Non-Invasive Localization of Atrial Flutter Circuit Using Recurrence Quantification Analysis and Machine Learning

Muhammad Haziq Kamarul Azman1, olivier Meste2, Decebal G. Latcu3, Kushsairy Kadir4
1Universite Cote d'Azur, Universiti Kuala Lumpur, I3S, 2Université Côte d'Azur, CNRS, I3S, 3Centre Hospitalier Princesse Grace, 4Universiti Kuala Lumpur


Atrial flutter (AFL) circuit localization is a crucial information, usually determined during the procedure itself. Knowledge of this information beforehand will help to improve procedural efficacy. AFL ECG recordings show the trajectory of the atrial dipole, which is cyclic due to the circular activation of the atrium. However, it is expected that the cycle-to-cycle properties of the trajectory is different in the right atrium than the left atrium due to the different physiology of the two atria. We investigate this difference using recurrence quantification analysis, suitable for capturing spatiotemporal information in the recording.

54 records of AFL were used (30 right, 24 left). F waves were detected and segmented from the ECG. Artifacts due to respiratory motion and T wave overlaps were removed accordingly from the F waves using previously described techniques. Recurrence signals were obtained from an unthresholded recurrence plots which uses normalized cross-correlation as a distance metric. The recurrence signal has a pseudo-periodic cosinusoidal form, with perfect correlation at lag=0. Two parameter series were constructed based on peak-to-peak amplitude (S1) and distance between successive maximum peaks (S2). 10 features were derived from statistics of each of the series (Mean, Standard Deviation, Variance, Skewness and Kurtosis). Two linear classifiers: logistic regression (LR) and SVM were used. A wrapper approach with exhaustive evaluation of possible feature combinations (from 1 to 10) was used to select the most relevant features.

The range of maximum accuracy for each classifiers were [0.72; 0.80] and [0.65; 0.80] from 1 to 10 features, for LR and SVM classifiers respectively. Maximum performance is achieved by both classifiers for a combination of 6 features. From the feature selection process, it was found that the mean of the series S2 was the most relevant feature (median of right vs. left AFL values: 230.63 vs. 203.5 ms (p<0.05)).