Prediction of Anesthesia Depth and Early Detection of State Transitions Using EEG-Derived Features

EEG-based depth of anesthesia (DoA) instrumentation facilitates the measurement of consciousness levels during surgical procedures; critical for patient safety and optimal anesthetic delivery. Existing instrumentation systems are limited by providing only retrospective DoA measurement and lack predictive measurement tools or early warning functions. This constrains their value during dynamic intraoperative events such as loss of consciousness (LoC) and recovery of consciousness (RoC), where timely adjustment of anesthetic delivery is crucial. To address this need, this study presents a prediction and early warning measurement system, named DoA-PrEWS. This framework integrates EEG-derived features with machine learning for DoA measurement at current and future time intervals. This framework extends to provide specialized measurement around the detection and prediction of transitional events. The framework comprises five components: DoA-Current-GPR, a Gaussian process regression model for measuring current DoA levels; DoA-State-BEM, a bagged ensemble model for anesthetic state classification; DoA-Transition-RF, a random forest model for classifying DoA state transition moments; DoA-TTT-NN, a neural network for measuring Time To Transition events; and DoA-Future-GPR, a GPR model for DoA prediction. Evaluated on EEG data from 120 surgical patients and benchmarked against the Bispectral Index (BIS), DoA-Current-GPR achieved a correlation of 0.85 (RMSE = 11.22), and DoA-Future-GPR predicted DoA 120 seconds ahead with a correlation of 0.78. DoA-Transition-RF detected loss of consciousness (LoC) and recovery of consciousness (RoC) with accuracies of 95.14% and 83.75%, respectively, while DoA-TTT-NN predicted time to transition with correlations of 0.81 (LoC) and 0.72 (RoC). Performance was consistent across demographic groups, surgical types, anesthetic agents, and under variable signal quality conditions. External validation using an independent UniSQ dataset yielded consistent results, supporting cross-institution generalizability. End-to-end inference latency was approximately 29.7 ms per window, indicating compatibility with real-time clinical deployment. These results demonstrate the feasibility of a unified DoA system for the detection of current and future consciousness levels. This work highlights the clinical value of improving intraoperative measurement to ensure responsiveness in managing anesthetic state transitions.