D-136 | Sleep stage classification using generalized ordinal patterns

Tools Development and Open Source Neuroscience
Author: Cristina Daiana Duarte | Email: cristinaduarte88@gmail.com

Cristina Daiana Duarte^1°, Marianela Pacheco^1°2°3°4°, Francisco Ramiro Iaconis^1°,  Claudio Augusto Delrieux^3°, Gustavo Gasaneo^1°

^1° Instituto de Física del Sur, Departamento de Física, Universidad Nacional del Sur (UNS) – CONICET, 8000 Bahía Blanca, Argentina
^2° Doctorado en Neurociencias, Universidad Nacional de Córdoba (UNC), Córdoba, Argentina.
^3° Instituto de Ciencias e Ingeniería de la Computación (ICIC), Universidad Nacional del Sur – CONICET, Bahía Blanca, Argentina.
^4° Universidad Provincial del Sudoeste (UPSO), Bahía Blanca, Argentina.

Sleep is a natural and reversible state. It is organized into cycles of about 120 minutes, consisting of rapid-eye-movement (REM) and non-REM (NREM) stages. NREM includes N1, N2 and N3 substages, each characterized by different physiological and brain activity patterns. In this context, polysomnography aims to identify these stages through the analysis of EEG signals, which is an essential tool for the diagnosis of sleep disorders. However, identifying and labeling sleep phases is a time-intensive task that requires considerable effort from experts. In this research we propose an unsupervised EEG signal analysis model that facilitates the identification of sleep stages efficiently and accurately.
We evaluate the statistical complexity of the EEG signal across sleep stages using the Generalized Ordinal Patterns (GOP), a generalization of the ordinal patterns of Bandt and Pompe’s permutation entropy. We determine the probability of each individual ordinal pattern, weighted by signal variance within the pattern, and raised to an entropic index. In this way, a given signal segment can be characterized by the set of weighted probabilities of each ordinal pattern, under the different entropic indices. A random forest classifier is trained on a labeled polysomnography dataset using these features. After training, the resulting classifier achieves remarkable accuracies, showing the potential of GOP as signal features for classification or other machine-learning based analysis.