Research On Monitoring Technology Of Depth Of Anesthesia Based On Deep Neural Network

Background and Objectives:Accurate monitoring of depth of anesthesia(Do A)is crucial to prevent intraoperative awareness,maintain an appropriate depth of anesthesia,reduce the dosage of anesthetics and complications related to general anesthesia,and enhance recovery after surgery.With the development of precision medicine and rapid rehabilitation concept,higher requirements are put forward for depth of anesthesia monitoring.Therefore,exploring new methods of high accuracy for depth of anesthesia monitoring has become a hot research topic in recent years.Currently,there is no"gold standard"for evaluating the depth of anesthesia.Some studies have found that electroencephalogram(EEG)signals can reflect the effects of general anesthetics on the central nervous system and accurately monitor changes in the state of consciousness.Therefore,the depth of anesthesia monitor based on the EEG signals has become the mainstream method of monitoring the depth of anesthesia.However,the accuracy and necessity of its monitoring are still controversial,and the use,attitudes,and demands of Chinese anesthesiologists regarding the depth of anesthesia monitoring are still unclear.In addition,the accuracies of domestic brands compared to imported brands of the depth of anesthesia monitors are still unclear.The bispectral index(BIS)developed in the United States is currently the widely used and higher accuracy index,and a common reference standard for evaluating the depth of anesthesia,while the depth of anesthesia index(Ai)is a self-developed depth of anesthesia index in China.The accuracy of the Ai index in monitoring the depth of anesthesia is unclear.Because EEG signal is susceptible to age and various pathophysiological conditions such as hypothermia,hypoglycemia,hypoxia,etc.,and the different proprietary algorithms of the depth of anesthesia indexes,and current monitoring methods of the depth of anesthesia have some limitations.In addition,heart rate variability(HRV)in electrocardiogram(ECG)signals is jointly regulated by the central nervous and autonomic nervous system,and closely related to the effects of general anesthetics and depth of anesthesia.Therefore,it is necessary to study EEG and ECG signals in-depth and explore new methods to further improve the accuracy of the depth of anesthesia monitoring.Artificial intelligence(AI)machine learning algorithms have been a hot research topic in the depth of anesthesia monitoring in recent years,and it is not clear which of these algorithms is more accurate to monitor the depth of anesthesia.Artificial neural network(ANN)is an artificial intelligence algorithm that simulates the neural network structure of human brain,while deep neural network(DNN)is an advanced artificial neural network with stronger learning and predictive capabilities.Studies have found that based on features of EEG or ECG signals combined with DNN can monitor the depth of anesthesia.However,there is still room for improvement in the accuracy of the depth of anesthesia monitoring.Therefore,as the core evaluation content of the depth of anesthesia monitors in the national key research and development project"Evaluation on instruments of perioperative vital sign monitoring based on Internet of things",this study intends to explore the deficiencies of the depth of anesthesia monitoring through a survey on the application status of the depth of anesthesia monitoring,the difference in accuracy between Ai index and BIS through a comparative study of the two indexes,and new algorithms of the depth of anesthesia monitoring based on EEG or ECG signals features combined with DNN to improve the accuracy of depth of anesthesia monitoring.PartⅠ:A survey on the application status of the depth of anesthesia monitoring in ChinaMethods:From July 2020 to September 2020,clinical anesthesiologists were invited to participate in the online survey,and the survey questionnaires were sent to them in the form of the We Chat survey form.All survey questionnaires were anonymous.The stratified analysis,correlation analysis,and chi-square test were used to analyze the usage patterns,attitudes,and demands of clinical anesthesiologists regarding the depth of anesthesia monitors.Results:A total of 4,037 feedback questionnaires were collected from the survey.For the use of depth of anesthesia monitors,only 9.1%of the anesthesiologists routinely use depth of anesthesia monitors.The anesthesiologists working in the teaching hospitals mainly used depth of anesthesia monitoring to prevent awareness,while the anesthesiologists working in the non-teaching hospitals mainly used depth of anesthesia monitoring to guide the delivery of anesthetics.However,the limited accuracy of the depth of anesthesia monitoring,poor anti-interference ability,inability to monitor analgesia,inability to bill insurance,or high costs were crucial factors influencing the application of depth of anesthesia monitors.For the attitudes of the depth of anesthesia monitors,67.3%of the anesthesiologists had no preference in domestic and imported brands of the depth of anesthesia monitors.BIS monitor was most commonly used by 81.0%of anesthesiologists.68.5%of anesthesiologists believed that BIS is the most accurate depth of anesthesia index.For the demands of the depth of anesthesia monitors,95.7%of the anesthesiologists demanded improvements in the accuracy of the monitors for depth of anesthesia monitoring and a broad application in patients of all ages(86.3%),analgesia monitoring(80.4%),and all types of anesthetics(75.6%).In total,65.0%of the anesthesiologists believed that depth of anesthesia monitors should be combined with EEG and vital sign monitoring,and 53.7%believed that advanced depth of anesthesia monitors should include artificial intelligence.Conclusions:The use of depth of