Detecting the state of drowsiness induced by propofol by spatial filter and machine learning algorithm on EEG recording

To accurately measure the depth of anesthesia has been a challenge for both anesthesiologists and engineers who work on developing tools of measurements. This study aims to use a machine-learning algorithm to predict the drowsy state, a transitional depth of sedation during propofol anesthesia. The data used in this study were scalp EEG (electroencephalogram) recordings selected from the University of Cambridge Repository. Raw EEG recordings were preprocessed into power spectrum matrices one second per sample. A total of 170 samples (110 awake samples and 60 drowsy samples) were used. A CNN (Convolutional Neural Network) for the MNIST (Modified National Institute of Standards and Technology) dataset was applied on these EEG power spectrum matrices. Due to the small dataset volume, Leave-One-Out cross-validation was used to train the data. Results of the training accuracy reached 99.69%. And test accuracy averaged 96.47%. Overall, the model is able to predict the state of drowsiness during propofol anesthesia. This provides the potential to develop EEG monitoring devices with closed-loop feedback of such a machine learning algorithm that controls the titration of the dosage of anesthetic administration and the depth of anesthesia.