A monitoring and regulating system and method for anesthesia maintenance period

Abstract

The application discloses a monitoring and regulation system and method for an anesthesia maintenance period, and the system comprises: a database for storing vital sign data and drug injection data of the whole process of the anesthesia maintenance period; and an anesthesia auxiliary platform configured with a monitoring and regulation model, which is used for intelligent regulation of an anesthesia cycle in the anesthesia maintenance period and provision of an anesthesia maintenance strategy. The application performs real-time monitoring on basic life functions of a human body in the anesthesia maintenance period, combines an anesthesia drug artificial intelligence intelligent regulation model, intelligently evaluates anesthesia risks, and intelligently regulates the dosage of anesthesia drugs, which is beneficial to control of anesthesia quality, can achieve optimal anesthesia effect by using the least anesthesia drugs, and shortens the postoperative recovery time.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical treatment, and particularly relates to a monitoring and regulation system and method for an anesthesia maintenance period. BACKGROUND

[0002] General anesthesia is a special and complex state, including sedation and hypnosis, amnesia, analgesia, stress inhibition and muscle relaxation, and many other factors. During the induction, maintenance and recovery of general anesthesia, the dosage of anesthetic drugs is controlled. Different surgical methods, different human individuals and different degrees of intraoperative stimulation also require different dosages of anesthetic drugs. Insufficient dosage of anesthetic drugs will cause shallow anesthesia depth, which will cause adverse reactions such as blood flow fluctuation, intraoperative awareness and body movement, and will cause serious mental and cardiovascular reactions or make the operation unable to proceed smoothly; excessive dosage of anesthetic drugs will cause deep anesthesia, which will cause serious inhibition of nervous system, circulatory system and other systems, affect the prognosis of patients, and even endanger life.

[0003] With the rapid development of machine learning technology, it has been applied to many fields such as image recognition, image processing, user analysis, etc. In the field of anesthesia, due to the scarcity of anesthesiologists, machine learning technology has been tried to be applied to anesthesia. That is, through machine learning, the amount of anesthetic drugs used is determined instead of anesthesiologists, so as to alleviate the problem of resource scarcity, improve the accuracy of drug use, and reduce the hidden dangers caused by inaccurate anesthetic dosage. SUMMARY

[0004] In view of the above deficiencies in the prior art, the monitoring and regulation system and method for an anesthesia maintenance period provided by the present application reduce the safety hazards caused by inaccurate anesthetic dosage due to fatigue or lack of experience in manual control.

[0005] In order to achieve the above-mentioned purposes, the technical solution adopted by the present application is as follows: a monitoring and regulation system for an anesthesia maintenance period, comprising:

[0006] a database for storing vital sign data and drug injection data of the anesthesia maintenance period;

[0007] an anesthesia auxiliary platform configured with a monitoring and regulation model, for intelligent regulation of anesthesia circulation during the anesthesia maintenance period, and providing an anesthesia maintenance strategy; wherein the anesthesia maintenance strategy includes regulation strategies for anesthesia depth and vital signs;

[0008] The monitoring and regulation model is obtained by learning the data stored in the database;

[0009] The monitoring and regulation model automatically regulates the depth of anesthesia by real-time monitoring of vital signs and depth of anesthesia data, provides recommended drug injection amount, and updates the monitoring and regulation model based on the changes in vital signs and depth of anesthesia data corresponding to the actual drug injection amount, thereby maintaining the depth of anesthesia and circulation.

[0010] Further, the monitoring and regulation model comprises a vital sign and depth of anesthesia monitoring unit and an anesthesia maintenance unit;

[0011] The monitoring unit is used for monitoring vital signs and depth of anesthesia, and providing data basis for anesthesia depth maintenance strategy;

[0012] The anesthesia maintenance unit is used for calculating the drug injection dose according to the current depth of anesthesia and vital sign data.

