Deep Learning-Based Classification of Anesthesia Depth Using EEG and Bispecral Index

Abstract

Electroencephalography (EEG) is a broadly used method designed for measuring the depth of anesthesia (DoA) due to its ability to replicate the brain's state and surgical pain. However, factors such as inaccurate assessment of DoA which can lead to unintended awareness and postoperative complications make precise monitoring difficult. Traditional methods of DoA monitoring are also time consuming, as they require continues attention to patient’s vital sign during surgery.  The objective of this study is to improve the classification of DoA by using deep learning techniques. We trained and evaluated Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), and a combined CNN-LSTM model on a dataset of EEG signals from 30 patients undergoing general anesthesia, alongside corresponding Bispectral Index (BIS) values. The CNN model achieved moderate performance, with its best results in the General anesthesia class with the precision: 65.95%, recall: 55.99%, F1-score: 60.56%. The LSTM model achieved an improved F1-score of 85.13% for the same class. The hybrid CNN–LSTM model produced the best overall performance, achieving 65.78% accuracy, 75.00% specificity, and an F1-score of 79.36%, demonstrating the efficiency of hybrid deep learning for EEG-based DoA classification.