Frontal electroencephalogram based drug, sex, and age independent sedation level prediction using non-linear machine learning algorithms

Abstract

Brain monitors which track quantitative electroencephalogram (EEG) signatures to monitor sedation levels are drug and patient specific. There is a need for robust sedation level monitoring systems to accurately track sedation levels across all drug classes, sex and age groups. Forty-four quantitative features estimated from a pooled dataset of 204 EEG recordings from 66 healthy adult volunteers who received either propofol, dexmedetomidine, or sevoflurane (all with and without remifentanil) were used in a machine learning based automated system to estimate the depth of sedation. Model training and evaluation were performed using leave-one-out cross validation methodology. We trained four machine learning models to predict sedation levels and evaluated the influence of remifentanil, age, and sex on the prediction performance. The area under the receiver-operator characteristic curve (AUC) was used to assess the performance of the prediction model. The ensemble tree with bagging outperformed other machine learning models and predicted sedation levels with an AUC = 0.88 (0.81-0.90). There were significant differences in the prediction probability of the automated systems when trained and tested across different age groups and sex. The performance of the EEG based sedation level prediction system is drug, sex, and age specific. Nonlinear machine-learning models using quantitative EEG features can accurately predict sedation levels. The results obtained in this study may provide a useful reference for developing next generation EEG based sedation level prediction systems using advanced machine learning algorithms. Clinical trial registration: NCT02043938 and NCT03143972.

Keywords: Anaesthesia; Consciousness Monitors; Electroencephalogram; Machine learning; Medical informatics.

Conflict of interest statement

M.M.R.F.S.: His research group/department received (over the last 3 years) research grants and consultancy fees from The Medicines Company (Parsippany, NJ, USA), Masimo (Irvine, CA, USA), Fresenius (Bad Homburg, Germany), Dräger (Lübeck, Germany), Paion (Aachen, Germany), and Medtronic (Dublin, Ireland). He receives royalties on intellectual property from Demed Medical (Temse, Belgium) and the Ghent University (Gent, Belgium). Other authors have no conflicts of interest to declare.

© 2020. The Author(s).

Figures

Fig. 1

(a) Architecture of the proposed sedation level estimator, and (b) Illustration of the EEG epoch selection, segmentation and feature extraction process. One-minute EEG segments preceding the time of MOAA/S assessments were used for the analysis. Each segment was further divided into non-overlapping 4 s short EEG epochs and 44 QEEG features were extracted from each 4 s epoch

Fig. 2

Illustration of the cross-validation strategy used in this study. A 10-fold cross validation using training data was used for model hyperparameters and feature selection and leave-one-subject-out cross validation was used to predict the sedation level for each subject

Fig. 3

The distribution of AUC’s for individual features across all drugs to discriminate between awake and sedated EEG epochs with (propofol, sevoflurane, dexmedetomidine and remifentanil) and without remifentanil (propofol, sevoflurane, dexmedetomidine). The performance of all features significantly dropped after the addition of remifentanil. Here the vertical solid line indicates mean AUC and horizontal bar refers to standard deviation. X-axis corresponds to features: 1–12 = time domain, 13–36 = frequency domain and 37–44 = entropy domain features

Fig. 4

Heatmap illustrating the weights (normalized to 1) assigned by the ensemble tree with bagging algorithm. Different features were selected when remifentanil was added to propofol, sevoflurane, dexmedetomidine. Here dark blue indicates highest weight assigned by the elastic-net regularization algorithm. Fractal dimension had highest weight in both cases