[0013] Further, the monitoring unit monitors the depth of anesthesia index as arterial blood pressure data.

[0014] Further, the anesthesia maintenance unit is an optimal adaptive control model based on reinforcement learning.

[0015] A monitoring and regulation method for anesthesia maintenance period, comprising the following steps:

[0016] S1, constructing a clinical anesthesia operation database, and training the monitoring and regulation model using the same;

[0017] S2, determining the target vital sign and depth of anesthesia range of the current surgical patient, and calculating the anesthesia drug injection dose;

[0018] S3, real-time monitoring of vital signs and depth of anesthesia data during the operation by the monitoring unit;

[0019] S4, adjusting the current anesthesia drug injection dose by the anesthesia maintenance unit based on the currently monitored vital sign and depth of anesthesia data;

[0020] S5, repeating steps S3-S4 to maintain the vital sign and depth of anesthesia data within the target range, and realizing the monitoring and regulation of the anesthesia maintenance period.

[0021] Further, in step S1, the data in the clinical anesthesia operation database includes patient's sign state, action and reward information; the action is the amount of anesthesia drug injection at each time;

[0022] The sign state includes the patient's vital sign data, including heart rate, oxygen saturation, respiratory rate, tidal volume, arterial blood pressure and BIS data;

[0023] The reward function corresponding to the reward information is:

[0024]

[0025] wherein, is the absolute target error, ρ1a k is the anesthetic drug dosage penalty, is the transition anesthetic drug penalty, r k is the reward value, r(·) is the reward function, o k is the k-th time point of the vital sign state vector, is the mean arterial blood pressure error at k+1 time point, a k is the action selected at k time point, i.e., the anesthetic drug infusion dosage, ρ1 is the anesthetic drug dosage penalty weight, and ρ2 is the excessive anesthetic drug dosage penalty weight.

[0026] Further, in the step S1, the objective function when the risk assessment and regulation model is trained is:

[0027]

[0028] wherein, α is the regularization weight, E S～D is the expectation of the vital sign state s, the vital sign state s is subject to the distribution of the data set D, and Q(s,a) is the state-action value function, is the expectation of the random variable s, and s is subject to the distribution of the data set D collected when the strategy is adopted, and Q is the Q function network, is the target Q value, s is the observed state, a is the action, and D is the training data set, is the output anesthetic drug injection strategy, and s' is the transition state at the next time point.

[0029] Further, the monitored anesthetic depth indicators in the step S3 include arterial blood pressure and BIS data, which include the previous time point mean arterial pressure, target mean arterial pressure, mean arterial pressure error, mean arterial pressure change, and current time point mean arterial pressure and BIS data.

[0030] The present application has the following advantages:

[0031] (1) The present application considers that only using one kind of linear or nonlinear analysis cannot effectively reflect the complex brain electrical activity. A new method of singular spectrum nonlinear analysis combined with Fourier transform linear analysis is adopted to denoise and extract the EEG features of the EEG signal. The singular spectrum analysis adopted can effectively remove the artifacts in the original EEG signal and extract α, β, θ, δ. Seven EEG signal features are obtained by calculating the β ratio, frequency domain and time domain sample entropy.

[0032] (2) The present application monitors the basic life functions of the human body in the anesthesia maintenance period in real time, combines an artificial intelligence intelligent control model of anesthetic drugs, intelligently assesses the anesthesia risk, and intelligently controls the dosage of anesthetic drugs, which is beneficial to control the anesthesia quality and can achieve the best anesthesia effect with the least anesthetic drugs, and shorten the postoperative recovery time. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The monitoring and control method flowchart of the anesthesia maintenance period provided by the present application. DETAILED DESCRIPTION

[0034] The specific embodiments of the present application are described below to facilitate those skilled in the art to understand the present application, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0035] Example 1:

[0036] The embodiment of the present application provides a monitoring and control system for the anesthesia maintenance period, comprising:

