Raw EEG as a viable alternative to engineered decompositions in anesthetic depth prediction

Assessing the depth of anesthesia remains a challenge in operating rooms worldwide, as hospitals often rely on proprietary monitors that are costly and inaccessible to low-resource institutions. This research explores whether machine learning can predict indicators of anesthetic depth from intraoperative EEG, and whether established preprocessing methods significantly improve performance.

Using EEG recordings from 143 patients receiving sevoflurane, we developed a deep neural framework that integrates convolutional, attention-based, and recurrent components. We hypothesized that empirical mode decomposition (EEMD) of EEG, a signal-processing approach in neurophysiology, would improve prediction of the bispectral index (BIS), a clinical measure of consciousness. Contrary to expectation, models trained on raw EEG achieved similar prediction error to EEMD-based pipelines.

These findings suggest that EEG signal decomposition may discard important temporal-spectral features that neural architectures can learn directly from raw data. Moreover, the raw-signal pipeline is computationally lighter, making it better suited for real-time deployment on standard hospital hardware.

These results challenge assumptions about the role of preprocessing in clinical neurophysiology and point toward more accessible and computationally efficient approaches to anesthesia monitoring.