Noise reduction in electrophysiological signals using transfer machine learning
Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): NIH Background/Introduction Reducing electrophysiological signal noise is essential for diagnosis, mapping and ablation, yet most approaches are suboptimal. Template matching re...
Saved in:
Published in | Europace (London, England) Vol. 24; no. Supplement_1 |
---|---|
Main Authors | , , , , , , |
Format | Journal Article |
Language | English |
Published |
19.05.2022
|
Online Access | Get full text |
Cover
Loading…
Summary: | Abstract
Funding Acknowledgements
Type of funding sources: Public grant(s) – National budget only. Main funding source(s): NIH
Background/Introduction
Reducing electrophysiological signal noise is essential for diagnosis, mapping and ablation, yet most approaches are suboptimal. Template matching requires libraries of known signal types, that are difficult to obtain. Beat averaging can reduce noise, yet cannot be applied to single beats and obscures beat-to-beat variations. Beat smoothing can lose critical and subtle signal features. We set out to use neural networks (NN) based on encoder-decoders, which are able to extract key signal features and hence reconstruct them without noise and artifact.
Purpose
We hypothesised that electrograms with varying sources of artifact can be denoised using autoencoder neural networks. We further hypothesised that this could be achieved in a small data set by developing the method in a larger dataset of related signals, then using transfer learning. We tested this approach for atrial monophasic action potentials (MAPs) that have verifiable shapes.
Methods
The NN was first trained with 5706 left and right ventricular MAPs from 42 patients with ischemic cardiomyopathy (age 65±13y; fig 1.A): 60% for training, 20% (validation) and 20% (testing). Transfer learning and parameter-tuning were then used to apply this NN to a smaller sample of atrial MAPs (N=641, 21 patients, 67±5y, 13 women; fig D,F,H).
Results
The autoencoder was able to learn key features of MAPs, and hence reconstruct them without artifacts. NN learned ventricular MAPs with similarity coefficient 0.91±0.16, Pearson correlation 0.99± 0.01 (fig A) and learned key features (upstroke, triangular descent, terminus) to reduce noise (fig B-C). Applying this trained NN to atrial MAPs, the approach automatically eliminated ventricular artifact (fig E), high frequency noise (fig G), truncation (fig I), saturation and other artifacts. After fine-tuning, the NN reconstructed atrial MAPs with Pearson correlation = 0.99±0.01 (p<0.001).
Conclusions
Machine learned encoder-decoders are powerful tools that can automatically eliminate diverse types of noise in single beats by learning essential signal features. Transfer learning makes this possible without large datasets for training, even from signals in a different cardiac chamber. This approach may have far-reaching applications for mapping and ablation. |
---|---|
ISSN: | 1099-5129 1532-2092 |
DOI: | 10.1093/europace/euac053.125 |