Reliable assessment of depth of anesthesia (DoA) remains a central challenge in both human and veterinary medicine, particularly in non-verbal subjects where behavioral feedback is unavailable. In this talk, I will present an integrated, data-driven framework for real-time DoA estimation in pigs undergoing propofol anesthesia, combining advanced EEG signal processing, pharmacological modeling, and deep learning.

First, I introduce GAM-GUI, an open-access, MATLAB-based toolbox designed for real-time and offline EEG acquisition and analysis. The platform interfaces directly with a BIOPAC MP160 system and integrates classical signal processing methods with pharmacokinetic/pharmacodynamic (PK/PD) models of propofol, enabling synchronized interpretation of brain activity, drug concentration, and anesthetic effect. At the core of the system lies a Long Short-Term Memory (LSTM) deep learning model trained to predict DoA continuously from EEG signals. The integration of this model into GAM-GUI allows immediate clinical validation and practical deployment, and the tool is currently used in veterinary settings.

Second, I will discuss a systematic study on feature selection for EEG-based DoA prediction, comparing multiple machine learning approaches grounded in variance-based principles, including Pearson and Spearman correlations, Principal Component Analysis (PCA), and ReliefF. Despite their methodological differences, these approaches converge toward a consistent set of informative EEG features. In particular, spectral power, power density ratios, and entropy measures in the gamma frequency range emerge as the most robust predictors of anesthetic depth.

Together, these results demonstrate how explainable feature selection and deep learning can be combined within an open, clinically oriented software framework, offering a transparent and flexible alternative to proprietary anesthesia monitoring systems and paving the way toward broader closed-loop anesthesia applications.