Deep learning based depth of anaesthesia monitoring using EEG: a 4-layer CNN model with PSD and BSR correlation features.

Abstract

Monitoring the depth of anaesthesia (DoA) through electroencephalogram (EEG) analysis plays a major role in maintaining patient safety and guiding optimal anaesthetic delivery during surgery. In this study, a convolutional neural network (CNN) framework was designed to classify DoA into three clinically relevant states: Deep Anaesthesia, adequate anaesthesia, and awake. EEG recordings were transformed into spectrograms, enabling the CNN to capture temporal-spectral representations. Key features such as power spectral density (PSD) and the correlation between PSD and burst suppression ratio (BSR) were incorporated to enhance discriminability across states. The model was evaluated on three representative patient cases, achieving accuracies of 93%, 84%, and 73%, respectively. Detailed performance analysis using confusion matrices and metrics such as Precision, Recall, and F1-score demonstrated strong generalization in detecting stable states (Deep Anaesthesia and awake) but limited sensitivity in transitional states (adequate/light anaesthesia) due to overlapping spectral signatures. These findings highlight both the promise and limitations of CNN-based DoA monitoring.