Model-Based Approaches for Classification of Levels of Consciousness

Abstract

Objective. Classification of different levels of consciousness is of great importance in the diagnosis of sleep disorders or monitoring during anaesthesia. Our current research comprises mainly of sleep scoring but it is planned to extend the work to classify vigilance states during anaesthesia as well. Sleep scoring can be very time-consuming. Hence, (semi-)automated strategies are advised to facilitate and speed up the procedure. Different methods which achieve a good accuracy have been proposed in research, most of them are based on machine learning algorithms. However, none of them are in widespread use in the clinical or preclinical field so far. For acceptance in the medical sector, a model has to be interpretable and comprehensible, properties that many of these models lack. Our approach aims to create an explainable model for understanding and analysing EEG data. Methods. Due to the very complex behaviour of the brain, a modelling strategy that is based on rules and equations is not possible in neuroscience. But a model description providing a technical understanding can help to describe processing in the brain. Machine learning combined with classical modelling supports the development of a mathematical and computational framework for classification of vigilance states. Given an EEG signal, features are extracted of the signal that should predict the level of consciousness. The focus lies on entropy-based parameters such as Permutation Entropy, Entropy of Difference and Kullback-Leibler Divergence. Granger Causality is also used which estimates the influence of one EEG signal to another one. Apart from that, statistical features are computed as well. Using these and the age of the patient different machine learning models are created, trained and tested and afterwards compared regarding their performance and interpretability. The influence of the features is evaluated as well. The implementation and tests are all run on MATLAB. Results. First results show that linear models, which are much more explainable and easier to interpret, can compete with more complex ones. However, the concept has yet to be refined and many steps to be taken to get a model that achieves good results and has the desired properties. Feature evaluation already gives us a hint which are more important and where the signal analysis can be optimised.