Predicting nitrogen narcosis using the electroencephalogram recorded from experienced divers

Nitrogen narcosis causes acute cognitive impairment in divers breathing compressed air at depth, increasing injury and fatality risk. Susceptibility varies widely between individuals, making real-time cortical monitoring a potential safety tool. We recorded 32-channel electroencephalography (EEG) from 45 divers (both sexes; aged 20-55 yr) at rest across four randomized crossover hyperbaric studies. We manipulated pressure (101-811 kPa) and breathing gas (air or heliox) to create two classes of recordings: narcotic (high-pressure air) or non-narcotic (low-pressure air; heliox at any pressure); our outcome was therefore gas composition, not direct cognitive impairment. We extracted 2,720 features (17 feature types × 32 channels × 5 frequency subbands) from preprocessed EEG epochs, trained six machine-learning classifiers using leave-one-participant-out cross-validation on a training set (24 participants), and evaluated the best-performing model on a test set (9 participants) and hold-out set (12 participants). Linear discriminant analysis (LDA) achieved a Matthews correlation coefficient (MCC) of 0.50 (chance ≈ 0; P = 0.001) and accuracy of 82% (baseline 76%; P = 0.001) on the test set, and MCC of 0.25 (chance ≈ 0; P = 0.001) and accuracy of 69% (baseline 69%; P = 0.77) on the hold-out set collected years later. Feature elimination retained 157 features; feature importance analysis identified zero-crossing rate, Hjorth mobility, and Katz fractal dimension at delta-band frequencies (< 4 Hz) over the anterior and posterior midline as the most important features. Easily computed EEG features with an LDA classifier can detect neurophysiological signatures of nitrogen narcosis, potentially enabling real-time passive monitoring of divers.NEW & NOTEWORTHY Nitrogen narcosis can impair a diver's cognitive performance in highly variable ways, influenced by multiple factors. To address this, real-time brain monitoring is desirable to inform divers of their cognitive state. By leveraging machine learning to detect subtle patterns in electroencephalography (EEG) signals, we can develop interpretable algorithms that may support a future monitoring and warning system.