Bechný, Michal (2025). From Automated Scoring to Digital Biomarkers: Computational Methods Towards Precision Sleep Medicine. (Thesis). Universität Bern, Bern
|
Text
25bechny_m.pdf - Thesis Available under License Creative Commons: Attribution-Noncommercial (CC-BY-NC 4.0). Download (34MB) | Preview |
Abstract
Sleep, together with diet and physical activity, is one of the fundamental pillars of health. Sleep disorders are highly prevalent, with rising incidence particularly among younger and economically disadvantaged populations, and they are closely linked to neurological, psychiatric, metabolic, and cardiovascular conditions. The clinical gold standard for assessing sleep is polysomnography (PSG), which records multiple biosignals such as brain activity, breathing, and movement during the night. Following established guidelines, every 30-second segment is scored by trained experts into one of five sleep stages, producing a stage-by-stage sequence called a hypnogram. From this representation, standard sleep metrics are calculated and combined with indices of breathing- or movement-related events to support diagnostics. While central to diagnosis, PSG is costly and labour-intensive, with manual scoring alone requiring up to two hours of expert time. With the growing use of Artificial Intelligence (AI), automated sleep scoring (ASS) powered by modern machine- and deep-learning algorithms achieves human-level agreement of 75–85%, but remains limited by the inter-rater variability inherent in training labels. This limits performance and requires mechanisms for effective human–AI collaboration. The first branch of this thesis (Chapters 2–4) addresses these challenges by developing methods for clinical integration of ASS, including uncertainty-guided oversight and a flexible statistical framework to quantify algorithmic bias. These approaches aim to support efficient physician review while promoting fairness, transparency, and reliability in clinical deployment. After improving the ASS process, the second branch (Chapters 5–7) focuses on deriving novel digital biomarkers from sleep-stage dynamics—an underutilised aspect of PSG with potential to capture subtle physiological signatures. Using clinical, observational, and prospective datasets, we investigated their value in sleep-disordered breathing, chronic fatigue and pain syndromes, and long-term cardiovascular risk. To address confounding and build predictive models, we employed causal inference methods, Bayesian networks, and forest-based classification and survival models, systematically examining the effects and risk profiles associated with sleep-stage transitions. These studies revealed disorder-specific alterations and showed that sleep-stage dynamics can also predict future cardiovascular events. Together, the findings demonstrate that sleep-stage dynamics represent a promising class of digital biomarkers that extend standard PSG metrics, improve risk assessment, and—when combined with wearable technology in the future—may enable unobtrusive yet sensitive long-term monitoring of diverse conditions.
| Item Type: | Thesis |
|---|---|
| Dissertation Type: | Cumulative |
| Date of Defense: | 24 October 2025 |
| Subjects: | 000 Computer science, knowledge & systems 600 Technology > 610 Medicine & health |
| Institute / Center: | 08 Faculty of Science > Institute of Computer Science (INF) |
| Depositing User: | Sarah Stalder |
| Date Deposited: | 12 Nov 2025 10:18 |
| Last Modified: | 12 Nov 2025 10:18 |
| URI: | https://boristheses.unibe.ch/id/eprint/6868 |
Actions (login required)
 |
View Item |