Research On Anesthesia Depth Recognition Based On Recurrent Neural Network

Nowadays,anesthesia monitoring gained increasing attention for its role in the guarantee of the intraoperative safety and postoperative recovery of patients.However,the improper dosage of anesthetics often leads to some known side effects.For instance,the low dosage can be associated with the risks of hemodynamic instability,or intraoperative awareness;while the postoperative recovery time was prolonged at the high drug dosage.Even in some serious situation when the respiratory depressions occurred,the patients’ safety were threatened.Therefore,the on-line and accurate monitoring of anesthetic depth is of a great significance for the intraoperative safety of the patients,along with a comfortable postoperative recovery.As a common anesthetic monitoring device,the current foreigner made BIS calculated the anesthetic depth based on the database of the foreigner patients,which limited the direct application of these devices in Chinese patients.In this regards,we developed a real-time anesthesia depth prediction algorithm based on Elman recurrent neural network model,which accommodates the Chinese patients.In addition,the relationship between multimodal signal and the anesthetic depth was assessed as well.Based on the EEG signal data of 62 patients with general anesthesia,the EEG signals were divided into four states:awake,mild anesthesia,moderate anesthesia and over-anesthesia.Firstly,in view of the low signal-to-noise ratio of EEG signals,this paper proposes a variety of signal preprocessing methods such as high-low-pass filtering,power frequency notching,baseline correction and independent component analysis.Then the electromyography,clectro-oculography,power frequency interference,baseline drift and electrical interference are effectively suppressed,and the signal-to-noisc ratio of the signal is improved.Then,the time and frequency domain characteristics and nonlinear dynamic characteristics of EEG signals under different anesthesia conditions are analyzed.The beta ratio,alpha ratio,burst suppression ratio,wavelet entropy,permutation entropy and myoelectric ratio are used as joint parameter feature inputs in Elman recurrent neural network model prediction algorithm.The research results show that the Elman recurrent neural network model proposed in this paper has higher prediction accuracy and prediction probability,which are 98.02%and 0.9522,respectively.At the same time,based on the multimodal data of 7 patients during the induction of clinical anesthesia,the relationship between changes of multimodal related characteristics and anesthesia state was analyzed.The accuracy of multimodal signal characteristics in the process of anesthesia identification was analyzed.The prediction probability of anesthesia status with multimodal characteristics is 0.9432.In summary,the anesthesia depth prediction algorithm proposed in this paper has a better prediction accuracy and prediction probability.The clinical experimental results show that the algorithm can reflect the depth of anesthesia in a better way,and its result has practical reference significance.At the same time,the prediction algorithm based on the characteristics of multimodal signals can preferably reflect the different anesthesia states during anesthesia induction.This conclusion provides a new idea for future research on the depth prediction of anesthesia based on multimodality.