Personalized anesthesia scheme prediction system and method based on multi-modal deep learning

Abstract

The invention provides a personalized anesthesia scheme prediction system based on multi-modal deep learning, and the system comprises a physiological data collection module which is used for periodically obtaining physiological data in an operation process; the network preprocessing module is used for iteratively updating the hyper-parameter of the neural network prediction module to obtain an optimal network hyper-parameter; the neural network prediction module is used for providing a neural network model, taking the physiological data and the network hyper-parameters as input of the neural network model, and outputting predicted physiological index trends or specific parameter values to obtain a continuous numerical value prediction result; the classification prediction module is used for carrying out binarization processing on the continuous numerical value prediction result according to a set threshold value and a classification rule to generate a corresponding binary classification result; and the medication output module is used for decoding the binary classification result to obtain a personalized use scheme of the anesthetic. According to the method, accurate and efficient personalized prediction and classification of the anesthetic are realized, and anesthesia decision can be assisted.

Description

Technical Field

[0001] The present invention relates to the field of biomedical engineering anesthesia technology, and in particular to a personalized anesthesia scheme prediction system and method based on multimodal deep learning. Background Art

[0002] With the continuous enhancement of people's health awareness and the increasing diversification of eating habits, according to survey data, the number of people undergoing gastrointestinal endoscopy in China has been increasing year by year. Anesthesia painless gastrointestinal endoscopy has been proven to significantly increase the positive detection rate of gastrointestinal diseases because it can significantly reduce the pain of patients. Among all patients undergoing gastrointestinal endoscopy, about 80% of the patients are in good physical condition and belong to the ASA assessment level I to II. Such patients have a lower risk of anesthesia and usually only need to maintain an appropriate depth of anesthesia by monitoring vital signs to ensure the smooth progress of the operation.

[0003] Although the above operations are routine work for anesthesiologists, their judgment still needs to be accurate and professional, which makes the already scarce anesthesiologist resources face more severe challenges. If the anesthesiologist misses the best time to inject anesthetic drugs, it may cause the patient to move or regain consciousness during the operation, which may not only lead to surgical failure, but also increase the risk of complications such as myocardial injury for certain high-risk patient groups (such as the elderly or patients with weak constitutions).

[0004] Existing research mainly focuses on the pharmacokinetic-pharmacodynamic (PK-PD) model of drugs and drug efficacy, traditional prediction methods based on intraoperative rules, and data-driven artificial intelligence prediction methods. However, these studies have not been optimized for specific surgical environments. Different types of surgery have different requirements for anesthesia depth and drug dosage, and general models are difficult to meet specific anesthesia prediction needs. In addition, since data collection during surgery requires informed consent from patients and the collection process is complicated, the amount of data actually available is limited. Traditional deep learning methods usually rely on large-scale data sets, so their performance on small data sets is limited, and real-time prediction is required during surgery. The general model forward propagation time is long, which is difficult to meet real-time requirements. Therefore, it is urgent to improve the algorithm to improve its convergence optimization ability and computing power on small data sets, so as to achieve high accuracy and high real-time performance indicators under limited data conditions; in addition, since the algorithm is an auxiliary system and does not participate in the anesthesia decision-making process, a redundant system needs to be designed. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a personalized anesthesia plan prediction system and method based on multimodal deep learning.

[0006] According to one aspect of the present invention, a personalized anesthesia plan prediction system based on multimodal deep learning is provided, comprising:

[0007] A physiological data acquisition module, which is used to periodically acquire physiological data during surgery;

[0008] The network preprocessing module is used to iteratively update the hyperparameters of the neural network prediction module to obtain the optimal network hyperparameters;

[0009] A neural network prediction module, which is used to provide a neural network model, take the physiological data and the network hyperparameters as inputs of the neural network model, and output predicted physiological index trends or specific parameter values ​​to obtain continuous numerical prediction results;

[0010] A classification prediction module, which is used to binarize the continuous numerical prediction results according to the set threshold and classification rules to generate corresponding binary classification results;

[0011] The medication output module is used to decode the binary classification results to obtain a personalized usage plan for anesthetic drugs.

[0012] Preferably, the above system further includes:

[0013] The data input module is used to uniformly input the acquired physiological data to the upper terminal through USB, read the data from the USB interface through the python pyUSB library, and transmit it back to python for input into the neural network prediction module.

