Research On Monitoring Of Depth Of Anesthesia Based On Convolutional Neural Network

In clinical surgery,it is necessary to accurately and accurately determine the depth of anesthesia(DOA)of the patient in order to guide the physician to anesthesia medication,and to prevent the secondary injury caused by the patient’s consciousness or brain trauma.In general,the assessment of patients’ DOA by an anesthesiologist based on a commercial anesthesia instrument is costly and doesn’t guarantee the robustness of prolonged anesthesia.With the progress of biomedicine,anesthesia analysis methods based on electroencephalography(EEG)signals have gradually emerged.However,most of them are based on small sample data,which cannot solve the differences among patients in practical application,making modern clinical Anesthesia trapped in the "over-fitting" phase.Therefore,the research and development of anesthesia depth monitoring for large sample data is significant to achieve accurate anesthesia in clinical procedures.Considering the problems above,this paper introduces the Convolutional Neural Network(CNN)model into DOA analysis,combining with multi-type patient cases and senior anesthesiologists’ clinical anesthesia experience,and thus establishing CNN model for DOA classification,and designing to implement the whole DOA monitoring process.The main contents are as follows:(1)The existence of the instability and low signal-to-noise ratio from EEG signal making it hard to intuitively reflect the anesthetic characteristics,researching the relationship between time domain and time-frequency domain of EEG signal in anesthesia,a time-varying window short-time Fourier transform is proposed to convert the EEG signal into anesthetic spectrum,and through the multi-segment EEG signal enhanced by the DOA of the patient during the clinical anesthesia surgery,analyzing the time-varying window of the adjacent time interval signal,selecting and adjusting the relevant parameters of the proposed method according to the experimental results.On this basis,the appropriate anesthetic spectrum is selected as the input of the twodimensional convolutional neural network,and the extraction method is theoretically verified and analyzed in anesthesiology.(2)In order to solve the problems caused by individualized differences in clinical anesthesia,studying a series of CNN models with different architectures based on a large number of anesthetic spectrum samples,combined with high-performance GPU and Theano platform,small samples under soft and hardware constraints are used to assess the applicability of the CNN network architecture.In the meantime,according to CNN’s performance feedback results in DOA classification,CNN architecture parameters and network layer depth are continuously adjusted to obtain a CNN model for generalized EEG-DOA evaluation.Finally,the cross validation and confusion matrix experiments were designed to analyze and verify the validity of the model.(3)Design and implement the EEG-DOA monitoring system based on the CNN model according to the precise monitoring requirements of patients in the clinical operation of DOA level.In this process,the physiological information database of the anesthesia patient is set up to collect and store the patient anesthesia data in real time,and then according to the relevant information transmitted,combined with the EEG data preprocessing module and the CNN model framework after the training to achieve the DOA levels of the clinical anesthesia patient in real-time monitoring.A simple intelligent conversion to DOA level is designed for the phenomenon of data loss during monitoring,which effectively improves system performance and paves the way for future research on intelligent products.