Exploring the Possibility of Measuring Arousal and Awareness in Altered States of Consciousness Using Interpretable Deep Learning
Quantifying arousal in altered states of awareness using interpretable deep-learning
In order to increase the accuracy of statistical estimations of these indices, the classical neurophysiological method for calculating PCI and power spectral densities, as well as spectral exponent, relies on a large number of epochs21. These methods can only be used to investigate averaged brain states, and only provide general information about neurophysiology. Machine learning (ML), even in a real-time trial, can decode and identify specific brain states as well as discriminate them from unrelated signals. This could potentially turn statistical results from a group into individual predictions9. Deep neural networks, a popular approach to ML, have been used to classify and predict brain states based on EEG data23. Convolutional neural networks (CNNs) are the most widely used deep learning technique and have proven effective at classifying EEG data24. A CNN does not provide any information as to why it made the prediction it did25. Recent layer-wise relevance propagation has been successful in demonstrating why classifiers like CNNs made a particular decision26. The relevance score that results from LRP is a measure of the contribution each input variable made to the decision. A high score for a certain area of an input variables indicates that the classifier made the classification or predicted using this feature. Neurophysiological data, for example, suggest that the left-hand motor region is active during right-hand imagery27. The LRP shows that the neural network classes EEG data as right hand motor imagery due to the activity in the left motor area28. The relevance score was therefore higher in the left-motor region than other regions. It is therefore possible to interpret neurophysiological phenomena that underlie the decisions made by CNNs with LRP.
In this work, we develop a metric, called the explainable consciousness indicator (ECI), to simultaneously quantify the two components of consciousness–arousal and awareness–using CNN. CNN was fed the time-series EEG processed data. ECI, unlike PCI which relies upon source modeling and statistical permutation analysis, used event-related potentia at the sensor-level for spatiotemporal dynamic and ML methods. For a generalized model, we used the leave-one-participant-out (LOPO) approach for transfer learning, which is a type of ML that transfers information to a new participant not included in the training phase24,27. The indicator proposed is a value of 2D consisting of indicators for arousal and awareness. We first used TMS/EEG data from healthy participants collected during NREM with no subjective experiences, REM with subjective experiences, and healthy awakeness to consider each component (i.e. low/high arousal, low/high consciousness) in order to analyze correlations with the proposed ECI. We then measured ECI by using TMSEEG data obtained under general anesthesia, with ketamine propofol and xenon. Again, this was done to determine correlations with the three anesthetics. TMS-EEG was also recorded before anesthesia during healthy wakefulness. Healthy participants who woke up reported conscious experiences during ketamine anesthesia, but no conscious experiences during propofol and xenon anesthesia. TMS-EEG was collected from patients who have disorders of consciousness, which include patients diagnosed with UWS or MCS. Our hypothesis was that our ECI could clearly distinguish between the two component of consciousness in physiological, pharmacological and pathological conditions.
We compared ECIawa to PCI, a reliable consciousness index, in order to verify the indicator. We then applied ECI to resting-state EEG acquired from anesthetized patients and participants with DoC. We hypothesize, if CNN is able to learn consciousness-related characteristics, it can calculate ECI accurately without TMS within the proposed framework. It is important, in terms of clinical application, to use the classifiers from previous LOPO-training of the old datasets to classify new data without additional training. We computed ECI for patients with DoC by using the hold-out method29. This is where the training and evaluation data were arbitrarily split, rather than cross-validation. We also investigated the reasons why this classifier made these decisions when using LRP as a way to interpret ECI30.