Inter-Patient Atrial Flutter Classification Using FFT-Based Features and a Low-Variance Stacking Classifier

Objective:
Atrial Flutter (AFL) is a supraventricular tachyarrhythmia typically arising from a macroreentry circuit that can have variable atrial anatomy. It is often treated by catheter ablation, the success of which depends upon the correct determination of the electroanatomic circuit, generally through invasive electrophysiological (EP) study. We hypothesized that machine learning (ML) methods applied to the diagnostic 12-lead surface electrocardiogram (ECG) could determine the specific circuit prior to any invasive EP study.
Methods:
The 12-lead ECGs were reduced to eight independent leads: I, II, V1 – V6. Through an algorithm using ventricular complex cancellation methods, windows of atrial activity in the ECG were uncovered and spectra were generated. Three ML classifier approaches were applied: Support Vector Machine (SVM), Random Forest (RF) and k-Nearest Neighbors (KNN), and their outputs combined using soft voting.
Results:
Ten-second surface ECGs taken from 419 AFL patients prior to invasive EP study and ablation were analyzed retrospectively. Of the 419 patients, 285 had typical cavotricuspid isthmus (CTI)-dependent AFL, 41 had atypical right-atrial AFL and 93 had left-atrial AFL, as determined during the subsequent EP study. Lead V5 was found to be most useful giving a test accuracy of 98% and f1 score of 0.97.
Conclusion:
We conclude that ML methods have the potential to automatically determine the AFL macroreentry circuit from the surface ECG.
Significance:
The AFL classification method presented in this investigation achieves 95+% accuracy on an unbalanced inter-patient dataset which has important clinical applications.