Implementation of Neural Networks to Frontal Electroencephalography for the Identification of the Transition Responsiveness/Unresponsiveness During Induction of General Anesthesia

ObjectiveGeneral anesthesia is a reversible drug-induced state of altered arousal characterized by loss of responsiveness (LOR) due to brainstem inactivation. Precise identification of the LOR during the induction of general anesthesia is extremely important to provide personalized information on anesthetic requirements and could help maintain an adequate level of anesthesia throughout surgery, ensuring safe and effective care and balancing the avoidance of intraoperative awareness and overdose. So, main objective of this paper was to investigate whether a Convolutional Neural Network (CNN) applied to bilateral frontal electroencephalography (EEG) dataset recorded from patients during opioid-propofol anesthetic procedures identified the exact moment of LOR.Material and methodsA clinical protocol was designed to allow for the characterization of different clinical endpoints throughout the transition to unresponsiveness. Fifty (50) patients were enrolled in the study and data from all was included in the final dataset analysis. While under a constant estimated effect-site concentration of 2.5 ng/mL of remifentanil, an 1% propofol infusion was started at 3.3 mL//h until LOR. The level of responsiveness was assessed by an anesthesiologist every six seconds using a modified version of the Richmond Agitation-Sedation Scale (aRASS). The frontal EEG was acquired using a bilateral bispectral (BIS VISTA™ v2.0, Medtronic, Ireland) sensor. EEG data was then split into 5-second epochs, and for each epoch, the anesthesiologist's classification was used to label it as responsiveness (no-LOR) or unresponsiveness (LOR). All 5-second epochs were then used as inputs for the CNN model to classify the untrained segment as no-LOR or LOR.ResultsThe CNN model was able to identify the transition from no-LOR to LOR successfully, achieving 97.90±0.07% accuracy on the cross-validation set.ConclusionThe obtained results showed that the proposed CNN model was quite efficient in the responsiveness/unresponsiveness classification. We consider our approach constitutes an additional technique to the current methods used in the daily clinical setting where LOR is identified by the loss of response to verbal commands or mechanical stimulus. We therefore hypothesized that automated EEG analysis could be a useful tool to detect the moment of LOR, especially using machine learning approaches.