Interpretation and Further Development of the Hypnodensity Representation of Sleep Structure

Objective: Data acquired during a sleep recording is typically compressed into a hypnogram; a visual representation of manually annotated sleep stages over the night. Recently, a richer hypnodensity representation was proposed that provides a probability distribution over these stages at each point in time. In this work we investigate how to interpret a hypnodensity plot, and reveal its implicit assumptions. We, moreover, seek alternative representations to acquire additional information about continuities in the sleeping brain.

Approach: We recap softmax classification theory, and empirically validate the interpretation of a hypnodensity plot. Unsupervised learning and the non-linear softmax activation are studied to find representations that are less dependent on the manual sleep staging decision process. Experiments are performed both in a synthetic setup, and on sleep recordings.

Main results: A hypnodensity plot, predicted by a supervised classifier, represents the probability with which the sleep expert assigned a label to an epoch. It thus reflects annotator behaviour, and is thereby only indirectly linked to underlying continuous dynamics of the brain. Unsupervised training was shown to result in hypnodensity plots that were less dependent on this annotation process. Moreover, pre-softmax predictions were found to better reflect continuous brain dynamics than the post-softmax counterparts (i.e. the hypnodensity plot).

Significance: This study provides insights in, and proposes new, representations of sleep that may enhance our comprehension about sleep and sleep disorders.