anesthesia monitors still needs to be improved,and there are many factors,including limited accuracy,affecting the use of these monitors.Most anesthesiologists have no preference in domestic and imported brands of the depth of anesthesia monitors,and are familiar with BIS,and have high acceptance of the index.The accuracy of the depth of anesthesia monitor is a significant indicator and demand for anesthesiologists to choose a depth of anesthesia monitor.The depth of anesthesia monitors with artificial intelligence may represent a new direction for future research on depth of anesthesia monitoring.PartⅡ:The accuracy of the depth of anesthesia index versus bispectral index:a prospective,multi-center,randomized controlled clinical studyMethods:From September 2020 to January 2021,145 patients with elective laparoscopic gastrointestinal surgery were recruited and included in 10 tertiary hospitals including the Second Affiliated Hospital of Army Medical University,the West China Hospital of Sichuan University,and others.The randomized,comparative study with BIS as a reference standard was conducted to explore the effectiveness and accuracy of the Ai index for depth of anesthesia monitoring.Results:The Bland-Altman consistency analysis revealed that the mean difference between the Ai index and BIS was-0.1747,95%CI(-0.6660～0.3166),P=0.4857.The regression equation of Ai index and BIS from the Deming regression analysis was:y=5.6387+0.9067x(y is BIS,x is Ai),and the slope and intercept were statistically significant.The ROC curve analysis revealed that the AUC of BIS monitoring the state of consciousness was 0.943,which was higher than 0.941 of the Ai index.There was no statistical difference between the AUCs of the two indexes(P=0.4705).The best cut-off values of monitoring loss of consciousness of the Ai index and BIS were both79.5.The AUCs of them for monitoring loss of consciousness were0.953 and 0.965,respectively,and there was a statistical difference(P=0.0013).The best cut-off values of the two indexes for monitoring recovery of consciousness were 71.5 and70.5,respectively.The AUCs of them for monitoring recovery of consciousness were 0.936and 0.934,and there was no statistical difference(P=0.5271).Conclusions:The Ai index is in good agreement with BIS,and the accuracies of the two indexes are similar in monitoring the depth of anesthesia.Both can accurately monitor the level of consciousness of patients under general anesthesia.PartⅢ:Research on monitoring depth of anesthesia based on EEG signals singular spectrum analysis combined with deep neural networkMethods:From September 2020 to January 2021,based on the EEG data of 78 subjects in the comparative study of depth of anesthesia monitors,theα,β,θ,δ,βratio,frequency domain and time-domain sample entropy were extracted by using singular spectrum analysis.These seven EEG features were used as the inputs for the deep neural network,and BIS was used as the output standard for the deep neural network to monitor the depth of anesthesia.The accuracies of the deep neural network and support vector machine in monitoring the depth of anesthesia were compared simultaneously.Results:Theα,β,θ,δ,βratio,frequency-domain,and time-domain sample entropy of EEG features were extracted by using singular spectrum analysis.These features were correlated with BIS target values,and especially theβratio and time-domain sample entropy were moderately correlated with BIS and correlation coefficients were 0.58 and 0.62,respectively.Compared with BIS target values,the mean square error of the predicted value of the deep neural network was significantly lower than that of the support vector machine,and the difference was statistically significant(c~2=15.2,P=0.0012).Conclusions:The features of EEG signal,includingα,β,θ,δ,βratio,frequency-domain,and time-domain sample entropy,can be extracted by using singular spectrum analysis.The accuracy of DNN in monitoring the depth of anesthesia outperforms the support vector machine.PartⅣ:Research on heart rate variability-derived features based on deep neural network for monitoring different anesthesia statesMethods:From March 2020 to April 2020,23 patients with elective laparoscopic surgery were recruited and enrolled in the second affiliated hospital of army medical university.The ECG data of subjects during surgery were recorded using a Philips monitor.There-sampling and data processing of original ECG data was performed.The four time-frequency features of heart rate variability,including high frequency,low frequency,high-to-low-frequency ratio,and sample entropy,were extracted as input.The expert assessment of consciousness level was used as the output standard.The four heart rate variability-derived features were combined with the deep neural network to monitor three anesthesia states:anesthesia induction,anesthesia maintenance,and anesthesia recovery.The accuracy of the deep neural network in monitoring different anesthesia states was compared with the logistic regression,support vector machine,and decision tree.Results:The accuracies of four heart rate variability-derived features,including high frequency,low frequency,high-to-low-frequency ratio,and sample entropy,combined with four machine learning algorithms in monitoring anesthesia states were 86.2%(logistic regression),87.5%(support vector machine),87.2%(decision tree),and 90.1%(deep neural network),respectively.The accuracy of the deep neural network was higher than those of the logistic regression(P<0.05),support vector machines(P<0.05),and decision trees(P<0.05).Conclusions:The incorporation of heart rate variability-derived features,including high frequency,low frequency,high-to-low-frequency ratio and sample entropy,and the DNN could accurately monitor between different anesthesia states.The accuracy of DNN outperforms the logistic regression,support vector machine,and decision tree.