Hypnodensities: Predicting probability graphs of sleep stages

The macro-structure of sleep dynamics is typically represented by a conventional hypnogram, which assigns a fixed label to every consecutive, non-overlapping 30 s windows of polysomnography (PSG) data. This convention is problematic for many reasons, one of which is the assumption that the underlying PSG content remains constant during the entirety of the epoch, which may not reflect true, underlying physiological activity. Current AASM guidelines on sleep stage scoring do not currently take this into account, resulting in epochs potentially including content from several sleep stages. This introduces uncertainties in the sleep stage scoring, which are not captured by the fixed-label hypnogram. In this roundtable session, we will present the hypnodensity as an alternative to conventional hypnograms. We posit that this will provide further insight into the dynamics of sleep by representing the hypnogram in a probabilistic sense. We will argue that a probabilistic representation of the hypnogram is desirable in characterizing sleep dynamics since the probabilistic representation of sleep epochs allow for mixed-state interpretations of sleep stages. We will show how this mixed-state representation can be used in downstream feature extraction schemes combined with probabilistic modeling for assisting in the diagnosis of narcolepsy type 1. Furthermore, by basing the hypnodensity prediction on an ensemble model, we can assess the uncertainty in sleep stage classification estimates by looking at the variance across the ensemble predictions. We will also discuss the important point of proper calibration of hypnodensities to ensure proper interpretation of sleep stage probabilities, as well as potential future research directions and applications.