Artificial Intelligence in Regulated Domains : Lessons from Sleep Medicine

Abstract

This thesis investigates how artificial intelligence (AI) and machine learning (ML) can be integrated into the regulated workflow of polysomnography-based sleep architecture analysis while ensuring outcomes are understandable and defensible. Rather than focusing solely on sleep-stage classification accuracy, the thesis argues that different parts of the workflow require different combinations of machine learning, explicit rules, uncertainty representation, and human oversight. Across six papers, the thesis develops AI methods for sleep analysis and, through simulation and empirical studies, examines how their outputs behave within the clinical workflow. These include an LLM-guided neural architecture search framework for autonomously developing multimodal sleep staging models; a rulebased automatic sleep staging system derived from clinical scoring rules, capable of producing auditable justifications in clinical language; simulation and empirical studies of how sleep staging error propagates to downstream sleep metrics (such as total sleep time and time spent in deep sleep); and methods for estimating sleep metrics without performing sleep stage classification. The results show that strong epoch-level agreement does not translate uniformly into reliable downstream sleep metrics and that some sleep metrics vary substantially across expert scorers. They further show that preserving probabilistic uncertainty can improve sleep metric estimation, that direct sleep metric estimation can be competitive with conventional hypnogram-based pipelines, and that parts of the formal scoring workflow can be translated into executable and auditable procedures. Interviews with sleep professionals also show that effective AI decision support requires interactivity, familiar clinical language, and system designs that preserve the human expert’s accountability. These findings are synthesized in the MARBLE framework (Multiple Abstraction-Level Rule-Based Learning Engine), which treats AI decision support in sleep medicine as the coordinated use of rule-based and pattern-based methods across multiple abstraction levels, from raw physiological signals and micro-events to sleep stages, sleep metrics, and higher-level interpretation. The thesis concludes that AI integration in clinical sleep analysis should not be framed as the automation of one prediction task, but as the careful design of decision support across a rule-governed workflow. Although developed in sleep medicine, this may also have implications for AI decision support in other regulated domains.