1013 Decoding Delirium: Machine Learning Classification From Inpatient Wearable EEGs

INTRODUCTION: Delirium is an acute neuropsychiatric disorder of attention and awareness that significantly impacts hospitalized patients, particularly older patients, and is associated with increased dependence and death. Despite its prevalence, delirium detection remains challenging due to its fluctuating nature and the limitations of current clinical diagnostic tools. METHODS: We conducted a prospective observational study at a tertiary care academic medical center (n = 115) to develop and validate EEG-based machine learning models for delirium detection. Recordings were collected from a wearable EEG device. We extracted 2196 EEG features from single-channel EEG recordings. Skewed feature data was log-transformed. We conducted feature selection using chi-squared univariate, lasso regression, and minimum redundancy maximum relevancy (mrmr) approaches. We then fit four different supervised ML models (SVM, logistic regression, gradient boosting, and random forest) to predict delirium occurrence using fivefold cross-validation. RESULTS: The model with the highest area under the receiver operating characteristic (AUROC) curve was an SVM model with 10 mrmr-selected features, demonstrating promising discriminative ability with an AUROC of 0.73, sensitivity of 65%, and specificity of 70.8%. Key features contributing to model performance included theta and delta band power metrics. The model performance surpassed the upper bound of a permutation test's null distribution, confirming its statistical significance over chance. CONCLUSIONS: Our findings demonstrate that wearable EEG combined with automated feature extraction and machine learning modelling can effectively detect delirium. Ongoing work aims to refine feature selection, extend analysis to larger datasets, and develop models which use continuous (rather than categorical) variables to predict delirium severity in hospitalized patients.