Prediction model of response to standardized painful stimulation during general anesthesia: Applying machine learning to the EEG signal

Abstract

Adequate analgesic and antinociception levels are necessary to avoid adverse events such as movement or hemodynamic reactions. Although different monitoring systems estimating antinociception effect during general anesthesia are available, the performance of these indices is limited. The aim of this study was to develop a predictive model of the probability of movement if a standardized tetanic stimulation is applied to the patient, using the electroencephalogram (EEG) signal as input. Seventy-seven female patients, scheduled for ambulatory gynecologic surgical procedures under general anesthesia, were included. EEG segments of 45s, preceding tetanic stimulation, were split into epochs of 5s with no overlapping. Features from both time and frequency domain, including continuous wavelet transform (CWT), were considered. Five different machine learning (ML) models were tested: 1D CNN model for raw EEG signal, random forest classifier, XGBoost classifier and support vector classifier for extracted features and finally 2D CNN model for CWT scalogram. The results showed that the performance of the proposed ML models to predict movement to standardized tetanic stimulation through the EEG signal was almost equal to the performance of commercially available indices (AUC=0.73). This performance was improved (AUC=0.84) with ML methods in ensemble mode with pupil diameter and remifentanil effect site concentration. Indeed, prediction of movement to noxious stimulation through the EEG signal matches state of the art and could potentially be used in combination of other variables.