Machine learning prediction of burst suppression under general anesthesia

Abstract

During propofol-remifentanil induced general anesthesia, burst suppression (BS) EEG patterns commonly occur in around 50% of the patients, with an increasing incidence with age. However, this phenomenon has been reported to be an indicator of too high anesthetic doses and produce adverse outcomes such as postoperative delirium, cognitive deficits, and it has even reported to be a postoperative mortality predictor. In light of the above, the present study aims to address the lack of predictive techniques for BS occurrence anticipation by developing Machine Learning predictive models such as SVM, KNN, RF, and XGB. Therefore, a large dataset including different monitored parameters during propofol-remifentanil induced general anesthesia from many patients has been used for both training and testing the models, as well as for final validation of the selected model. Obtained results present an acceptable overall performance of the SVM model with a ROC-AUC score of 0.829, and a feature importance analysis shows a high influence of age and BIS value for the final prediction. Nonetheless, 25% of the predictions have been reported to have accuracies under 0.6, questioning the reliability of the model and making it useful as an orientative aiding tool for anesthesiologists, but not the ultimate decisive factor. Hence, further studies involving more variability on the data, validation techniques and confidence intervals for each process, and an exhaustive feature selection analysis, along with the repetition of the study with different ML algorithms should be performed to improve the predictive ability of the current model and achieve better performances.