Quantitative electroencephalogram and machine learning to predict expired sevoflurane concentration in infants

Abstract

Processed electroencephalography (EEG) indices used to guide anesthetic dosing in adults are not validated in young infants. Raw EEG can be processed mathematically, yielding quantitative EEG parameters (qEEG). We hypothesized that machine learning combined with qEEG can accurately classify expired sevoflurane concentrations in young infants. Knowledge from this may contribute to development of future infant-specific EEG algorithms. Frontal EEG collected from infants ≤ 3 months were time-matched as one-minute epochs to expired sevoflurane (eSevo). Fifteen qEEG parameters were extracted from each epoch and eight machine learning models combined the qEEG to classify each epoch into one of four eSevo levels (%): 0.1-1.0, 1.0-2.1, 2.1-2.9, and > 2.9. 64 epochs formed the post hoc SHAP dataset to determine the qEEG that contributed most to the model. The remaining epochs were randomly split 50 times into 80/20 training/testing sets. Accuracy and F1-score determined model performance. 42 infants provided 4574 epochs. The top classifiers K-nearest neighbors, default multi-layer perceptron, and support vector machine achieved 67.5-68.7% accuracy. Burst suppression ratio and entropy β were the top contributors to the models. Post hoc analysis performed without burst suppression ratio yielded similar prediction performance. In young infants, machine learning applied to qEEG predicted eSevo levels with moderate success. Burst suppression ratio, the most important contributor, represented an efficient EEG feature that encapsulated underlying EEG changes seen on other qEEG features. These results provided insight into EEG parameter selection and optimal machine learning models used for future development of infant-specific EEG algorithms.

Keywords: Infant electroencephalogram EEG; Machine learning; Pediatric EEG anesthesia; Quantitative EEG anesthesia; SHAP EEG analysis.

Fig. 1

Eligibility flowchart

Fig. 2

A Histogram of the distribution of analyzed expired sevoflurane concentrations. B Histogram of the number of epochs contributed by each subject

Fig. 3

A Plot of the first two principal components of the quantitative EEG parameters colored by different infants, with little separation seen between the epochs from different infants. Each dot of the same color represents an epoch from the same infant. B Plot of the first two principal components of the quantitative EEG parameters colored by different sites, with little separation seen by site. Each dot of the same color represents an epoch from the same study site

Fig. 4

Boxplots of model overall accuracy (A) and F1-score (B) on holdout sequences over 50 different iterations. Center horizontal line represents median; top and bottom of box are 1st and 3rd quartile, respectively; top and bottom whiskers are 95th and 5th percentile, respectively; points outside of whiskers are outliers. LR logistic regression; DT decision tree, SVM support vector machine, KNN K-nearest neighbors, GNB Gaussian Naïve Bayes, DMLP default multi-layer perceptron (with default Adam optimizer), MLP multi-layer perceptron (with stochastic gradient descent optimizer), ADA AdaBoost of a decision tree

Fig. 5

Plots of Shapley values for the three best-performing models (top = KNN, middle = DMLP, bottom = SVM). The four colors represent contribution to correctly classifying to one of the four levels of expired sevoflurane

Fig. 6

Ranked histogram of proportion (%) of each EEG features’ contribution to the model (top = KNN, middle = DMLP, bottom = SVM)