[0014] According to another aspect of the present invention, a method for predicting a personalized anesthesia regimen based on multimodal deep learning is provided, comprising:

[0015] Periodically obtain physiological data during surgery;

[0016] Providing a neural network model, iteratively updating the hyperparameters of the neural network model to obtain optimal network hyperparameters;

[0017] The physiological data and the network hyperparameters are used as inputs of the neural network model, and the predicted physiological index trend or specific parameter value is output to obtain a continuous numerical prediction result;

[0018] Binarizing the continuous numerical prediction result according to the set threshold value and classification rule to generate a corresponding binary classification result;

[0019] The binary classification result is decoded to obtain a personalized use plan of the anesthetic drug.

[0020] According to a third aspect of the present invention, there is provided a computer terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, can be used to execute the system described above in the present invention, or to execute the method described above in the present invention.

[0021] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it can be used to run the system described above in the present invention, or to execute the method described above in the present invention.

[0022] Due to the adoption of the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art:

[0023] The present invention adopts an overall auxiliary anesthesia system architecture, which solves the problem of misjudgment that may be caused by doctor fatigue during traditional anesthesia and improves the efficiency and accuracy of anesthesia decision-making.

[0024] The present invention adopts the hyperparameter geometric manifold optimization technology, which can realize the automatic update of hyperparameters and solve the problem of low model generalization ability caused by non-updated hyperparameters.

[0025] The present invention adopts an LSTM network that combines patient information and has iterative feedback, which can realize an LSTM network personalized for patients to perform anesthesia prediction, and does not need to divide the verification set, test set and training set, thus solving the problem that the traditional LSTM network cannot customize the prediction model for patients and has a small number of training data sets;

[0026] The present invention adopts sparse matrix classification technology, which can achieve high-accuracy medication classification, and solves the problems of weak generalization ability of small data sets, low prediction accuracy and long time consumption of traditional machine learning such as SVM and traditional deep learning such as CNN;

[0027] Compared with the traditional CNN prediction model, the present invention adopts an LSTM network based on a combination of data-driven and clinical experience, and its results are more accurate.

[0028] The optimization prediction classification of the present invention is simpler and more real-time, requiring only simple matrix operations, and can actually reduce the surgical burden on anesthesiologists.

[0029] The present invention not only realizes accurate, efficient and personalized prediction and classification of anesthetic drugs, significantly improves the generalization ability and prediction accuracy of the model, but also can assist in anesthesia decision-making, effectively reduce the work pressure and misjudgment risk caused by fatigue of anesthesiologists, and overcomes the problems of insufficient personalization, strong dependence on data sets and low prediction accuracy in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0031] Figure 1 This is a schematic diagram of the component modules of a personalized anesthesia plan prediction system based on multimodal deep learning in a preferred embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of information flow transmission of a personalized anesthesia plan prediction system based on multimodal deep learning in a preferred embodiment of the present invention.

[0033] Figure 3 This is a workflow diagram of a personalized anesthesia plan prediction method based on multimodal deep learning in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following is a detailed description of the embodiments of the present invention: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. It should be pointed out that for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

[0035] At present, there is a shortage of anesthesiologists, while the number of anesthesia surgeries is large. This situation can easily lead to misjudgments by anesthesiologists due to fatigue. Therefore, it is urgent to design an anesthesia prediction assistance system that can assist anesthesiologists in making judgments to improve the accuracy of judgments. There are already some technologies in the prior art for predicting the depth of anesthesia, but these technologies usually ignore the need for anesthesia for specific surgeries, have limited generalization capabilities for small data sets, and are difficult to achieve personalized neural network predictions for different patients. At the same time, the prediction accuracy still needs to be improved.

[0036] In response to the above problems, an embodiment of the present invention provides a personalized anesthesia plan prediction system based on multimodal deep learning. The system adopts an overall auxiliary anesthesia system architecture, which not only realizes accurate, efficient and personalized prediction and classification of anesthetic drugs, significantly improves the generalization ability and prediction accuracy of the model, but also can assist in anesthesia decision-making, effectively reduce the work pressure and misjudgment risk caused by fatigue of anesthesiologists, overcomes the problems of insufficient personalization, strong dependence on data sets and low prediction accuracy in the prior art, and can effectively solve the limitations of anesthesia depth prediction in the prior art.

[0037] Specifically, Figure 1 As shown, the personalized anesthesia scheme prediction system based on multimodal deep learning provided by this embodiment may include:

[0038] A physiological data acquisition module, which is used to periodically acquire physiological data during surgery;

[0039] The network preprocessing module is used to iteratively update the hyperparameters of the neural network prediction module to obtain the optimal network hyperparameters;

[0040] A neural network prediction module, which is used to provide a neural network model, take physiological data and network hyperparameters as inputs of the neural network model, and output predicted physiological index trends or specific parameter values ​​to obtain continuous numerical prediction results;

[0041] A classification prediction module, which is used to binarize the continuous numerical prediction results according to the set threshold and classification rules to generate corresponding binary classification results. The binary classification results are used to characterize whether anesthetic drugs need to be used and what kind of anesthetic drugs to use;

[0042] Medication output module, which is used to decode the binary classification results and obtain a personalized use plan for anesthetic drugs.

