Abstract
Frequent sleep stage transitions and abnormal sleep stage distribution are features of sleep disorders, including insomnia, sleep apnea, and narcolepsy, and have been associated with altered affective processing, including mood disorders. Research on the role of sleep stage transitions is nascent, and mixed operationalizations abound. In this comparative methods study, we overview the ways that prior research has operationalized sleep stage transitions, propose guidelines for four types of metrics, and compare the relevance of each for different analytic purposes. We then discuss three definitional and methodological "hard problems" for research on sleep stage transitions: bias due to scoring discrepancies, low temporal resolution, and opposing definitions of transitions. We discuss the pros and cons of several solutions that use machine learning (ML) scoring algorithms, with examples derived from the Stanford Sleep and Affect polysomnography dataset scored with validated ML algorithms (Stanford STAGES, U-Sleep, and YASA), and conclude with a call to return to descriptive physiological studies to shift the current framing of sleep stage transitions away from categorical state changes. This intends to lay the foundation for further insight into the role of sleep stage transitions in affective function and in clinical dysfunction.