Representation learning for multivariate time series : artefact analysis on sleep recordings

The application of machine learning is becoming a major area of interest within the field of sleep science. The domain of sleep science is shifting, due to the automation machine learning introduces. Automatic analyses are vital to minimise diagnostic time and provide as many people as possible with the help they require. Artefact detection plays a key role in the development of automatic tools especially concerning the potential misdiagnosis by automatic models. This study explores the use of a representation reinforcement learning model on multivariate time series for automatic analyses on artefacts in sleep study recordings. This study concerns sleep data from a dataset of 100 full-night recordings from a standard equipment setup for diagnosing sleep-related breathing disorders. The scoring was performed by expert sleep analysts using a new proposed artefact standard. Two approaches were developed, both an unsupervised approach and a supervised one. The unsupervised approach used clustering on the features that resulted from the representation learning model. To validate the resulting features several supervised classification models were implemented, and their performance evaluated. The finding suggested that representation learning has immense potential in the development of automatic analyses on data from sleep studies. The models validated the benefits in the application of representation learning on multivariate time series and both the unsupervised and supervised models showed high association to the artefact annotation. The models developed in this study can be further improved upon for even better results in the analyses of artefacts in sleep data.

Authors dc:creator

• Björn Breki Þorbjörnsson 1997-
• Haukur Húni Árnason 1997-
• Óðinn Hjaltason Schiöth 1998-

Contributors dc:contributor

• Háskólinn í Reykjavík

• Tölvunarfræði
• Vélrænt nám
• Svefnrannsóknir
• Computer science
• Machine learning
• Sleep
• Research

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/39348

OAI identifier oai:identifier

oai:skemman.is:1946/39348