Machine learning for predicting and understanding oxygen toxicity using EEG brain recordings of divers
Abstract
A diver may experience oxygen toxicity dependent on ambient pressure and the oxygen fraction in the gas mixture breathed. There is an almost inevitable risk of toxicity at the operationally relevant depths at which navy divers breathe oxygen for attack or reconnaissance missions using rebreathers. Manifestations of oxygen toxicity include tunnel vision, tinnitus, nausea, lip twitching, irritability, dizziness, or even more serious symptoms such as convulsions or loss of consciousness. Mild symptoms do not always precede the seizure and hence cannot be used as an early warning system. Unconsciousness while underwater can cause loss of the underwater breathing apparatus mouthpiece and hence the diver has a high chance of drowning. Therefore, studying oxygen toxicity is of great importance for operational military divers undertaking hazardous missions (e.g., conducting attack/reconnaissance swims using oxygen rebreathers). The group at Duke University has conducted a hyperbaric oxygen toxicity study investigating the onset time of oxygen toxicity. Fifty participants were exposed twice to breathing pure oxygen at 2.06 Atmospheric Absolute (ATA) inside the hyperbaric chamber, while exercising in a pool. As part of the extensive participant#s physiology monitoring, an electroencephalogram (EEG) was recorded. Some subjects developed oxygen toxicity symptoms, but preliminary analyses have not provided consistent associations between EEG patterns and oxygen toxicity symptoms. Machine learning and quantitative analysis of the EEG (qEEG) could provide for more information about the subtle neurophysiological changes in the EEG signal accompanying oxygen toxicity. We have four goals in the analysis of the oxygen-toxicity EEG recordings made by the Duke group: (1) to develop automated preprocessing methods to minimize the artifacts found in all EEG recordings (e.g., movement artifacts, which are not related to brain physiology); (2) to conduct qEEG analysis of these data, specifically, the computation of temporal complexity, and understand the relationship between complexity and neurophysiological changes associated with toxicity; (3) to train a convolutional neural network (CNN) (a form of machine learning algorithm) to predict the onset of oxygen toxicity; (4) to analyze the spatial and temporal computations instantiated by the trained CNN with the aim of understanding the neurophysiological changes which accompany oxygen toxicity. In the future, the algorithms developed in this effort could be used for real-time oxygen seizure prediction in divers, and the techniques used in understanding the developed algorithm could improve understanding of the neurophysiological changes which accompany oxygen toxicity. Approved for Public Release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2023
- Source ID
- N000142312467
Entities
People
- Xavier Vrijdag
Organizations
- Office of Naval Research
- United States Navy
- University of Auckland