Local activation time sampling density for atrial tachycardia contact mapping: how much is enough?

Abstract

AimsLocal activation time (LAT) mapping forms the cornerstone of atrial tachycardia diagnosis. Although anatomic and positional accuracy of electroanatomic mapping (EAM) systems have been validated, the effect of electrode sampling density on LAT map reconstruction is not known. Here, we study the effect of chamber geometry and activation complexity on optimal LAT sampling density using a combined in silico and in vivo approach.Methods and results In vivo 21 atrial tachycardia maps were studied in three groups: (1) focal activation, (2) macro-re-entry, and (3) localized re-entry. In silico activation was simulated on a 4×4cm atrial monolayer, sampled randomly at 0.25–10 points/cm2 and used to re-interpolate LAT maps. Activation patterns were studied in the geometrically simple porcine right atrium (RA) and complex human left atrium (LA). Activation complexity was introduced into the porcine RA by incomplete inter-caval linear ablation. In all cases, optimal sampling density was defined as the highest density resulting in minimal further error reduction in the re-interpolated maps. Optimal sampling densities for LA tachycardias were 0.67 ± 0.17 points/cm2 (focal activation), 1.05 ± 0.32 points/cm2 (macro-re-entry) and 1.23 ± 0.26 points/cm2 (localized re-entry), P = 0.0031. Increasing activation complexity was associated with increased optimal sampling density both in silico (focal activation 1.09 ± 0.14 points/cm2; re-entry 1.44 ± 0.49 points/cm2; spiral-wave 1.50 ± 0.34 points/cm2, P < 0.0001) and in vivo (porcine RA pre-ablation 0.45 ± 0.13 vs. post-ablation 0.78 ± 0.17 points/cm2, P = 0.0008). Increasing chamber geometry was also associated with increased optimal sampling density (0.61 ± 0.22 points/cm2 vs. 1.0 ± 0.34 points/cm2, P = 0.0015).ConclusionOptimal sampling densities can be identified to maximize diagnostic yield of LAT maps. Greater sampling density is required to correctly reveal complex activation and represent activation across complex geometries. Overall, the optimal sampling density for LAT map interpolation defined in this study was ∼1.0–1.5 points/cm2.

Physics-Informed Neural Networks for Cardiac Activation Mapping

“…We train the network for 50,000 ADAM iterations with a batch size N mb = 100 and then train with the L-BFGS method [28]. We compare our method against three other methodologies: a neural network with the same architecture and parameters except without including the physics, linear interpolation [2], and Gaussian process regression [3,29]. In the neural network without physics, we compute the conduction velocity analytically as V = 1/ ∇T .…”

“…In the neural network without physics, we compute the conduction velocity analytically as V = 1/ ∇T . In the linear interpolation case, we use the scatteredInterpolant function from MATLAB with linear extrapolation [2]. We compute the conduction velocity by approximating the gradient of the activation time with finite differences on a regular grid across the domain.…”

“…This process is repeated at multiple sites to cover the entire atrium. Finally, these measurements are interpolated to create a complete electro-anatomic map of the chamber [2]. The most common approach to interpolate the data is to use linear functions [2,3] or radial basis functions [4].…”

“…Finally, these measurements are interpolated to create a complete electro-anatomic map of the chamber [2]. The most common approach to interpolate the data is to use linear functions [2,3] or radial basis functions [4]. There is a recent focus into incorporating the uncertainty associated with these maps [3,5].…”

“…There is also uncertainty in the interpolation method that is used, which can be relevant particularly in regions of low data density. However, most current techniques [2,3] ignore the underlying physics of the electrical wave propagation [6]. This can result in unrealistic interpolations with artificially high conduction velocities.…”