An investigational study of electroencephalography variability in emergency care physicians: a machine learning approach (Preprint)

BACKGROUND Emergency Department physicians are continuously exposed to traumatic events during their work and are at high risk of developing neurological disorders such as burnout and Acute Stress Disorder (ASD). The clinical evaluation process for diagnosing these disorders is based on history and physical examination. Screening and questionnaires are the most common diagnostic tools that require significant training and often miss initial physiological symptoms. The addition of Artificial Intelligence (AI) data-driven decision support using passive EEG monitoring of biopotential brainwave activity may be the key to unlocking early signs of trauma-related neurological disorders for healthcare professionals exposed to high levels of work-related stress. OBJECTIVE The objective of this research is to test the design of a minimally invasive mobile electroencephalography (EEG) biomonitoring device with emergency physicians (EPs), analyze biopotential brainwave activity collected from EPs while experiencing stressful events, and develop an AI-driven decision support system prototype to analyze near-real-time neurological responses and monitor biomarkers associated with stressful events. METHODS Twenty-three emergency department physicians were monitored while on shift at the Greenville Memorial Hospital Emergency Department, a Level 1 Trauma Center. Participants continuously wore the study's EEG headset both before their shift, to establish a reference baseline signal, and throughout their shift when exposed to stressful events. Biopotential data were transferred to a secure university research server, and stress levels were determined using personalized statistical thresholds. Machine learning models were created to predict stressful events, and waveform analysis revealed key predictors. RESULTS Neurological response data show increased frequency domain waveform amplitudes as participants are exposed to traumatic events. Machine learning models (ANN, SVM, NB, LR, DT) were trained using 70% of the data and tested with 30%. ANN achieved the highest prediction accuracy at 92% (64,626/70,246). All models demonstrated an impressive ability – greater than 81% accuracy – to use neurological responses to predict stressful events. In each model, Gamma waveforms were over 80% important for stress prediction, highlighting the importance of high-frequency brainwaves in stress detection. CONCLUSIONS Neurological response data show that stressful events can be parameterized by the instantaneous and progressive outputs of the monitored spectral waveforms as they relate to ER physicians exposed to traumatic environments. Comparing EEG waveforms from multiple participants highlights the need for personalized baselines in stress prediction. Analysis of machine learning model performance reveals that non-linear models, such as ANN and DT, are better suited for biopotential signal analysis due to their ability to account for complex relationships with the data. The investigation into the frequency domain of EEG spectra revealed that higher-frequency bands, such as Gamma and Beta, are important for predicting stress. In contrast, lower-frequency bands like Theta and Delta are less influential. CLINICALTRIAL N/A