Improving Flutter Localization Performance by Optimizing the Inverse Dower Transform

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


Previous research showed that utilizing features derived from vectorcardiographic (VCG) loop parameters of atrial flutter (AFL) F waves allowed classification of right or left circuit localization. It is known that the Inverse Dower Transform (IDT) matrix, used to obtain the VCG from the ECG was obtained with strict assumptions on the volume conductor. Hence, it does not reflect real-life use cases. In this study, we aim to correct this inaccuracy by finding an optimal transform to be coupled with the IDT to maximize classification accuracy using previously described features.


56 records of AFL were used (31 right, 25 left). F waves were detected and segmented using a detector described previously. Only pure F waves not overlapped with QRST complexes were considered. Segmented F waves were transformed into VCG loops using IDT and corrected for respiratory motion using a technique detailed previously.


A parametrized model for the transformation of the IDT accounting for scaling and rotation of the three individual leads (XYZ) was proposed. However, no closed-form solution exists. The problem is treated instead using global optimization approach. Cuckoo search is a simple iterative, multi-instance search procedure and was used to find the optimum parameter values, using maximum SVM classifier accuracy as the objective. At each iteration, the transform was calculated and applied to all 56 recordings, VCG loop features were calculated, and an exhaustive wrapper search of feature combinations were performed.


Results showed an increase of up to 4% for maximum SVM accuracy of up to 10 features. Using logistic regression classifier on the same transformed features, maximum performance of 0.98 was achieved for up to 10 features (before transformation: 0.93). Inspection of the values showed that the transform acts most on Lead X (scaled by a factor of 0.72) and Lead Z (rotation of 12.52 degrees).