[0037] A database for storing vital sign data and drug injection data of the anesthesia maintenance period;

[0038] An anesthesia auxiliary platform configured with a monitoring and control model for intelligently controlling the anesthesia cycle in the anesthesia maintenance period and providing anesthesia maintenance strategies; wherein the anesthesia maintenance strategies include anesthesia depth and vital sign control strategies;

[0039] The monitoring and control model is obtained by learning the data stored in the database;

[0040] The monitoring and control model automatically controls the anesthesia depth by real-time monitoring of the vital signs and anesthesia depth data, provides the recommended drug injection amount, and updates the monitoring and control model based on the changes of the vital signs and anesthesia depth data corresponding to the actual drug injection amount, thereby realizing the maintenance of anesthesia depth and cycle.

[0041] The monitoring and control model in the embodiment of the present application includes a vital sign and anesthesia depth monitoring unit and an anesthesia maintenance unit; the monitoring unit is used for monitoring the vital signs and anesthesia depth and providing data basis for anesthesia depth maintenance strategies; the anesthesia maintenance unit is used for calculating the drug injection dose according to the current anesthesia depth and vital sign data.

[0042] The anesthesia depth index monitored by the monitoring unit in the embodiment of the present application is arterial blood pressure data, which includes previous time average arterial pressure, target average arterial pressure, average arterial pressure error, average arterial pressure change and current time average arterial pressure.

[0043] The anesthesia depth index of the monitoring unit in the embodiment of the present application can also be brain electrical double frequency index BIS data, which mainly reflects the brain electrical activity after the inhibition of the action of anesthetic drugs on the cerebral cortex, indirectly reflects the anesthesia depth, and appropriate BIS maintenance interval can reduce the incidence of postoperative cognitive dysfunction, accelerate anesthesia recovery, reduce postoperative anesthesia-related mortality and prevent intraoperative awareness.

[0044] The monitoring and control model in the embodiment of the present application is an optimal adaptive control model based on reinforcement learning. In use, according to the target blood pressure value fluctuation range, the system automatically controls the anesthesia depth, and at the same time cooperates with the size of the heart rate for collaborative judgment. When the blood pressure and heart rate are within the target range, the anesthesia depth BIS detection value is controlled within the interval of 40-60, and when it exceeds the normal interval during the operation, the system automatically controls the dosage of propofol and remifentanil.

[0045] Embodiment 2:

[0046] The embodiment of the present application provides a monitoring and control method based on the monitoring and control system of the anesthesia maintenance period in embodiment 1, as shown in Figure 1 , comprising the following steps:

[0047] S1, constructing a clinical anesthesia operation database, and training the monitoring and control model by using the same;

[0048] S2, determining the target vital signs and anesthesia depth range of the current surgical patient, and calculating the anesthesia drug injection dose;

[0049] S3, monitoring the vital signs and anesthesia depth data in the operation process in real time through the monitoring unit;

[0050] S4, based on the currently monitored vital signs and anesthesia depth data, adjusting the current anesthesia drug injection dose through the anesthesia maintenance unit;

[0051] S5, repeating steps S3-S4, so that the vital signs and anesthesia depth data are maintained within the target range, and the monitoring and control of the anesthesia maintenance period are realized.

[0052] The data in step S1 of the embodiment of the present application

[0053] The data in the clinical anesthesia operation database includes the sign state, action and reward information of the patient during the operation; the action is the amount of anesthesia drug injection at each time;

[0054] The vital sign state includes patient's vital sign data, including heart rate, oxygen saturation, respiratory rate, tidal volume, arterial blood pressure and BIS data;

[0055] The reward function corresponding to the reward information is:

[0056]

[0057] In the formula, is an absolute target error, and ρ1a k is a penalty of anesthetic drug dosage, is a penalty of excessive anesthetic drug, and r k is a reward value, r(·) is a reward function, o k is a vital sign state vector at the kth moment, is an error of mean arterial blood pressure at the k+1th moment, a k is an action selected at the kth moment, i.e., an anesthetic drug infusion amount, ρ1 is a penalty weight of anesthetic drug dosage, and ρ2 is a penalty weight of excessive anesthetic drug dosage.

