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

Darek Rubio Bizcaino, Kamil Bujnarowski, Max Matyschik, Henry Mauranen, Ruhui Zhao, Pietro Bonizzi, Joël Karel
Department of Data Science and Knowledge Engineering, Maastricht University


Abstract

Aims: Body surface potential mapping (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 less electrodes and reconstructing signals from the missing electrodes with a neural network. The minimal configuration consists of the electrodes that are most relevant for reliable reconstruction. Methods: A variational autoencoder was trained on BSPM recordings from the Nijmegen-2004-12-09 dataset available on EDGAR (BSPM system with 65 channels; 14 different procedures from 1 subject) and used to reconstruct the signals of all 65 channels from a reduced number of input channels. The autoencoder was trained on a procedure, and accuracy was assessed by computing reconstruction errors on data from a different procedure. Additionally, the effect of systematically increasing numbers of missing electrodes and the corresponding reconstruction errors was studied. Finally, a qualitative comparison with the 12-lead ECG was made, by selecting the 12 most important leads and mapping their locations on the upper torso. Results: The optimal settings for the autoencoder was 2 seconds rolling window and a latent dimension of 24. The variational autoencoder could reconstruct sufficiently well BSPM signals from the test set using as few as 12 electrodes (average mean square error 11.210 ± 0.408 over 100 predictions). Minimal configuration electrode positions coincided well with those of precordial leads from the 12-lead ECG, with additional positions appearing on the lower back. Conclusion: Results shows that by using an autoencoder it is possible to achieve accurate reconstruction of 65-BSPM channels from a minimal configuration of 12 electrodes on the body surface. The positions of these electrodes overlap well with the standard 12-lead ECG, with the addition of 2 electrodes on the lower back.