CNN Variational autoencoders' reconstruction ability of long ECG signals
Keywords: VAE, CNN, electrocardiogram, reconstruction, compression, interpretability
TL;DR: A manifold encoding decoding architecture of CNN VAE for long ECG signals which facilitates interpretable model design.
Abstract: Can variational auto-encoders (VAEs) generate flexible continuous latent space for long electrocardiogram (ECG) segments and reconstruct the input? A folded VAE architecture is introduced in this study which is able to encode long ECG segments by splitting an input segment into folds and process them in sequence using a narrow field-of-view in the encoder and concatenate them at the end, instead of processing the long segment at a time. The VAE decoder follows similar folding and concatenation strategy for reconstruction of the original ECG segments. The proposed folded VAE architecture is able to generate better reconstruction of long 30-second ECG segments compared to unfolded classical VAE approach which often produce trivial reconstruction of long ECG segments. Experimental results show that the latent representation generated by our folded VAE architecture not only retains rich compressed information but also aids designing interpretable models by providing decision-making insights.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10227
Loading