[0058] In step S1 of the embodiment of the application, the objective function when the risk assessment and regulation model is trained is:

[0059]

[0060] In the formula, α is a regularization weight, E S～D is an expectation of a vital sign state s, the vital sign state s is subjected to a distribution of a data set D, and Q(s,a) is a state-action value function, is an expectation of a random variable s, and s is subjected to a distribution of a data set D collected when a strategy is adopted, and Q is a Q function network, is a target Q value, s is an observed state, a is an action, and D is a training data set, is an output anesthetic drug injection strategy, and s' is a transition state at the next moment.

[0061] The objective function in the embodiment of the application includes two parts, the latter part is to estimate the state-action value function using the time difference method, and the former part is to increase the distance between the Q estimation value and the true value, so that the state-action value function Q estimation is more conservative, thereby avoiding the problem that the Q network overestimates the value function.

[0062] In the training process, the exponential moving average method is used to update the target feature network and the target Q function network respectively until the model converges or a specified training number is reached, and the output strategy network is the current anesthetic drug injection amount.

[0063] In step S3 of the embodiment of the present application, the monitored anesthesia depth index includes arterial blood pressure, which includes the previous time point mean arterial pressure, target mean arterial pressure, mean arterial pressure error, mean arterial pressure change, and current time point mean arterial pressure. Among them, the target arterial pressure is the mean arterial pressure that the doctor expects to maintain during the operation; the mean arterial pressure error = current time point arterial pressure error-target arterial pressure error; the mean arterial pressure change = current time point arterial pressure error-previous time point arterial pressure error.

[0064] In step S3 of the embodiment of the present application, the anesthesia depth index can also be BIS data; the monitoring method of BIS data is specifically:

[0065] A1, acquiring an EEG original signal;

[0066] A2, processing the EEG original signal by a singular spectrum analysis method, and reconstructing the EEG signal;

[0067] A3, performing power spectrum calculation, beta ratio calculation, and sample entropy calculation on the reconstructed EEG signal to obtain corresponding power spectrum, beta ratio, and sample entropy in time domain and frequency domain;

[0068] A4, taking the calculated power spectrum, beta ratio, and sample entropy in time domain and frequency domain as inputs of a BIS prediction model to obtain BIS data.

[0069] In step A2 of the embodiment of the present application, the singular spectrum analysis is used to process the original EEG signal, the single-channel original EEG signal is converted into a trajectory matrix, the trajectory matrix is singular value decomposed in a high-dimensional space, the decomposed feature vectors are grouped according to a rule, and the original signal is reconstructed. The singular spectrum analysis in the embodiment mainly includes two parts of decomposition and reconstruction, the singular spectrum decomposition includes time delay embedding matrix construction and singular value decomposition, and the singular spectrum reconstruction includes feature vector grouping and diagonal line averaging.

[0070] In the embodiment, the EEG signal is processed by singular spectrum analysis for 0-50Hz band-pass filtering, the energy spectrum of alpha, beta, theta, and delta bands is extracted by using discrete Fourier transform, the beta ratio, frequency domain sample entropy, and time domain sample entropy are calculated, and then seven EEG extraction and derived features are taken as input features of a deep neural network model. The time processing window of the EEG signal transformation is 1 second, that is, 500 continuous sampling points are used for Fourier transform and sample entropy calculation.

[0071] Therefore, in step A3 of the embodiment, the power spectrum includes the power spectrum corresponding to the alpha, beta, theta, and delta bands;

[0072] The method for calculating the power spectrum is specifically: performing Fourier transform on the EEG original signal, dividing the transformed signal into several segments, using a window function for each segment of signal to obtain a corresponding power spectrum, wherein the window function is a Hanning window function.

