An Intraoperative EEG Biomarker for Postoperative Delirium Predicting Based on Interpretable Deep Learning Framework

Abstract

Postoperative delirium (POD) is a common complication in older surgical patients and substantially worsens clinical outcomes, yet existing intraoperative electroencephalography (EEG) monitoring tools lack spatial and temporal specificity, creating a need for interpretable biomarkers. We prospectively analyzed 32-channel intraoperative EEG from 71 patients aged ≥ 60 undergoing noncardiac surgery, trained an interpretable spatiotemporal convolutional network (ST-CN), derived a best temporal filter (BTF), and evaluated model performance with region-specific tests and independent external validation. The ST-CN classified POD with 97.52% accuracy and an ROC of 0.996. The BTF alone discriminated POD with an AUC of 0.911 and achieved 85.12% accuracy using frontal EEG alone. It captured a distinct 2-12 Hz (δ-θ-α) oscillation in a spindle-like envelope, which occurred at a significantly higher rate in POD patients (4.62 ± 0.15 vs. 3.88 ± 0.13 waves/min in the frontal region) and differed in central frequency and spectral power. Independent external validation further confirmed robust generalizability, with frontal EEG achieving 89.01% accuracy and an AUC of 0.950. This framework enables accurate, interpretable POD risk stratification and identifies a reproducible frontal EEG biomarker, supporting objective intraoperative early warning and individualized perioperative care.