Artificial intelligence to reduce artifact in cardiac electrophysiological signals. (3rd October 2022)

• Record Type: Journal Article
• Authors: Ruiperez-Campillo, S; Deb, B; Feng, R; Ganesan, P; Tjong, F V Y; Clopton, P; Rogers, A J; Narayan, S M
• Is Part Of: European heart journal. Volume 43(2022)Supplement 2
• Publication Date: 2022-10-03
• Subjects: Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehac544.422
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background/Introduction: Signals in Electrophysiology cases are often noisy despite laboratory shielding and filtering, and current noise-reduction methods are suboptimal. Template matching can identify a "nearest type" of electrogram, but libraries of signal shapes may be unavailable. Beat averaging can reduce noise but obscures beat-to-beat variations and is not optimal to analyze dynamically changing signals, such as when moving a catheter in the heart. Smoothing reduces noise yet blurs high frequency components. Purpose: We set out to test if machine learned autoencoders could reduce noise in single beats without requiring massive training data or beat libraries. Specifically, we hypothesised that noisy electrograms in small datasets of atrial signals could be de-noised using an encoder-decoder neural network (NN) using transfer learning of machines trained to recognize key features in larger datasets of related signals. Methods: We applied NN to monophasic action potentials (MAPs), because they have visually verifiable shapes. The NN was first trained to reconstruct 5706 left and right ventricular MAPs in 42 patients (67±13y; Fig. 1A). Transfer learning was then used to apply the NN to a much smaller dataset of 641 atrial MAPs in 21 patients (67±5y, 13 women; Fig. 1B, D, F). Results: NN reconstructed atrial MAPs with a Pearson correlation of 0.87±0.11. After fine-tuning, NN reconstruction accuracy improved dramatically (Pearson 0.99±0.01; p<0.001). In Fig. 1B–G the NN learned key MAP features (upstroke, triangular descent, terminus) and thus could eliminate ventricular artifact and electrical circuit noise without specific training or manual annotation. Conclusion: Machine learned autoencoders are a novel and powerful approach to de-noise electrophysiological signals in a dynamic, beat-to-beat fashion. The ability to learn fundamental signal features from models trained in large datasets, and apply them via transfer learning to small datasets in different heart chambers may have wide ranging applications for automated signal annotation, mapping and ablation. Funding Acknowledgement: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): NIH