).

Main Results:

• The method ranked 97 features, identifying the top 10 per lead and 20 overall with high consistency.
• The interquartile range of RR-intervals showed the highest normalized OFIS, indicating strong discriminative power.
• Feature space dimensionality was reduced by nearly 80%, preserving interpretability and physiological meaning.

Conclusions:

• The proposed methodology offers a reproducible, interpretable, and statistically grounded framework for ECG feature discovery.
• This approach serves as a valuable preprocessing step for ML/DL models in AFib detection.
• The findings aid clinicians in achieving more accurate real-time AFib detection.