[0043] In some preferred embodiments, the above system may further include:

[0044] The data input module is used to uniformly input the acquired physiological data to the upper terminal through USB, read the data from the USB interface through the python pyUSB library, and transmit it back to python for input into the neural network prediction module.

[0045] The following is a detailed description of the technical means by which the functional modules of the system provided in this embodiment implement corresponding functions in conjunction with the preferred implementation manner.

[0046] A preferred embodiment of the present invention provides a personalized anesthesia plan prediction system based on multimodal deep learning, including: a physiological data acquisition module, a network preprocessing module, a neural network prediction module, a classification prediction module and a medication output module.

[0047] In some preferred embodiments, the physiological data acquisition module: collects each physiological data according to the set period, and updates the collected data in real time; wherein, for physiological data for which no new data is collected, the data bit is temporarily covered by the last collection result until new valid data is collected; wherein, the physiological data includes: heart rate, blood oxygen saturation, systolic pressure, diastolic pressure and / or bispectral index BIS. Further preferably, the heart rate, blood oxygen saturation, systolic pressure, diastolic pressure, and bispectral index BIS are collected, and the data is updated with a collection period of every 10 seconds by default. For some new data that are difficult to obtain in a short time (such as systolic pressure and diastolic pressure), if the new value cannot be obtained in real time, the data bit will be temporarily covered by the last collection result until a new round of valid data collection is successful.

[0048] In some preferred implementations, the network preprocessing module preprocesses the preset hyperparameters to obtain the most suitable network hyperparameters, which is one of the inputs to the neural network.

[0049] The information flow of the network preprocessing module is as follows Figure 2 As shown in part A, it is necessary to input the hyperparameters of the previous round and send the three-dimensional hyperparameters to the hyperparameter geometric manifold optimization network to obtain the hyperparameters of the new round. This module further includes:

[0050] Packaging submodule: Packaging the preset or previous round of hyperparameters (h n-1 ,l n-1 ,e n-1 ) is packaged into the loop (h 0 ,l 0 ,e 0 );

[0051] Hyperparameter optimization submodule: 0 ,l 0 ,e 0 ) Input a geometric manifold optimization algorithm F, update the hyperparameters to obtain (h 1 ,l 1 ,e 1 ), where F is represented by:

[0052]

[0053] Where N is the number of local optimal solutions; A i is the height of the i-th local maximum, representing the corresponding local optimal performance; μ i is the position of the i-th local maximum in the hyperparameter space; σ i is the width around the i-th peak. Use the gradient descent update method to update (h 0 ,l 0 ,e0 ) is updated to get (h 1 ,l 1 ,e 1 );

[0054] Evaluation submodule: 1 ,l 1 ,e 1 ) is sent into an evaluation function E to obtain the evaluation index c 1 , the evaluation function E is expressed as:

[0055]

[0056] In the formula, t represents the time consumed for each calculation, ξ represents the bi-norm of the hyperparameter, and C 1 , C 2 is the correction constant. The real-time performance and stability of the algorithm can be comprehensively considered through the evaluation function. This submodule obtains the evaluation index c 1 ;

[0057] Iterative optimization submodule: Save c 1 and its corresponding (h 1 ,l 1 ,e 1 ), and (h 1 ,l 1 ,e 1 ) Input the hyperparameter optimization submodule for the next round of iteration and get c 2 and its corresponding (h 2 ,l 2 ,e 2 );

[0058] Compare 1 and c 2 If the value of c 1 >c 2 , save c 1 and (h 1 ,l 1 ,e 1 ), if c 1 ≤c 2 , save c 2 and (h 2 ,l 2 ,e 2 );

[0059] Re-enter the iteration, loop to get the smallest c and its corresponding (h, l, e), and output it as (h n ,l n ,e n ).

[0060] In some preferred implementations, the neural network model provided by the neural network prediction module adopts an LSTM network model.

[0061] In some preferred embodiments, the neural network prediction module: after completing the hyperparameter preprocessing, the physiological data (heart rate, blood oxygen, blood pressure, BIS, etc.) and the optimized network hyperparameters are input into the neural network prediction module. In this process, the network model continuously iterates and updates the parameters, and outputs the predicted physiological index trend or specific parameter value, which will serve as the basis for subsequent decision-making.

