Assessment of the depth of anesthesia with hidden Markov model based on cardiopulmonary variables

Abstract

BackgroundPrecise monitoring of the Depth of Anesthesia (DoA) is essential to prevent intra-operative awareness (in case of underdosage) or increased post-operative morbi-mortality (in case of overdosage). The recording of a high- frequency multimodal monitoring during general anesthesia (GA) and the capability of classification of dynamic networks should have the potential to help predicting the DoA in a clinical practice. In this study, we aimed at predicting the DoA according four levels (Awake, Loss of Consciousness (LOC), Anesthesia, Return of Consciousness (ROC), Emergence) thanks to a Hidden Markov Model (HMM) relying on four common physiologic variables: Mean Blood Pressure (MBP), Heart Rate (HR), Respiratory Rate (RR), and end-expiratory concentration of sevoflurane (AAEt).MethodsAfter induction by sufentanil and propofol, the anesthesia was maintained by sevoflurane. We recorded the physiological variables at a high frequency during all the procedure [cardiopulmonary variables, AAEt, 2- channel ElectroEncephaloGraphy (EEG) data, and BIS values]. In the training phase, the different states (Awake, LOC, Anesthesia, ROC, Emergence) were identified according to the reading of the spectrograms of the two EEG channels. However, the prediction with the HMM were only based on the four physiological variables.ResultsOn a dataset consisting of 60 patients under general anaesthesia, results suggested that the HMM had a true positive rate (TPR) for identifying Awake, Anesthesia and Emergence of 88%, 72% and 58%, respectively.ConclusionTo our knowledge, this is the first application of such a model to identify the DoA without relying on EEG data. We suggest that a HMM can help the anesthetist monitoring the DoA out of a set of current physiologic variables without necessity of brain monitoring. The model could be improved by increasing the number of patients in the database and accuracy would probably benefit from adding in the model the data of a single EEG channel.