236
Automatic classification of 20 different types of atrial tachycardia using 12-lead ECG signals

Supported by the European Union"s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No.766082 (MY-ATRIA)

 Atrial Flutter (AFl) as a common reentrant atrial tachycardia is driven by self-sustainable mechanisms that cause excitation to propagate along pathways different from sinus rhythm. Intracardiac electrophysiological mapping and catheter ablation is often performed without prior knowledge of the mechanism perpetuating AFl in a given patient, likely prolonging the procedure time of these invasive interventions. We investigated the feasibility of automatically identifying 20 different AFl types based on the non-invasive 12-lead electrocardiogram (ECG) using machine learning. 

 Electrophysiological fast marching computer simulations of 20 different atrial tachycardia scenarios (micro-/macro-reentry, scar-related/anatomical/functional, figure-of-eight, focal, different locations) were performed and propagated to the standard 12-lead ECG based on the Courtemanche atrial action potential model. The virtual study population comprised combinations of 8 different anatomical bi-atrial models with 2 orientational variants each and 8 different torso models yielding a total of 2512 ECGs. From each ECG, we extracted 114 features from different domains (e.g., time, frequency, entropy, wavelet, non-linear recurrence analysis). The dataset was randomly split into 1256 training samples, 628 validation samples and 628 test samples while maintaining a balanced AFl type distribution. A radial basis neural network (RBNN) was trained as a classifier after selection of the most informative features. 

The RBNN yielded a test set accuracy of 90% regarding the identification of the AFl mechanism using 10 features (from different domains). The most discriminative single feature was the cycle length that alone led to a test set accuracy of 74%, while the remaining feature set without cycle length (9 features) reduced the test set accuracy to 33%. The machine learning approach generalized well regarding unseen torso geometries (90% accuracy if training was performed on only 7 torso models) but rather poor regarding atrial anatomies (23% if the atrial anatomical model was not seen during training) indicating that more than the currently used 8 atrial models should be included during training to cover the relevant anatomical variability. 

Our results show that a machine learning classifier can potentially identify a high number of different AFl types using the 12-lead ECG. This non-invasive method can aid in planning and tailoring AFl treatment for patients. Application to clinical data is necessary as a next step to pave the way for clinical translation.

Abstract Figure.