Exploration of sleep events in the latent space of variational autoencoders on a breath-by-breath basis

Abstract

In this exploratory paper, we attempt to address a growing demand for unsupervised machine learning techniques on sleep data by applying a variational autoencoder on respiratory sleep data on a breath-by-breath basis. We transform respiratory signals into a latent representation and cluster them together using K Means clustering. We calculate the cluster preference of scored events and attempt to explain their position in the latent space. We show that a variational autoencoder can accurately reconstruct three respiratory signals from individual breaths despite being sampled through a latent dimension 384 times smaller than the input data. Our results also indicate that respiratory events in particular show a tendency to cluster together in the latent space despite a purely unsupervised learning approach. Finally, we lay the groundwork for future work made possible in this paper.

More information

Identifiers

ORCID for Maria Oskarsdottir: orcid.org/0000-0001-5095-5356

Catalogue record

Export record