Towards the Prediction of Atrial Fibrillation Based on Interpretable ECG Features

Alexander Hammer, Hagen Malberg, Martin Schmidt
TU Dresden


Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmic disease in our society. Since early detection of AF episodes allows timely interventions, the prediction of AF episodes during sinus rhythm based on electrocardiograms (ECGs) is of great importance. Contrary to the often superior deep learning approaches, approaches based on handcrafted features provide physiologically interpretable information to clinicians. Therefore, we investigated the value of a variety of ECG features during sinus rhythm for predicting AF episodes.

We used the MIT-BIH AF database containing 25 long-term ECG recordings (10 hours each) with AF episodes. Sinus rhythm episodes of 60 minutes before AF have been split into segments of 5 minutes with 50% overlap. In total, 644 segments were used to calculate 212 features, including signal quality indices, heart rate (variability) features, and beat characteristics as well as their beat-to-beat changes. To investigate the importance of each feature for AF prediction, we performed logistic regression analyses between the segments of AF onset and temporally prior segments.

Logistic regression analyses showed the greatest number of significant effects for segments that ended 5 minutes before AF onset. Next to PQ interval slope (p < 0.01), median RR time (p < 0.01), and PQ interval correlation (p < 0.05) are indicated as most relevant. Additionally, 21 significant effects (p < 0.05) were found for signal quality indices and P/QRS/T wave features. Other segments before the AF episode showed significant effects but less frequently.

Our results confirm expected changes in the considered ECG features (e.g., P wave features) before AF episodes, indicating impaired atrial excitation. This demonstrates the potential of feature-based approaches using machine learning methods for the interpretable prediction of atrial fibrillation from ECG. For this purpose, however, extensive databases should be included to verify the value of the respective ECG features for AF prediction.