Development of a deep learning-based prediction model for postoperative delirium using intraoperative electroencephalogram in adults

Abstract

Postoperative delirium (POD) is associated with increased morbidity and mortality. This study aims to develop a deep learning-based model (DELPHI-EEG) to predict postoperative delirium using intraoperative electroencephalogram (EEG) waveform. A total of 34,550 surgical cases (267 event cases), with 6-lead intraoperative EEG monitoring between 2022 and 2024, were included for model development. During 5-fold cross-validation, the DELPHI-EEG model showed an area under the receiver operating characteristic (AUROC) curve of 0.870 (95% confidence interval [CI]: 0.789-0.935) and the area under the precision-recall curve (AUPRC) of 0.038 (95% CI: 0.017-0.084), significantly outperforming the logistic regression model using burst suppression ratio with AUROC of 0.729 (95% CI: 0.624-0.825, p = 0.004) and AUPRC of 0.013 (95% CI: 0.007-0.026, p = 0.002). The DELPHI-EEG model might serve as a risk predictor for postoperative delirium, potentially enabling targeted preventive interventions for surgical patients; nonetheless, external validation in diverse clinical settings is required.