Machine Learning for EEG-Based Depth of Anesthesia Assessment
Accurately measuring the depth of anesthesia is vital to its success, given the negative impacts of both overly light and overly deep anesthesia on patient outcomes; however, it remains a significant clinical challenge.1 Traditional assessment methods based on physiological and behavioral indicators (such as muscle reflexes, heart rate, and blood pressure) are limited due to substantial inter-individual variability, which can be influenced by age, comorbidities, and pharmacological sensitivity. Neurophysiological techniques, particularly electroencephalography (EEG), are the next level of monitoring.2
EEG directly captures cortical electrical activity and exhibits characteristic changes in frequency, amplitude, and complexity as anesthesia deepens, including a shift from low-amplitude, high-frequency activity during wakefulness to high-amplitude, low-frequency oscillations during unconsciousness.3 However, the complexity of this data is imperfectly translated into current metrics such as the Bispectral Index. Alongside the multifactorial determinants of unconsciousness, the nonlinear and dynamic nature of EEG signals have driven interest in advanced analytical approaches that use machine learning methods to improve the accuracy and precision of depth of anesthesia monitoring techniques.1
EEG measures the brain’s electrical activity, providing a more direct, non-invasive assessment of anesthetic states than traditional methods which rely on anesthesia’s effects on the body. Depth of anesthesia assessment with machine learning involves collecting raw EEG data, denoising the signals, extracting relevant features, and applying supervised machine learning to associate EEG changes with anesthetic states.
Variability in feature extraction and index design has led to multiple commercial monitoring systems (such as SedLine, BIS Vista, Narcotrend, NeuroSENSE, and qCON 2000), which each produce distinct indices that generally correlate with alertness and can reduce anesthetic use and recovery time. Although the BIS monitor remains the most widely used device in clinical practice,4 its limitations (such as age and drug dependence, time delay, and susceptibility to noise) highlight the ongoing need for innovation in EEG-based algorithm development.
Traditional machine learning approaches used in depth of anesthesia research include regression models (linear, logistic, and Gaussian processes), support vector machines (SVM), decision trees, and k-nearest neighbors (KNN). Linear regression offers low computational cost and potential suitability for real-time monitoring,5 though its predictive performance is often surpassed by more complex models such as SVM and Gaussian process regression.
Several studies report strong correlations between these advanced models and the Bispectral Index (BIS), with Gaussian process regression and SVM demonstrating high accuracy, particularly when combined with sophisticated feature extraction techniques such as spectral graph wavelet transforms or empirical wavelet transformation.1 Decision trees and random forests have also shown promising classification accuracy, with ensemble methods helping to reduce overfitting and improve robustness across datasets. Meanwhile, KNN algorithms have achieved high accuracy in small cohorts, although there are still some concerns regarding their generalizability to larger and more diverse populations.6
More advanced approaches incorporating high-density EEG and functional connectivity measures further enhance the performance of machine learning models of depth of anesthesia but introduce greater model complexity and interpretability challenges. Studies using graph theory metrics, directed coherence, and phase-based connectivity features have reported high accuracy and strong predictive probabilities. However, many findings are derived from small sample sizes, raising concerns about overfitting and limited external validity.1
EEG-based monitoring combined with machine learning provides a more accurate real-time depth of anesthesia assessment compared to traditional methods. Although advanced models like SVM, Gaussian process regression, and ensemble techniques show high predictive performance, current research does not demonstrate a strong ability to generalize models. Continued development and validation of these methods is necessary and may lead to advancements in anesthesia management.
References
1. Schmierer T, Li T, Li Y. Harnessing Machine Learning for EEG Signal Analysis: Innovations in Depth of Anaesthesia Assessment. Artificial intelligence in medicine. 2024;151:102869-102869. https://doi.org/10.1016/j.artmed.2024.102869
2. Lee KH, Egan TD, Johnson KB. Raw and Processed Electroencephalography in Modern Anesthesia Practice: A Brief Primer on Select Clinical Applications. Korean Journal of Anesthesiology. 2021;74(6):465-477. https://doi.org/10.4097/kja.21349
3. Francisco, Maria A, Nardiello I, Brandão J, Osborn IP. Electroencephalogram Monitoring in Anesthesia Practice. 2021;11(3):169-180. https://doi.org/10.1007/s40140-021-00461-6
4. Zhan J, Yi TT, Wu ZX, et al. A Survey of Current Practices, Attitudes and Demands of Anaesthesiologists Regarding the Depth of Anaesthesia Monitoring in China. BMC Anesthesiology. 2021;21:294. https://doi.org/10.1186/s12871-021-01510-7
5. Li T, Sivakumar P, Tao X. Anesthesia Assessment Based on ICA Permutation Entropy Analysis of Two-Channel EEG Signals. Lecture Notes in Computer Science. 2019:244-253. https://doi.org/10.1007/978-3-030-37078-7_24
6. Nguyen-Ky T, Hoang Duong Tuan, Savkin AV, Do MN, Thu T. Real-Time EEG Signal Classification for Monitoring and Predicting the Transition Between Different Anaesthetic States. IEEE Transactions on Biomedical Engineering. 2021;68(5):1450-1458. https://doi.org/10.1109/tbme.2021.3053019