Introduction: Delirium is an acute neuropsychiatric disorder of attention and awareness that significantly impacts hospitalized patients and is associated with increased dependence and death. Despite high prevalence, delirium detection remains challenging due to limitations of current clinical tools. In this work, we developed a novel wireless wearable EEG device and deployed it among hospitalized inpatients. Using collected EEG data, we built machine learning (ML) pipelines to predict validated delirium.
Methods: We conducted a prospective observational study (n = 155) to develop and validate EEG-based machine learning models for delirium detection. Recordings were collected from a novel wireless wearable EEG device we developed. We extracted 2196 EEG features from single-channel recordings. Feature selection was conducted using LASSO regression, f_classifier, minimum-redundancy-maximum-relevancy (mRMR), and other approaches. We then fit four different supervised ML models (XGBoost, random forest, logistic regression, support vector machines) to predict delirium occurrence using fivefold cross-validation (80:20 split). Feature generation was conducted in MATLAB R2022b and ML analysis was conducted with scikit-learn in Python 3.11.5. We further assessed device comfort, tolerability, and impact on sleep from patient surveys.
Results: The XGBoost model with all selected features demonstrated promising discriminative ability in detecting delirium with an area under the receiver operator characteristic (AUROC) of 0.81, sensitivity of 70%, and specificity of 75%. Shapley analysis of feature importance found that predictions are based on a complex use of many features including statistical metrics, fractal dimensions, power spectral density for specific EEG frequency bands, features related to de-trending signals, and measures extracted from EEG analysis packages including FOOOF, Catch-22, and NEURAL. Based on patient surveys, our device demonstrated high device comfort, high willingness to wear again, and minimal impact on sleep.
Conclusion : Our wireless wearable EEG device with automated feature extraction and ML modeling can effectively detect delirium from EEG data alone. The ability to track neural activity over time can help develop preventative strategies and targeted treatments. We further conclude that our device is comfortable, tolerable, and minimally impacts sleep.
.jpg)
