Determination of the excitation origin in the ventricles from the ECG using support vector regression

• Published in: Current directions in biomedical engineering Vol. 3; no. 2; pp. 257 - 260
• Main Authors: Pilia, Nicolas, Ritter, Christian, Potyagaylo, Danila, Schulze, Walther H. W., Dössel, Olaf, Lenis, Gustavo
• Format: Journal Article
• Language: English
• Published: De Gruyter 01.09.2017
• Subjects: ECG; ECG Imaging; inverse problem; machine learning; Support vector regression
• ISSN: 2364-5504; 2364-5504
• DOI: 10.1515/cdbme-2017-0180

A common treatment of focal ventricular tachycardia is the catheter ablation of triggering sites. They have to be found manually by the physician during an intervention in a catheter lab. Thus, a method for determining the position of the focus automatically is desired. The inverse problem of electrocardiography addresses this problem by reconstructing the source of the ectopic beats using the surface ECG. This problem is ill-posed and therefore needs specific methods for solving it. We propose a machine learning approach for localisation of the ectopic foci in the heart to assist cardiologists with their therapy planning.We simulated 600 120-lead ECGs with different known excitation origins in the heart using a cellular automaton followed by a forward calculation. Features from the ECGs were used as input for a support vector regression (SVR). We assumed a functional relation between features from the ECG and the excitation origin. To benchmark SVR, we also used the well-known Tikhonov 0th order regularisation to reconstruct the transmembrane potentials in the heart and detect the location of the ectopic foci. Parameters for SVR and regularisation were chosen using a grid search minimising the error between estimated and true excitation origin. Compared to the Tikhonov regularisation method, SVR achieved a smaller deviation between estimated and real excitation origin evaluated with 6-fold cross validation. Future work could investigate on the behaviour on data from simulations with other torso and electrophysiological models, the influence of other methods for feature extraction and finally the evaluation with clinical data.