On Tuesday 22 September, Alexander Neergaard Olesen will defend his PhD thesis: Deep Learning Methods for Clinical Sleep Analysis
Time: Tuesday 22 September, at 15:00
Place: The defence will take place via Zoom.
Zoom: Register for the defence here: https://dtudk.zoom.us/meeting/register/u5Mpc--uqz4qE9HrIiIZM3axbyqeQR9c0Mkg
Principal supervisor: Associate Professor Helge B. D. Sørensen
Co-supervisor: Professor Poul Jørgen Jennum
Co-supervisor: Professor Emmanuel Mignot
Examiners:
Associate Professor Kaj Bjarne Jakobsen, DTU Elektro
Professor Søren Holdt Jensen, Aalborg University
Professor Thomas Penzel, Charité – Universitätsmedizin Berlin
Chairperson at the defence:
Associate Professor Kaj-Åge Henneberg
Abstract:
Sleep disorders are common in the general population and have major implications for personal health and mortality including increased risk of cardiovascular, metabolic, and psychiatric complications, while also being a major economic burden on society.
The current standard for sleep disorder diagnosis is based on visual analysis of brain activity (EEG), eye movements (EOG), muscle activity (EMG) and cardiorespiratory variables recorded during sleep. This process is expensive, timeconsuming, and prone to subjective interpretation of analysis guidelines. Introducing automated analysis methods in the clinics has the potential to overcome these issues by virtue of being fast, consistent, and objective.
This thesis proposes new automatic methods for clinical sleep analysis based on artificial intelligence algorithms. First, two models for automatic sleep stage classification based on deep neural networks and EEG, EOG, and EMG data are proposed. Such models are crucial, as the dynamics of sleep stages can be indicative of certain sleep disorders. Second, a model for detection of sleep microevents using raw EEG, EOG, EMG, respiratory variables, and state-of-the-art computer vision algorithms is presented, which can precisely annotate several different types of events. Excessive amounts of sleep micro-events can also be indicative of certain sleep disorders, making such a model valuable for clinical use. Third, a model is presented for fast and accurate classification of narcolepsy, a sleep disorder where the basic mechanisms regulating sleep and wake are disrupted. This model was able to identify patients with narcolepsy as well as current gold standard methods, while being faster and cheaper in use.
In conclusion, this thesis presents new automatic methods for clinical sleep analysis based on artificial intelligence. Compared to current methods, the proposed models have the advantage of being faster and more reliable, allowing clinicians an increased focus on patient care.