On Tuesday 31 May 2022, Andreas Brink-Kjær will defend his PhD thesis "Design of Interpretable End-to-End Deep Learning Models for Diagnosis of Sleep Disorders and Sleep Quality Evaluation".
Supervisor: Helge B.D. Sørensen
Co-Supervisor: Poul Jørgen Jennum and Emmanuel Mignot
Assessment committee:
- Associate Professor Kaj Bjarne Jakobsen, DTU Department of Elektro Technology
- Professor Henrik Karstoft, Aarhus University
- Assist. Professor Theerawit Wilaiprasitporn
Chairperson:
- Associate Professor Kaj-Åge Henneberg, DTU Health Tech
Title: Design of Interpretable End-to-End Deep Learning Models for Diagnosis of Sleep Disorders and Sleep Quality Evaluation
Abstract:
Sleep clinics worldwide evaluate millions of patients every year. The standard diagnostic test for sleep disorders is a visual analysis of physiological measures during sleep such as brain activity (EEG). Currently, both visual and even automatic analyses follow a set of guidelines that may neglect relevant information that could improve the diagnosis of certain sleep disorders such as REM sleep behavior disorder (RBD). Diagnosis of RBD is particularly important as it is the strongest biomarker of Parkinson’s disease (PD).
This thesis aims to utilize interpretable end-to-end machine learning to evaluate sleep health and identify sleep disorders directly from raw data, thereby considering all relevant information. First, automatic scoring of arousals during sleep was validated and used to find biomarkers of narcolepsy, RBD, and PD. Second, a model of healthy aging based on sleeping patterns was developed. An elevated age estimate was associated with increased all-cause mortality risk; thus, it is a biomarker of overall health. This same framework was optimized to identify RBD patients with high accuracy. Lastly, a dual approach using wrist actigraphy and questionnaires was developed to screen for RBD. This approach was highly specific, which makes it suitable for population-based studies.
Altogether, this thesis presents new methods for analyzing sleep recordings to extract more relevant information from the data. The methods could be used to assist sleep clinicians in profiling their patients and accelerate research in neuroprotective treatments for PD.