Prediction of Burst Suppression Occurrence Under General Anaesthesia Using Pre-operative EEG Signals.

Intraoperative burst suppression (BS) is associated with postoperative neurocognitive disorders, which could lead to increased mortality, morbidity and longer hospitalisation. The main objective of this study is to build a machine learning model capable of predicting the BS pattern between the phases of induction and incision, based on EEG and patient characteristics acquired preoperatively. To this end, several models, namely decision trees, random forest, XGBoost, SVM and logistic regression, were trained on the data with varying window sizes. The performance of the trained models is evaluated using stratified 5-fold cross-validation and a 5 × 2 cv t-test was applied to test significance. The main objective is to study, among other factors, the potential impact of the anaesthetic dosage on BS occurrence with the aim of personalising this process. The results obtained indicate that a Logistic Regression approach trained on the data epoched into 7-s windows can achieve a precision score of 0.63 and an ROC-AUC score of 0.61 while predicting future BS occurences. Furthermore, the explanations of the feature importance obtained from the SHapley Additive exPlanations (SHAP) demonstrate that the mean absolute power delta and alpha bands contribute the most to the predictions made as well as the dosage of the anaesthetic agent propofol.