Abstract 4142494: Prediction Of Atrial Fibrillation From Beat-To-Beat Correlations During Normal Rhythm

Background: Atrial fibrillation (AF) is the most common type of cardiac arrhythmia, affecting over 59 million individuals worldwide. Even wearable heart rate monitors can detect ongoing AF episodes, but none of the devices are able to assess the susceptibility to impending episodes. The prediction of AF episodes during normal rhythm would significantly improve the early treatment of AF. Research Hypothesis: Heart diseases are known to affect the short- and long-range correlations in beat-to-beat intervals, i.e., RR interval (RRI) time series. The hypothesis is that persistent alterations remain in the RRI correlations even outside of AF episodes. The study examines whether the RRI correlations can indicate the susceptibility to AF during normal rhythm. Methods: We analyze the correlations in RRIs with detrended fluctuation analysis (DFA). We enhance the analysis with its dynamical (DDFA) variant, which assesses time-dependent correlations over several ranges of scale. We attempt to discriminate the AF population with varying ratios of AF episodes (from <20% to >99%) from healthy controls by their RRI correlations with XGBoost classifier. Data and results: The data consist of long ECG recordings of 274 healthy controls and 84 patients with AF from Physionet and THEW datasets. The data is split into subclasses according to the relative time of the recording spent in AF: (a) >99% (N=34), (b) 20-99% (N=24), (c) <20% (N=26). The classification according to DDFA calculations leads to the following results in each subclass for sensitivity and specificity, respectively: (a) 94% and 97%; (b) 88% and 94%; (c) 73% and 96%. Figure 1 illustrates the differences in DDFA results between each of the studied subclasses and healthy controls. There is significant qualitative similarity between the subclasses (a-c) compared to the healthy controls in (d), especially at high heart rates. We confirmed that a small ratio of AF episodes in the data, such as in subclass (c), does not impair the DDFA results. Conclusions: The RRI correlations in the AF population show marked deviation from healthy controls even outside of AF episodes. A significant portion of the AF cases could be isolated outside the episodes with excellent specificity, paving the way for practical early screening without overflowing the healthcare system with false positives.