[0062] The information flow of the neural network prediction module is as follows Figure 2 As shown in Part B, the neural network prediction module needs to call the hyperparameter results (h n ,l n ,e n ), and feed these parameters into the neural network model (such as LSTM network) for iteration. In addition, this module also needs to call the patient's real-time victory data input by the data input module. This module further includes:

[0063] Input submodule: network hyperparameters, physiological data from the 1st to the n-2th time points (such as bispectral index value, blood oxygen value, heart rate value, systolic blood pressure value, diastolic blood pressure value), and the corresponding patient's BMI value and age are sent to the neural network model;

[0064] Preliminary prediction submodule: The n-1th BIS value is predicted by the neural network model, the RMSE between it and the true value is calculated, and the RMSE and the corresponding weight value are saved;

[0065] Iterative prediction submodule: Update the physiological data at different time points and the corresponding patient's BMI value and age, repeat the prediction process of the preliminary prediction submodule, obtain the new RMSE and the corresponding weight value, compare the new RMSE with the old RMSE, and save the smaller RMSE value and the corresponding weight;

[0066] Continue to iterate to obtain the minimum RMSE value and the corresponding weight file;

[0067] Final prediction submodule: Using the weight file, the physiological data from the 1st to the n-1th time points (such as bispectral index value, blood oxygen value, heart rate value, systolic blood pressure value, diastolic blood pressure value) are used to predict the physiological data at the nth time point (such as bispectral index value, blood oxygen value, heart rate value, systolic blood pressure value, diastolic blood pressure value).

[0068] In some preferred embodiments, the classification prediction module receives the continuous numerical prediction results output by the neural network prediction module, and binarizes them through specific thresholds and classification rules, and finally generates a classification result (expressed in binary code) on whether anesthetic drugs are needed and what kind of anesthetic drugs to use.

[0069] The information flow of the classification prediction module is as follows Figure 2 As shown in Part C, this module uses a sparse matrix based on experience and data-driven classification prediction. It needs to call the prediction results of the neural network prediction module and send the prediction results to the sparse matrix prediction to obtain the predicted binary results. The binary results are then sent to the medication output module for decoding to obtain the final results of whether to use the drug and what drug to use. This module further includes:

[0070] Setting submodule: setting thresholds and classification rules, including: the threshold of BIS value is 0 to 1 when it is less than 60, the threshold of heart rate is 0 to 1 when it is less than 50, and the threshold of systolic blood pressure is 0 to 1 when it is less than 90;

[0071] Sparse matrix submodule: According to the set threshold and classification rules, a sparse matrix based on experience and data is provided. In the specific sparse matrix, the BIS threshold is represented and segmented by the function R(x), the heart rate threshold is represented and segmented by the function F(x), and the systolic blood pressure threshold is represented and segmented by the function P(x). At the same time, according to clinical experience, in the matrix M 5×7 Each row in is set with a different divisor, representing the order of weights from low to high over time. The sparse matrix based on experience and data driving provided by the sparse matrix submodule is as follows:

[0072] M prediction =M 5×7 ·D 7×5 ·W 5×1

[0073]

[0074]

[0075] W 5×1 =[1 0 1 0 1] T

[0076]

[0077] Where M prediction Represents the prediction matrix, M 5×7 represents a matrix of 5 rows and 7 columns used to summarize data, D 7×5 represents a real-time matrix of patient physiological data with 7 rows and 5 columns, W 5×1represents a matrix with 5 rows and 1 column, F(x) represents the function of classifying heart rate threshold, P(x) represents the function of classifying systolic blood pressure threshold, R(x) represents the function of classifying BIS threshold, s(x) represents the value function, R n ,S n ,U n ,O n ,B n Indicates the heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen, BIS value at the nth time node, pre r Represents the predicted value of heart rate, pre s Represents the predicted value of systolic blood pressure, pre u represents the predicted value of diastolic blood pressure, o Indicates the predicted value of blood oxygen, pre b It represents the predicted value of BIS value;

[0078] Classification submodule: Input the prediction results into a sparse matrix to obtain the predicted binary classification results.

[0079] In some preferred embodiments, the medication output module decodes the binary results output by the classification prediction module to clarify the type of anesthetic drugs and dosage recommendations to be used at the next moment, providing support and reference for actual medical decision-making.

[0080] In some preferred embodiments, the above system may further include:

[0081] Data input module: The data collected by each instrument is uniformly input into the computer through USB. The Python pyUSB library reads the data from the USB interface and transmits it back to Python as one of the inputs of the neural network.

[0082] Based on the same inventive concept, an embodiment of the present invention also provides a personalized anesthesia plan prediction method based on multimodal deep learning.

