Abstract:
In modern practice of major surgery using anesthesia is entirely mandatory. But due to the
failure of optimal dose of anesthetic dose delivery it is also common to the patients to face
intraoperative and postoperative complications. The main cause of the imbalance dose of
anesthesia is not being sure to assess the depth of sleep of the patient or the depth of
anesthesia or. Therefore, precise prediction of the depth of anesthesia or the proper
assessment of transitional sleep state (from deep sleep to awake) could be a way out to set the
optimal anesthetic dose by the anesthesiologist. In this work, a different approach of feature
extraction and classification method is proposed to predict three different sleep states during
surgery from the EEG signal. This work used an open-source database containing the EEG
data of anesthetic patients during surgery. The data were separated into three states: into the
deep-sleep state (IntoDeep), the deep-sleep state (InDeep), and the awake state (InAwake).
The raw EEG signals were filtered and their power spectral (PSD) densities were calculated
using MUSIC (multiple signal classification) model, a parametric method. These MUSIC
based PSD values are taken as the features of the EEG signal. An artificial neural network
model was trained to develop a machine learning based predictive model with the MUSIC
based PSD features. Finally, the predictive model was verified by the data separated for
testing and evaluated the prediction accuracy in subject-dependent and subject-independent
approach. Eventually, it is found that the results are better than the existing works those
worked on the same dataset.
