Generalized Prediction of Unconsciousness during Propofol Anesthesia using 3D Convolutional Neural Networks.

Neuroscience has generated a number of recent advances in the search for the neural correlates of consciousness, but these have yet to find valuable real-world applications. Electroencephalography under anesthesia provides a powerful experimental setup to identify electrophysiological signatures of altered states of consciousness, as well as a testbed for developing systems for automatic diagnosis and prognosis of awareness in clinical settings. In this work, we use deep convolutional neural networks to automatically differentiate sub-anesthetic states and depths of anesthesia, solely from one second of raw EEG signal. Our results with leave-one-participant-out-cross-validation show that behavioral measures, such as the Ramsay score, can be used to learn generalizable neural networks that reliably predict levels of unconsciousness in unseen transitional anesthetic states, as well as in unseen experimental setups and behaviors. Our findings highlight the potential of deep learning to detect progressive changes in anesthetic-induced unconsciousness with higher granularity than behavioral or pharmacological markers. This work has broader significance for identifying generalized patterns of brain activity that index states of consciousness.Clinical Relevance- In the United States alone, over 100,000 people receive general anesthesia every day, from which up to 1% is affected by unintended intraoperative awareness [1]. Despite this, brain-based monitoring of consciousness is not common in the clinic, and has had mixed success [2]. Given this context, our aim is to develop and explore an automated deep learning model that accurately predicts and interprets the depth and quality of anesthesia from the raw EEG signal.