Multiscale depth of anaesthesia prediction for surgery using frontal cortex electroencephalography

Abstract

Hypnotic and sedative anaesthetic agents are employed during multiple medical interventions to prevent patient awareness. Careful titration of agent dosing is required to avoid negative side effects; the accuracy thereof may be improved by Depth of Anaesthesia Monitoring. This work investigates the potential of a patient specific depth monitoring prediction using electroencephalography recorded neural oscillation from the frontal lobe of 10 patients during sedation, where a comparison of the prediction accuracy was made across five different approaches to post-processing; Noise Assisted-Empirical Mode Decomposition, the Raw Signal, Linear Series Decomposition Learner, Deep Wavelet Scattering and Deep Learning features. These methods towards anaesthesia depth prediction were investigated using the Bispectral Index as ground truth, where it was seen that the Raw Signal, enhanced feature set and a low complexity classification model (Linear Discriminant Analysis) provided the best classification accuracy, in the region of 85.65 % ±10.23 % across the 10 subjects. Subsequent work in this area would now build on these results and validate the best performing methods on a wider cohort of patients, investigate means of continuous DoA estimation using regressions, and also feature optimisation exercises in order to further streamline and reduce the computation complexity of the designed model.