The Taub Faculty of Computer Science Events and Talks

Atrial fibrillation (AF) is the most prevalent form of heart arrhythmia and is associated with a fivefold increase in stroke incidence. In the context of AF detection, some patients experience sporadic AF events. This makes the Holter electrocardiogram (ECG) examination, which captures longer-term heart activity, essential to capture these irregular events. Automatic detection of AF in Holter recordings has the potential to reduce clinician workload. On the ECG, AF is characterized by an irregular rhythm and by the presence of fibrillatory waves (f-wave). While the standard Holter examination typically includes three leads, single-lead ECGs have become increasingly more common thanks to the development of patches and smartwatches for remote health monitoring and screening. This research makes two scientific contributions towards creating AI driven systems to support the detection and analysis of AF in single lead ECG. First, we propose a new method for ranking f-wave extraction methods. Second, we develop a robust, i.e., highly performing and generalizable model, for AF events detection and benchmark this new model to state-of-the-art.

The new method for ranking f-wave extraction algorithms is based on the hypothesis that better-performing AF classification using features computed from the extracted f-waves implies better-performing extraction. Three independent Holter datasets and four f-wave extraction algorithms were used for this experiment. The results showed that the PCA-based f-wave extraction approach was superior on all datasets and for all leads. A significant advantage of our evaluation method lies in its ability to leverage real datasets without the need for ground truth f-waves.

The elaboration of robust algorithms for AF events detection from single lead ECG present several challenges. These include distribution shifts across lead, hardware, ethnicity and the inherent presence of noise within the ECG signal. A new deep learning model, called RawECGNet, was developed for the task of AF detection. In order to enhance the generalization capabilities of the model, it was trained on single lead ECG input from different lead position. In addition, the domain shift uncertainty (DSU) layer was included to introduce a degree of uncertainty into the encoding process. RawECGNet demonstrates exceptional generalization both in the source domain and two target domains and outperformed a state-of-the-art deep learning model taking as input the beat-to-beat interval time series. RawECGNet harnesses the complete potential of ECG waveform morphology for enhanced diagnostic accuracy.