[0073] The calculation formula of the beta ratio BR is:

[0074]

[0075] The frequency band of the beta ratio in the embodiment is the logarithm of the power spectrum ratio of the frequency band 30-47Hz to the frequency band 11-20Hz, and the index can reflect the activity of the beta frequency band.

[0076] The sample entropy in the embodiment is a nonlinear dynamic parameter for quantifying the occurrence rate of new subsequences in a time sequence, and is also a measure index of irreversibility or randomness.

[0077] The calculation formula of the sample entropy SE(m, r, N) in the embodiment is:

[0078]

[0079] In the formula, m and r are the allowable standard deviations of the time sequence, N is the length of the time sequence, c m (r) is a similar sequence of m consecutive data points;

[0080] The BIS prediction model in step A4 is a deep neural network including 5 hidden layers; specifically, the deep neural network framework of Tensorflow is used for training and testing, and finally the obtained deep neural network includes 5 hidden layers, the number of neural nodes of each hidden layer is 2048, 1024, 512, 256 and 128 respectively, the epoch of the model is 100000, the initial learning rate is 0.001, the dropout ratio is 0.3 (i.e. 30% of the neural nodes will be randomly shielded in the training process to prevent the model from overfitting), and the loss function is a logarithmic likelihood loss function.

[0081] In the embodiment of the present application, the EEG signal is weak and is easily affected by other interference signals such as electrocardiogram signals, electromyogram signals and medical electric knife frequencies. In addition, general anesthetics, opioid drugs and muscle relaxants can all affect the EEG signal. Therefore, only using a linear or nonlinear analysis cannot effectively reflect the complex brain electrical activity. The embodiment adopts a new method of singular spectrum nonlinear analysis combined with Fourier transform linear analysis to denoise and extract the EEG features of the EEG signal. The singular spectrum analysis adopted can effectively remove the artifacts in the original EEG signal and extract alpha, beta, theta and delta, and seven EEG signal features are obtained by calculating the beta ratio, frequency domain and time domain sample entropy. Previous studies have also shown that singular spectrum analysis can effectively denoise the original EEG signal and is an effective nonlinear analysis method for extracting the frequency of the EEG signal, which can extract the alpha frequency from the original EEG signal and accurately distinguish the open-eye state and the closed-eye state.

[0082] In the optimal adaptive control model in the embodiment of the present application, the Agent is a learner and a decision maker, similar to the controller in the traditional control theory. Unlike the supervised learning method, no examples of desired behavior are provided in the structured training process. On the contrary, the desired result is achieved through positive or negative reinforcement assigned by the critic according to the desirability of the result. Therefore, the RL controller learns in a biologically inspired habit, which experiments on its environment in a trial-and-error manner. The Agent selects a control action and observes its consistency, obtains positive reinforcement for favorable results and negative reinforcement for unfavorable results, and iteratively updates to meet the given performance metrics.

[0083] In the embodiment, it is assumed that there is a narcotic drug dosage control task, and a RL algorithm is considered. For example: a patient receiving BIS-guided anesthesia can be observed in 3 states: BIS increases (ΔBIS>0), BIS is stable (ΔBIS=0) or BIS decreases (ΔBIS<0); it can adjust the BIS level of the patient during the operation by administering a dose of propofol from the following set of actions: 0mg, 10mg or 40mg. The optimal control strategy is found through the above control model, and the experimental data show that under the condition of reducing the BIS value, the selection of 40mg has the highest utility; similarly, when the BIS value increases, the selection of 0mg has the highest utility. By comparing the action value function, the propofol dose with the maximum utility can be determined for each patient state. In this way, the optimal action selection for each patient state is determined. The optimal control strategy is to collect the optimal operation for all possible patient state sets.

