Neural Network-Based Matrix Completion for Minimal Configuration of Body Surface Potential Mapping

Abstract

Body surface potential mapping (BSPM) provides high spatial resolution recordings of the electric potential of the heart on the body surface. BSPM can involve up to 200 electrodes, in contrast to standard 12-lead ECG. The costs and complexity of a BSPM procedure are a limiting factor to its use in clinical practice. Both can be reduced by using fewer electrodes and reconstructing signals from the missing electrodes with an artificial neural network. The minimal configuration consists of the electrodes that are most relevant for reliable reconstruction. We propose an architecture for a variational autoencoder, trained on BSPM procedures from the Nijmegen dataset: EDGAR [1], to reconstruct a full 65-lead system from a reduced number of input electrodes. Further, we determine the effect of an increased numbers of missing electrodes on the corresponding reconstruction error, and show that it is possible to achieve a good 65-lead reconstruction from as few as 12 electrodes. We consider the implication of our research in the scope of current BSPM practice, as well as the limitations of using neural networks for this task.

Feasibility of ECG Reconstruction from Minimal Lead Sets Using Convolutional Neural Networks

“…Convolutional VAEs (ConVAEs) based on the structure in [2] were used to investigate the lead significance and amount of dimensionality reduction on the input data. For this purpose, the investigated hyperparameters were the repetition of the convolutional blocks, dense blocks and the size of the latent layer.…”

“…The input data was windowed for these experiments based on the parameters found in [2]. This ensures manageable training times without compromising the reconstruction quality.…”

“…We devised two CNN architectures to reconstruct missing ECG data. This choice was made based on our previous work [2]. Existing architectures such as AlexNet [8] were used as inspiration for our design due to their performance in image-related tasks.…”

“…[1] investigated lead reduction in 64-lead body surface potential maps (BSPMs). Other ECG procedures have been shown to contain redundant information [2], which shows promising results for applying neural networks in compressing and reconstructing ECG information. [2] implemented neural networks in the context of 65-lead BSPMs and significantly reduced the lead count required for reliable reconstructions.…”

“…This raises the question if similar approaches could be applied to the traditional 12-lead ECG to achieve more portable monitoring system. In [2], a heuristic method based on VAE is used for lead selection and assessment of lead significance for the reconstruction process. In this study VAEs were analyzed directly to understand which leads had highest contribution to reconstruction quality.…”