Skip to search boxSkip to navigationSkip to main content

Optimizing Automated Sleep Staging

Output:
Thesis
Phd Thesis 1 (Research TU/e / Graduation TU/e)
Academic

Open access

Information

Output type

Output:
Thesis
Phd Thesis 1 (Research TU/e / Graduation TU/e)
Academic

Original language

English

Qualification

Doctor of Philosophy

Publication milestones

  • Published - 04/12/2025

Publication status

Published - 04/12/2025

Supervisors/Advisors

Place of publication

Eindhoven

Award date

04/12/2025

Publisher

Eindhoven University of Technology

Publication series

  • Publication series name: The TU/e Sleep Series
    Volume: 10
978-94-92303-49-3

Abstract

Polysomnography (PSG) is the gold-standard for sleep monitoring and sleep disorder diagnostics. However, PGSs usually only cover a single-night and are therefore unable to capture night-to-night variability. Furthermore, the obtrusive sensors can impact sleep quality. To address these limitations, wearable monitoring technologies have been developed. Automated sleep stage classification (ASSC) is typically needed for these technologies due to the lack of standardized guidelines for manual scoring with alternative or reduced sensor montages. Various deep learning ASSC models are available, utilizing e.g., cortical activity from wearable electroencephalography (EEG), or heart rate variability (HRV) from wrist-worn photoplethysmography (PPG). Applying wearable technologies and their accompanying ASSC models to the clinical population is complex. A major barrier is the need for large-scale data for model training, including concurrent gold-standard PSG. While other available datasets can be used to overcome limited data availability, suboptimal performance may occur from mismatches between source dataset (on which the model is trained) and target dataset (on which the model is applied), such as differences in population characteristics (e.g., an unrepresented sleep disorder) or sensors (e.g., alternative sensors or channel locations). Therefore, this dissertation aims to contribute to the optimization of wearable technologies and their ASSCs in the clinical population, in terms of classification performance, data availability for model training, and the use of new analysis methods that are enabled by ASSCs. In general, there are three strategies to train a ASSC model: pretraining (i.e., training on the source dataset), training-from-scratch (i.e., training on the target dataset), and finetuning (i.e., training on the source dataset, finetuning on the target dataset). In Chapter 2, we systematically assessed which training strategy is favourable for training a single-channel conventional EEG ASSC model when population and channel mismatches between the source and target dataset are introduced. We showed that, when channel and/or population data mismatches cause suboptimal sleep staging performance, a finetuning approach can yield similar to superior performance compared to training-from-scratch, while requiring fewer sleep recordings. Furthermore, our research indicated that unlike insomnia and obstructive sleep apnea (OSA), REM sleep behavior disorder (RBD) demands targeted model training due to sleep-related distinct cortical characteristics. Next, we applied the obtained knowledge on model training to a scenario in which limited data availability and data mismatches are common obstacles, namely the implementation of ASSC for a new sensor. In Chapter 3, we applied each of the three model training strategies to single-channel wearable EEG recordings that we acquired in a heterogeneous patient population. We showed that leveraging source datasets through transfer learning is essential in data-limited scenarios to optimize sleep staging in on a new, wearable sensor. Additionally, our results indicated similarity in data characteristics between conventional and wearable EEG. Moreover, results suggested the combination of the wearable EEG, ASSC model, and training strategy can be viable for sleep monitoring in the clinical population. In the previous chapters, we tuned ASSC models to population- and sensor-specific characteristics. In Chapter 4, we wanted to stretch the idea of model tuning and questioned whether it was possible to optimize HRV-based ASSC performance by tuning to subject-specific characteristics. We acquired concurrent wrist-worn PPG and wearable EEG during prolonged in-home sleep monitoring, and finetuned the HRV-based ASSC model on the individual’s PPG data combined with the wearable EEG labels. We showed that a personalized HRV-based ASSC model outperformed a generalized model after three or more training nights, where especially younger individuals and insomnia patients benefitted from personalization. Results indicated individual sleep stage-specific autonomic representations were learned by tuning the model to subject-specific thresholds in the sympathetic-parasympathetic balance.In general, ASSC models classify a discrete sleep stage based on the stage with the highest probability, mirroring the process in manual scoring. However, ASSCs offer novel opportunities to study sleep structure, including utilizing the underlying probability distribution (i.e., the hypnodensity). In Chapter 5, we used hypnodensity-derived metrics to improve understanding of the observed suboptimal ASSC performance in RBD, a sleep disorder strongly associated with Parkinson’s disease. Analysis was performed with both EEG- and HRV-based ASSC models, which leverage different physiological processes that are potentially impaired in individuals with RBD. For both models, we showed that in RBD, stage classification was more ambiguous, and stage transitions had a more continuous profile, potentially complicating stage classification. These results suggest that while conventional discrete sleep stage analysis can be inadequate in capturing sleep macrostructure characteristics, hypnodensity analysis can enhance the characterization of (RBD) sleep.In sum, this dissertation explored how wearable technologies and their ASSC models can be implemented effectively in the sleep-disordered population by tuning ASSC models to sensor-specific, population-specific, and subject-specific characteristics. These automated models enable probability-based analyses to reveal new characteristics of sleep macrostructure, which can improve our understanding of sleep staging difficulties and pathophysiology, such as in RBD.

Bibliographical Note

Proefschrift.

Documents and Links

Final published version, 9.97 MB

Metrics

Metrics

Download statistics
Download count
114