Improving Flutter Localization Performance by Optimizing the Inverse Dower Transform

A previous study showed the possibility to localize right or left flutter circuit origin using variability contained in vectorcardiographic loop parameters. The Inverse Dower Transform, used to obtain the vectorcardiograms is based on a very simplistic torso conductor model, and hence not optimized. The present study aims to optimize the transform to maximize classifier accuracy. A parametric optimization model was proposed, as well as an optimization scheme. Model parameters were obtained by iteratively optimizing the linear SVM classifier accuracy until convergence. The goal can be shown to be multimodal and non-smooth. Therefore, a multi-instance and derivative-free method was considered. Previous dataset of 56 flutter recordings (31 right, 25 left) was used, considering only non-overlapped and respiratory motion-corrected F loops. For the SVM classifier, a 3.8% increase in accuracy was observed (max 0.95). When the logistic regression classifier was used, an increase of 7.8% was observed (max 0.98). Comparison to a targeted transform previously developed showed an improvement by 17−19%. Observation of the model parameter values showed amplitude reduction applied to Lead X and rotation applied to Lead Z.