Unraveling the Electro-Pathophysiology of Atrial Fibrillation

Abstract

Introduction Atrial fibrillation (AF) is the most common age-related, progressive tachyarrhythmia in the USA and in European countries. AF is associated with an increased risk of stroke, heart failure, impaired cognitive function, and increased mortality. An obstacle for optimal diagnosis and treatment is the relatively unknown (electro-)pathophysiology of AF. In combination with intra-operative cardiac mapping, accurate analysis of the AF burden using post-operative continuous rhythm registrations might provide great insight into the underlying mechanisms of AF development. However, manual analysis of these continuous rhythm registrations is both time-consuming and subject to interpretation. Therefore, the aim of this study is to develop an automated AF detection algorithm for use in the research setting. Methods Using 6,400 manually annotated 30-seconds electrograms (ECGs) derived from the post-operative continuous rhythm registrations in the Erasmus Medical Center (Rotterdam), and 192 annotated records from standard MIT-BIH ECG databases, a classifier was developed with three output classes: AF, No AF, and Unusable (due to noise/artefacts). QRS-complexes were detected using a method based on the Pan-Tompkins algorithm. Subsequently, P- and T-waves were detected and features were extracted, which can be grouped into eight groups: RR-interval characteristics, peak-interval characteristics, amplitude characteristics, P-wave characteristics, T-wave characteristics, QRS-morphology characteristics, autocorrelation characteristics, and noise. Multiple classifiers were trained using a training set containing 4,800 post-operative ECGs and a hidden test set containing the remaining 1,600 post-operative ECGs. The optimal classifier in terms of accuracy was further optimized. Results Optimal classification was achieved using boosted decision trees. For the hidden test set, this resulted in an accuracy of 96.44% (95% CI: 95.41% - 97.24%) for detection of AF with a false negative rate of 2.8% (95% CI: 1.5% - 4.9%) and a false positive rate of 3.8% (95% CI: 2.9% - 5.1%). Of all 74 misclassifications, 36 (49%) were made in the group with irregular rhythms without AF. Classification was mainly based on the RR-interval characteristics. Conclusion An automated AF classifier based on post-operative continuous rhythm registrations for use in the research setting was proposed. Careful use of the classifier in combination with manual validation of detected AF segments makes the classifier suitable for supervised research purposes.