[0083] Specifically, Figure 3 As shown, the personalized anesthesia scheme prediction method based on multimodal deep learning provided by this embodiment may include:

[0084] S1, periodically obtain physiological data during surgery;

[0085] S2, providing a neural network model, iteratively updating the hyperparameters of the neural network model to obtain the optimal network hyperparameters;

[0086] S3, taking the physiological data and the network hyperparameters as the input of the neural network model, and outputting the predicted physiological index trend or specific parameter value to obtain a continuous numerical prediction result;

[0087] S4, binarizing the continuous numerical prediction result according to the set threshold and classification rule to generate a corresponding binary classification result, and the binary classification result is used to characterize whether anesthetic drugs need to be used and what kind of anesthetic drugs to use;

[0088] S5, decode the binary classification results to obtain a personalized use plan for anesthetic drugs.

[0089] In some preferred embodiments, the above S1, periodically acquiring physiological data during surgery, may further include:

[0090] Physiological data, including: heart rate, blood oxygen saturation, systolic blood pressure, diastolic blood pressure and / or bispectral index (BIS);

[0091] Each physiological data is collected according to the set period, and the collected data is updated in real time; among them, for physiological data for which no new data is collected, the data bit is temporarily covered by the last collection result until new valid data is collected.

[0092] In some preferred implementations, the above S2, the neural network model, adopts an LSTM network model.

[0093] In some preferred implementations, the above S2, iteratively updating the hyperparameters of the neural network model to obtain the optimal network hyperparameters, may further include:

[0094] S21, set the preset or previous round hyperparameters (h n-1 ,l n-1 ,e n-1 ) is packaged into the hyperparameters of the loop (h 0 ,l 0 ,e 0 );

[0095] S22, the hyperparameters of the loop (h 0 ,l 0 ,e 0 ) Input a hyperparameter geometric manifold optimization algorithm F and update the new hyperparameter (h 1 ,l 1 ,e 1 ), where the hyperparameter geometric manifold optimization algorithm F is expressed as:

[0096]

[0097] In the formula, h represents the hidden layer dimension, l represents the number of network layers, e represents the number of cycles, N represents the number of local optimal solutions, and A represents the number of local optimal solutions. i is the height of the i-th local maximum, representing the corresponding local optimal performance; μ iis the position of the i-th local maximum in the hyperparameter space; σ i is the width around the i-th peak; n represents the number of n-th cycle rounds; the gradient descent update method is used to update (h 0 ,l 0 ,e 0 ) is updated to obtain the new hyperparameter (h 1 ,l 1 ,e 1 );

[0098] S23, (h 1 ,l 1 ,e 1 ) is sent into an evaluation function E to obtain the evaluation index c 1 ; Among them, the evaluation function E is expressed as:

[0099]

[0100] In the formula, t represents the time consumed for each calculation, ξ represents the bi-norm of the hyperparameter, and C 1 , C 2 is the correction constant, n is the number of n-th cycle rounds, and Z is the integer domain; the real-time performance and stability of the algorithm can be comprehensively considered through the evaluation function E. This step obtains the evaluation index c 1 ;

[0101] S24, save the evaluation index c 1 and its corresponding hyperparameters (h 1 ,l 1 ,e 1 ), and the hyperparameter (h 1 ,l 1 ,e 1 ) to the next iteration, repeat S22 and S23 to get the evaluation index c 2 and its corresponding hyperparameters (h 2 ,l 2 ,e 2 );

[0102] S25, comparison c 1 and c 2 If the value of c 1 >c 2 , save c 1 And the hyperparameters (h 1 ,l 1 ,e 1 ), if c 1 ≤c 2 , save c 2 And the hyperparameters (h 2 ,l 2 ,e 2 );

[0103] S26, re-enter the iteration, loop to obtain the minimum evaluation index c and its corresponding hyperparameters (h, l, e), and output the optimal network hyperparameters (h n ,l n ,e n ).

[0104] In some preferred embodiments, the above S3, taking the physiological data and the network hyperparameters as the input of the neural network model, and outputting the predicted physiological index trend or specific parameter value to obtain a continuous numerical prediction result, may further include:

[0105] S30, inputting the network hyperparameters into the neural network model;

[0106] S31, inputting the physiological data from the 1st to the n-2th time points and the BMI value and age of the corresponding patient into the neural network model;

[0107] S32, predicting the BIS value at the n-1th time point through the neural network model, calculating the RMSE between the predicted BIS value and the true BIS value, and saving the RMSE and the corresponding weight value;

[0108] S33, repeat S31 and S32 to obtain a new RMSE and a corresponding weight value, compare the new RMSE with the old RMSE, and save the smaller RMSE value and the corresponding weight;

[0109] S34, continuously iterate to obtain the minimum RMSE value and the corresponding weight file;

[0110] S35, using the weight file and the physiological data from the 1st to the n-1th time points as inputs of the neural network model, and outputting the prediction result at the nth time point.

