Design of Interpretable End-to-End Deep Learning Models for Diagnosis of Sleep Disorders and Sleep Quality Evaluation

Sleep clinics worldwide evaluate millions of patients every year. The current gold standard diagnostic test for sleep disorders is a visual analysis of a polysomnography (PSG) recording, which follows guidelines from the American Academy of Sleep Medicine. Through visual analysis, various manual annotations are made based on patterns in electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), and cardio-respiratory signals. However, visual and even automatic analyses that follow these guidelines neglect some 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).
Through three research parts, this thesis aims to utilize interpretable end-to-end deep learning models and other machine learning-driven methods to better summarize sleep health and to assist clinicians in the diagnosis of RBD.
First, an automatic method for scoring sleep-wake periods and arousal was validated in PSG data recorded from patients with narcolepsy, RBD, and PD. These patient groups displayed various abnormal patterns of arousals that serve as biomarkers of disease, which could assist clinicians in the diagnosis of RBD and narcolepsy. Second, interpretable end-to-end deep learning models were used to i) estimate age, as a proxy for mortality risk, and ii) classify RBD. The age estimation model was tested in 11,998 PSGs and had a mean absolute error of 5.81 years. An age estimate higher than the chronological age was associated with high mortality risk, thus it constitutes a biomarker of health. Additionally, the end-to-end framework was optimized to detect RBD in a sample of 290 PSGs. The model identified RBD patients with an accuracy of 89 %. Each signal modality could provide an estimated RBD probability, which could be used by clinicians to profile each patient in more detail. Moreover, these methods were able to highlight regions of interest in the PSG and how the regions affected predictions.
Lastly, a machine learning-based screening method for RBD was designed using both wrist actigraphy and questionnaires that had a sensitivity of 87.8 % and specificity of 100 %. The high specificity makes the method 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 neurodegenerative disease.