Self-supervised learning for EEG-based automatic sleep staging

Estevan Tomás, Emilio
Montesano del Campo, Luis (dir.)

Universidad de Zaragoza, EINA, 2025
Informática e Ingeniería de Sistemas department, Lenguajes y Sistemas Informáticos area

Máster Universitario en Robótica, Gráficos y Visión por Computador

Abstract: Sleep is crucial within the spectrum of human health. A significant number of people suffer from sleep-related disorders, representing a growing public health concern. Polysomnography (PSG) is the gold standard for sleep evaluation, which begins with the sleep staging procedure. PSG recordings are segmented into 30-second intervals and categorized into one of the well-defined sleep stages. However, it is a very expensive process performed manually by expert technicians, and is highly sleep-invasive on patients due to the large number of sensors placed on the body. Recent advancements in deep learning are revolutionizing automatic sleep staging, achieving, and even surpassing, expert-level accuracy, thereby overcoming some limitations of the manual scoring process. Given the drawbacks of traditional sleep monitoring and the potential of learning-based sleep staging approaches, Bitbrain Technologies has developed a fully textile-based, wearable, and user-friendly EEG garment to serve as a scalable solution for sleep analysis, also including a deep learning system with medical-grade precision trained on labeled datasets. Nevertheless, deep learning methods are notoriously data-hungry, necessitating from large volumes of labeled data for effective training. The process of signal annotation is as resource-intensive as the manual scoring procedure carried by technicians. Therefore, the mass adoption of home sleep monitoring enabled with this novel device generates a massive amount of EEG data that is impossible to label, resulting in enormous datasets that Bitbrain cannot leverage for supervised training of its models. As a consequence, the objective of this work is to develop and evaluate self-supervised learning (SSL) methods as a pre-training step within the learning pipeline to learn electroencephalogram representations from unlabeled data, reducing the costs of the manual annotation process. To achieve this, state-of-the-art deep learning and self-supervised methods for automatic sleep staging are reviewed, implemented, and evaluated across different scenarios, comprising two different datasets recorded using Bitbrain's device: the HOGAR study, an unlabeled dataset acquired under domestic conditions to quantify cognitive function in populations at risk of dementia, and the BOAS dataset, a labeled resource collected in a controlled laboratory environment with the aim of closing the gap between traditional clinical PSG technologies and portable EEG solutions. As a result, the transferability between both datasets was thoroughly assessed and successfully demonstrated, obtaining significant improvements by incorporating SSL techniques into the learning pipeline. This is further supported by a feature visualization analysis performed using t-SNE and UMAP tools. On the other side, a comprehensive understanding of the performance and limitations of SSL relative to the labeled data available was provided, exhibiting particularly pronounced accuracy gains in low-data regimes. Consequently, this work serves as a critical step forward in the design and validation of label-efficient, self-administered, and large-scale automatic sleep monitoring systems in home environments under uncontrolled conditions, advancing in a scalable solution for the substantial proportion of the word population suffering from serious sleep disorders that require medical attention.

Tipo de Trabajo Académico: Trabajo Fin de Master