[0111] In some preferred implementations, the above S4, binarizing the continuous numerical prediction result according to the set threshold and classification rule to generate the corresponding binary classification result, may further include:

[0112] Setting thresholds and classification rules, including: the threshold of BIS value is less than 60 and changes from 0 to 1, the threshold of heart rate is less than 50 and changes from 0 to 1, and the threshold of systolic blood pressure is less than 90 and changes from 0 to 1;

[0113] According to the set threshold and classification rules, a sparse matrix based on experience and data driving is provided. Specifically, the BIS threshold is represented and segmented by the function R(x), the heart rate threshold is represented and segmented by the function F(x), and the systolic blood pressure threshold is represented and segmented by the function P(x). In addition, according to clinical experience, anesthesiologists are more likely to consider more recent data, so we use the matrix M5×7 Each row in the table has a different divisor, such as 60, -60, 30, -30, 15, -15, which represents the order of weights from low to high over time, which is closer to the actual decision-making process;

[0114] Input the prediction results into the sparse matrix to obtain the predicted binary classification results;

[0115] Among them, the sparse matrix is:

[0116] M prediction =M 5×7 ·D 7×5 ·W 5×1

[0117]

[0118] W 5×1 =

[10101] T

[0119]

[0120] Where M prediction Represents the prediction matrix, M 5×7 represents a matrix of 5 rows and 7 columns used to summarize data, D 7×5 represents a real-time matrix of patient physiological data with 7 rows and 5 columns, W 5×1 represents a matrix with 5 rows and 1 column, F(x) represents the function of classifying heart rate threshold, P(x) represents the function of classifying systolic blood pressure threshold, R(x) represents the function of classifying BIS threshold, s(x) represents the value function, R n ,S n ,U n ,O n ,B n Respectively represent the heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen and BIS value at the nth time node, pre r Represents the predicted value of heart rate, pre s Represents the predicted value of systolic blood pressure, pre u represents the predicted value of diastolic blood pressure, o Indicates the predicted value of blood oxygen, pre b Indicates the predicted value of BIS value.

[0121] It should be noted that the steps in the method provided by the present invention can be implemented by using the corresponding components in the system, and those skilled in the art can refer to the technical solution of the system to implement the step flow of the method, and can also refer to the technical solution of the method to implement the composition of the system, that is, the embodiments in the system and the embodiments in the method can be understood as preferred examples of each other, which will not be elaborated here.

[0122] The technical solution provided by the above embodiment of the present invention is further described in detail below in conjunction with a specific application example.

[0123] In this specific application example, the subject is a patient with ASA level I to II, and the environment is a gastrointestinal endoscopy. The patient is 58 years old, 171 cm tall, 85 kg in weight, and has a BMI index of 29.1.

[0124] During the operation, physiological data (including heart rate, blood pressure, blood oxygen saturation and bispectral index BIS) are collected in real time through sensors and transmitted to the computer every 10 seconds. Since anesthetics have been injected at the beginning of the operation, in order to avoid the fluctuations in the initial anesthesia stage affecting the prediction results, the data collected one minute before the operation will not be included in the prediction. One minute after the start of the operation, physiological data such as heart rate, blood oxygen, and BIS values ​​are updated every 10 seconds, while blood pressure requires additional time for detection and is updated every minute. The above collected data are summarized in the computer to form a 1×5 input matrix, and each column of the matrix is: [heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen, BIS] T , after obtaining the input data, the geometric manifold optimization method is first used to update the hyperparameters of the model to ensure the efficiency and accuracy of the prediction process. Subsequently, the previous BIS value, the patient's age, and the BMI value are input into the LSTM network as features. Through the cyclic calculation of LSTM, the model gradually optimizes its weights to obtain the best prediction performance. Using the optimized model weights, the vital signs of the next moment are predicted, including heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen saturation, and BIS value. These predicted values ​​are combined into a new 1×5 matrix again. During the prediction process, the values ​​of the output matrix are meaningful only in the first, third, and fifth columns, corresponding to whether the heart rate, systolic blood pressure, and BIS value are abnormal: the value of the first column is 1, indicating that the heart rate is too low and heart rate drugs (such as atropine) need to be injected; the value of the third column is 1, indicating that the systolic blood pressure is too low and pressor drugs (such as ephedrine or norepinephrine) need to be injected; the value of the fifth column is 1, indicating that the BIS value is too low and anesthetic drugs (such as propofol, sufentanil, or remifentanil) need to be injected. For example, if the final output matrix is ​​[1,0,0,0,0] T , indicating that the patient's heart rate is too low and atropine needs to be injected; while the blood pressure and BIS values ​​are normal, so no other drugs need to be injected. Through this process, the model realizes real-time prediction of whether drugs need to be injected and what drugs to inject.

