Author:

Keywords:

STADIUS-24-148

Abstract:

Sleep is a fundamental biological process essential for overall health and well-being. Studying it to diagnose sleep-wake disturbances requires overnight measurement of physiological signals, and annotation of these measurements through sleep staging. Traditionally, sleep studies are performed in the hospital using specialized polysomnography equipment and labor-intensive manual sleep staging. Wearable devices measuring brain activity and other physiological signals provide a more comfortable alternative and enable at-home sleep studies, but they necessitate automated sleep staging approaches. This thesis focuses on the development and enhancement of automated sleep staging methods, which are crucial for home-based sleep studies and can reduce the workload in hospital-based sleep studies. The primary aim is to address challenges in performance, generalization, and trust in these methods to facilitate their adoption in clinical practice, and to validate the algorithms in clinical datasets. These objectives are attained in three main parts. The first part explores techniques to improve the generalization of automated sleep staging methods trained on polysomnography, to achieve better performances on data from wearables. Two key methods are introduced: (1) supervised transfer learning via feature matching, which enhances performance using small amounts of labeled data by aligning features from different domains, and (2) adversarial domain adaptation, which minimizes domain mismatch by leveraging unlabeled data, and can be further improved with pseudo-labels and careful source domain selection. These methods demonstrate significant improvements in multiple clinical datasets. The second part tackles the issue of trust and transparency in automated decision support systems by incorporating uncertainty estimation. The U-PASS framework is developed to optimize the reliability of sleep staging models through uncertainty-based training data selection and curation, active learning with clinician feedback, and deferral of uncertain predictions to experts. This approach enhances both model performance and user trust, especially in clinical settings. The application of U-PASS to wearable sleep monitoring highlights the need to adjust handling of uncertainty in home-based wearable sleep monitoring applications. The third part validates the developed algorithms across various clinical populations, demonstrating their effectiveness in characterizing sleep in patients with epilepsy, Alzheimer's disease, and Parkinson's disease. The results highlight that sleep staging can not only reproduce established clinical correlations but also contribute to therapeutic applications and aid in the diagnostic screening of brain disorders. Notably, sleep measurements using wearables are shown to be valuable in Alzheimer's disease screening. Depending on data availability, different transfer learning approaches are recommended: fine-tuning for polysomnography datasets and either feature matching or adversarial domain adaptation for wearable datasets. In summary, this work contributes significantly to the integration of artificial intelligence in sleep medicine by addressing performance and generalization challenges, improving trust and transparency, and validating methods in clinical settings. The proposed solutions pave the way for more reliable and widespread use of automated sleep staging tools, potentially transforming both clinical diagnostics and large-scale sleep studies.