Publication

Abstract : Depth of anaesthesia (DOA) monitoring is very crucial during surgical procedures to prevent complications like intraoperative awareness or excessive anaesthetic drug administration. The Bispectral Index (BIS), extracted from the electroencephalogram (EEG) signals, is a widely used indicator for monitoring the level of consciousness during general anaesthesia. However, a reliable DOA prediction remains challenging due to nonlinear drug dynamics and inter-patient variability. This study presents a comparative framework that evaluates classical pharmacokinetic–pharmacodynamic (PK–PD) models alongside data-driven machine learning (ML) models for BIS prediction. A real-world clinical dataset is analyzed using strict patientwise partitioning, for training and testing, ensuring subject independent validation. BIS values are predicted at multiple horizons of 1 second, 30 second, 60 second, and 90 second ahead using Nonlinear Autoregressive model with eXogenous inputs (NARX) of Random Forest (RF) and Long Short-Term Memory (LSTM) networks to examine short- and medium-term forecasting performance. Model performance is assessed using various performance indices, along with phase-wise evaluation across induction, maintenance, and recovery stages of anaesthesia. The results show that ML models, particularly LSTM networks, achieve lower prediction errors and a high correlation with measured BIS values. The findings demonstrate that ML models effectively capture temporal dynamics and inter-patient variability, highlighting their potential for enhanced anaesthesia monitoring and future closed-loop anaesthesia control systems.