[0125] An embodiment of the present invention further provides a computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it can be used to run the system of any one of the above embodiments of the present invention, or execute any one of the methods of the above embodiments of the present invention.

[0126] Optionally, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDR SDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory is used to store computer programs (such as applications, functional modules, etc. that implement the above method), computer instructions, etc., and the above computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. And the above computer programs, computer instructions, data, etc. can be called by the processor.

[0127] The processor is used to execute the computer program stored in the memory to implement the various steps of the method or various modules of the system involved in the above embodiments. For details, please refer to the relevant descriptions in the above method and system embodiments.

[0128] The processor and the memory may be independent structures or integrated structures. When the processor and the memory are independent structures, the memory and the processor may be coupled and connected via a bus.

[0129] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it can be used to run the system of any one of the above embodiments of the present invention, or to execute the method of any one of the above embodiments of the present invention.

[0130] Among them, computer-readable media include computer storage media and communication media, wherein the communication media include any media that facilitates the transmission of computer programs from one place to another. The storage medium can be any available medium that can be accessed by a general or special-purpose computer. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC. In addition, the ASIC can be located in a user device. Of course, the processor and the storage medium can also be present in a communication device as discrete components.

[0131] The personalized anesthesia scheme prediction system and method based on multimodal deep learning provided by the above-mentioned embodiment of the present invention adopts an overall auxiliary anesthesia system architecture, solves the problem of misjudgment that may be caused by doctor fatigue during traditional anesthesia, and improves the efficiency and accuracy of anesthesia decision-making; adopts a hyperparameter geometric manifold optimization method, which can realize automatic updating of hyperparameters, and solves the problem of low model generalization ability caused by non-updated hyperparameters; adopts an LSTM network that combines patient information and has iterative feedback, which can realize personalized patient LSTM network for anesthesia prediction, and does not need to divide the verification set, test set and training set, which solves the problem that the traditional LSTM network cannot customize the prediction model for the patient and the training data set is small; adopts a sparse matrix classification method, which can achieve high accuracy The medication classification solves the problems of weak generalization ability of traditional machine learning such as SVM and traditional deep learning such as CNN with small data sets, low prediction accuracy and long time consumption; compared with the traditional CNN prediction model, it adopts the LSTM network based on data-driven and clinical experience, and its results are more accurate; the optimization prediction classification is simpler and has better real-time performance, only requires simple matrix operations, and can actually reduce the surgical burden of anesthesiologists; it not only realizes accurate, efficient and personalized prediction and classification of anesthetic drugs, significantly improves the generalization ability and prediction accuracy of the model, but also assists in anesthesia decision-making, effectively reduces the work pressure and misjudgment risk of anesthesiologists caused by fatigue, and overcomes the problems of insufficient personalization, strong dependence on data sets and low prediction accuracy in existing technologies.

[0132] The personalized anesthesia plan prediction system and method based on multimodal deep learning provided in the above embodiments of the present invention are particularly suitable for using an artificial intelligence network to assist in predicting whether to use anesthetics and what kind of anesthetics to use during the anesthesia process of gastrointestinal endoscopy.

[0133] All matters not covered in the above embodiments of the present invention are well known in the art.

[0134] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art may make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A personalized anesthesia plan prediction system based on multimodal deep learning, characterized in that: include: A physiological data acquisition module, which is used to periodically acquire physiological data during surgery; The network preprocessing module is used to iteratively update the hyperparameters of the neural network prediction module to obtain the optimal network hyperparameters; A neural network prediction module, which is used to provide a neural network model, take the physiological data and the network hyperparameters as inputs of the neural network model, and output predicted physiological index trends or specific parameter values ​​to obtain continuous numerical prediction results; A classification prediction module, which is used to binarize the continuous numerical prediction results according to the set threshold and classification rules to generate corresponding binary classification results; The medication output module is used to decode the binary classification results to obtain a personalized usage plan for anesthetic drugs.

2. The personalized anesthesia plan prediction system based on multimodal deep learning according to claim 1, characterized in that: The physiological data acquisition module collects each physiological data according to a set period and updates the collected data in real time; wherein, for physiological data for which no new data is collected, the data bit is temporarily covered by the last collection result until new valid data is collected; The physiological data include: heart rate, blood oxygen saturation, systolic blood pressure, diastolic blood pressure and / or bispectral index (BIS).

