Deep Learning for Electrophysiological Investigation and Estimation of Anesthetic-Induced Unconsciousness

Neuroscience has made a number of advances in the search for the neural correlates of consciousness, but our understanding of the neurophysiological markers remains incomplete. In this work, we apply deep learning techniques to resting-state electroencephalographic (EEG) measures of healthy participants under general anesthesia, for the investigation and estimation of altered states of consciousness. Specifically, we focus on states characterized by different levels of unconsciousness and anesthetic depths, based on definitions and metrics from contemporary clinical practice. Our experiments begin by exploring the ability of deep learning to extract relevant electrophysiological features, under a cross-subject decoding task. As there is no state-of-theart model for EEG analysis, we compare two widely used deep learning architectures - convolutional neural networks (cNNs) and multilayer perceptrons (MLPs) - and show that cNNs perform effectively, using only one second of the raw EEG signals. Relying on cNNs, we derive a novel 3D architecture design and a standard preprocessing pipeline, which allows us to exploit the spatio-temporal structure of the EEG, as well as to integrate different acquisition systems and datasets under a common methodology. We then focus on the nature of different predictive tasks, by investigating classification and regression algorithms under a variety of clinical ground-truths, based on behavioral, pharmacological, and psychometrical evidence for consciousness. Our findings provide several insights regarding the interaction across the anesthetic states, the electrophysiological signatures, and the temporal dynamics of the models. We also reveal an optimal training strategy, based on which we can detect progressive changes in levels of unconsciousness, with higher granularity than current clinical methods. Finally, we test the generalizability of our deep learning-based EEG framework, across subjects, experimental designs, and anesthetic agents (propofol, ketamine and xenon). Our results highlight the capacity of our model to acquire appropriate, task-related, cross-study features, and the potential to discover common cross-drug features of unconsciousness. This work has broader significance for discovering generalized electrophysiological markers that index states of consciousness, using a data-driven analysis approach. It also provides a basis for the development of automated, machine-learning driven, non-invasive EEG systems for real-time monitoring of the depth of anesthesia, which can advance patients' comfort and safety.