[0084] Based on this, the anesthesia maintenance unit in the embodiment of the present application considers the clinical differences of gender, age, weight, pharmacokinetics and pharmacodynamics of patients, first wants to use a series of clustering methods to preprocess historical anesthesia maintenance period patient vital sign data and drug injection data, then builds the corresponding lifting network framework, and carries out reinforcement training on the data, evaluates the injection strategy of the anesthesia maintenance unit by analyzing the anesthesia continuous processing strategy of the anesthesia maintenance period that has been successfully completed, and calculates the average return of each treatment scheme. If the infusion strategy provided causes the patient's vital signs to exceed the normal fluctuation range, a certain penalty coefficient will be given, and conversely if it is maintained within the normal fluctuation range, a reward coefficient will be given. Learn the model through this idea, the actor-critic structure of the algorithm evaluates the current intraoperative control vital sign strategy, updates the intraoperative patient continuous injection anesthesia drug dose in real time to maintain normal vital signs and anesthesia level, and finally obtains an anesthesia maintenance unit.

[0085] In the description of the present application, it should be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features. Therefore, the features limited by "first", "second", "third" can explicitly or implicitly include one or more of the features.

Claims

1. A monitoring and regulation system for an anesthesia maintenance period, characterized in that, The method comprises the following steps: a database for storing vital sign data and drug injection data during the whole process of anesthesia maintenance; an anesthesia auxiliary platform configured with a monitoring and control model for intelligent control of anesthesia circulation during anesthesia maintenance and providing anesthesia maintenance strategies; wherein the anesthesia maintenance strategies include control strategies for anesthesia depth and vital signs; the monitoring and control model is obtained by learning the data stored in the database; the monitoring and control model automatically controls the anesthesia depth by real-time monitoring of vital sign and anesthesia depth data, provides recommended drug injection amount, and updates the monitoring and control model based on the changes of vital sign and anesthesia depth data corresponding to the actual drug injection amount, thereby realizing the maintenance of anesthesia depth and circulation; the data in the database includes the state of the patient's vital signs, actions and reward information during the operation; the action is the amount of anesthetic drug injection at each time point; wherein the vital sign data includes heart rate, oxygen saturation, respiratory rate, tidal volume, arterial blood pressure and BIS data; the reward function corresponding to the reward information is: wherein, is the absolute target error, is the anesthetic drug dose penalty, is the transition anesthetic drug penalty, is the reward value, is the reward function, is the state vector of the patient at the k time instant, is the mean arterial blood pressure error at the k+ 1 time instant, is the action selected at the k time instant, i.e. the anesthetic drug infusion amount, is the anesthetic drug dose penalty weight, is the excessive anesthetic drug dose penalty weight; the objective function for training the monitoring and control model is: wherein is a regularization weight, is a sign state s is a desire, sign state s is a distribution of the dataset D is a distribution of the dataset is a state-action value function, is a random variable s is a desire, and s is subject to a distribution of the dataset collected when policy D is employed, is a Q-function network, is a target Q-value, is an observed state, a is an action, D is a training dataset, is an output anesthetic injection policy, is a transition state at the next time instant.

2. The system for monitoring and regulation of the maintenance phase of anaesthesia according to claim 1, characterized in that, the monitoring and control model includes a vital sign and anesthesia depth monitoring unit and an anesthesia maintenance unit; the monitoring unit is used for monitoring vital signs and anesthesia depth and providing data basis for anesthesia depth maintenance strategies; the anesthesia maintenance unit is used for calculating the drug injection dose according to the current anesthesia depth and vital sign data.

3. The system for monitoring and regulation of the maintenance phase of anaesthesia according to claim 2, characterized in that, The anesthesia depth index monitored by the monitoring unit is arterial blood pressure data.

4. The system for monitoring and regulation of the maintenance phase of anaesthesia according to claim 2, characterized in that, The anesthesia maintenance unit is an optimal adaptive control model based on reinforcement learning.

Images