Automated machine learning classification of electrodermal activity for prediction and detection of symptoms related to the central nervous system oxygen toxicity including seizures

Background: Prolonged and high pressure (deep) diving leads to various physiological challenges including the risk for decompression sickness (DCS) and death [1][7]. DCS is a significant concern for individuals performing tasks after which a decrease in ambient perbaric oxygen (HBO2) prebreathing [10][12]. Breathing HBO2 decreases leukocyte adhesion and platelet aggregation [13], [14]er, divers breathing HBO2 are at risk for developing central nervous system oxygen toxicity (CNSOT), which can manifest as symptoms that might impair a divers performance, such as headache, nausea, tinnitus, lip twitching, tingling of the limbs, or even more serious symptoms such as seizure or altered consciousness [15], [16].Previous work and preliminary data: We have recently collected a tasks at increased oxygen partial pressures following a ketogenic and another HBO2 after consumption of a normal diet in a hyperbarironment to detect and predict symptoms related to CNS-OT, including the onset of seizures, prior to a human expert noticing the symptoms. Some of the results on clean EDA data showed an exciting potential to predict symptoms related to oxygen toxicity, as we observed a several-fold increase in EDA values prior to human expert adjudication that CNS-OT had begun. While we have obtained some clean EDA data, many of the subjects data were not usable due to either low signal fidelity or technical issues at the onset of our study.Proposed Aims: The first objective of the study is to develop machine and deep learning for automated detection of clean EDA data segments even when confronted with motion-corrupted signals. This first objective will be in collaboration with Dr. Bruce Derrick at Duke University. We will leverage his ongoing study on the efficacy of a ketogenic diet in suppressing CNS-OT in human subjects. As EDA data collection is ongoing at his lab, we will have an ample supply of data to analyze and train for motion artifact detection. The second objective is to evaluate various machine learning methods for automated classification of whether or not a given rise in TVSymp value is due to CNS-OT. We will compare machine learning results to an arbitrarily-set TVSymp threshold value to examine whether or not a spike is truly due to CNS-OT. The second objective will also be in collaboration with Dr. Bruce Derrick at Duke University and Dr. Dawn Kernagis at the University of North Carolina. The third objective is to build a wearable device to collect EDA and accelerometer data, and to wirelessly transmit the data collected to either a smartphone or divers heads-up display to provide an advanced CNS-OT warning system. We have developed a prototype that is able to collect EDA and accelerometer data, among other signals, so we intend to finish the construction and validation of such device for the deployment of the envisioned system to predict CNS-OT.Impact: If EDA is sensitive to pre-seizure and seizure conditions with enough warning time to take action, the proposed approach has the potential to save divers lives when the EDA sensor is used in conjunction with O2 prebreathing and ketogenic nutrition. If this work is successful in predicting onset of symptoms related to CNS-OT including seizures using EDA sensors, the development of wearable systems for future deployment will be much simpler to implement, better accepted by divers, and more cost-effective than wearing impractical electroencephalographic (EEG) sensors. Importantly, we have already demonstrated the functionality of our waterproof EDA sensors in immersed conditions whereas such underwater capability has not yet been developed for EEG sensors.