Enhanced Depth of Anesthesia Classification Using Graph Neural Networks and EEG Features

Abstract

During surgical procedures, precise monitoring of the depth of anesthesia (DoA) is crucial for patient safety; yet, conventional techniques, such as the Bispectral Index (BIS), are expensive and opaque. Based on EEG signals from a dataset of 24 patient cases, this study suggests a unique method for classifying DoA into four states: general anesthesia, profound anesthesia/burst suppression, moderate sedation, and awake/light sedation. It does this by employing a Graph Neural Network (GNN). Before extracting features including spectral entropy, Hjorth parameters, and peak- specific metrics, the methodology preprocesses EEG data using bandpass filtering (1– 50 Hz), DC offset removal, downsampling to 128 Hz, normalization, and smoothing. To represent feature interdependencies, a correlation-based graph was built. It outperformed more conventional models such as ANN (78.83%) and ResNet50 (62.75%) with a classification accuracy of 91.46%. With the potential to improve patient outcomes and accessibility in environments with limited resources, the open- source GNN paradigm provides a reasonably priced and interpretable substitute for proprietary solutions. Subsequent research will concentrate on verifying the model using bigger datasets and investigating the viability of real-time deployment.