3. The personalized anesthesia plan prediction system based on multimodal deep learning according to claim 1, characterized in that: The network preprocessing module comprises: Packaging submodule: Packaging the preset or previous round of hyperparameters (h n-1 ,l n-1 ,e n-1 ) are packaged into the hyperparameters of the loop (h0, l0, e0); Hyperparameter optimization submodule: The hyperparameters (h0, l0, e0) of the loop are input into a hyperparameter geometric manifold optimization algorithm F, and the new hyperparameters (h1, l1, e1) are updated; wherein the hyperparameter geometric manifold optimization algorithm F is expressed as: In the formula, h represents the hidden layer dimension, l represents the number of network layers, e represents the number of cycles, N represents the number of local optimal solutions, and A represents the number of local optimal solutions. i is the height of the i-th local maximum, representing the corresponding local optimal performance; μ i is the position of the i-th local maximum in the hyperparameter space; σ i is the width around the i-th peak; n represents the number of n-th cycles; the gradient descent update method is used to update (h0, l0, e0) to obtain new hyperparameters (h1, l1, e1); Evaluation submodule: (h1, l1, e1) is sent to an evaluation function E to obtain the evaluation index c1; wherein the evaluation function E is expressed as: t represents the time consumed for each calculation, ξ represents the second norm of the hyperparameter, C1 and C2 are correction constants, n is the number of cycles, and Z is the integer domain. The evaluation function E can comprehensively consider the real-time performance and stability of the algorithm, and obtain the evaluation index c1; Iterative optimization submodule: save the evaluation index c1 and its corresponding hyperparameters (h1, l1, e1), and input the hyperparameters (h1, l1, e1) into the hyperparameter optimization submodule for the next round of iteration to obtain the evaluation index c2 and its corresponding hyperparameters (h2, l2, e2); compare the values ​​of c1 and c2, if c1>c2, save c1 and hyperparameters (h1, l1, e1); if c1≤c2, save c2 and hyperparameters (h2, l2, e2); re-enter the iteration, loop to obtain the minimum evaluation index c and its corresponding hyperparameters (h, l, e), and output the optimal network hyperparameters (h n ,l n ,e n ).

4. The personalized anesthesia plan prediction system based on multimodal deep learning according to claim 1, characterized in that: The neural network model provided by the neural network prediction module adopts an LSTM network model.

5. The personalized anesthesia plan prediction system based on multimodal deep learning according to claim 1, characterized in that: The neural network prediction module comprises: Input submodule: inputting the network hyperparameters, the physiological data from the 1st to the n-2th time points, and the BMI value and age of the corresponding patient into the neural network model; Preliminary prediction submodule: predict the BIS value at the n-1th time point through the neural network model, calculate the RMSE between the predicted BIS value and the true BIS value, and save the RMSE and the corresponding weight value; Iterative prediction submodule: Update the physiological data at different time points and the corresponding patient's BMI value and age, repeat the prediction process of the preliminary prediction submodule above, obtain the new RMSE and the corresponding weight value, compare the new RMSE with the old RMSE, save the smaller RMSE value and the corresponding weight; continue to iterate to obtain the minimum RMSE value and the corresponding weight file; Final prediction submodule: takes the weight file and the physiological data from the 1st to the n-1th time points as inputs of the neural network model, and outputs the final prediction result at the nth time point.

6. The personalized anesthesia plan prediction system based on multimodal deep learning according to claim 1, characterized in that: The classification prediction module comprises: Setting submodule: setting thresholds and classification rules, including: the threshold of BIS value is 0 to 1 when it is less than 60, the threshold of heart rate is 0 to 1 when it is less than 50, and the threshold of systolic blood pressure is 0 to 1 when it is less than 90; Sparse matrix submodule: provides an experience-based and data-driven sparse matrix according to the set threshold and classification rules; Classification submodule: input the prediction result into the sparse matrix to obtain the predicted binary classification result; The sparse matrix submodule provides an empirical and data-driven sparse matrix, specifically: M prediction =M 5×7 ·D 7×5 ·W 5×1 W 5×1 ＝[10101] T Where M prediction Represents the prediction matrix, M 5×7 represents a matrix of 5 rows and 7 columns used to summarize data, D 7×5 represents a real-time matrix of patient physiological data with 7 rows and 5 columns, W 5×1 represents a matrix with 5 rows and 1 column, F(x) represents the function of classifying heart rate threshold, P(x) represents the function of classifying systolic blood pressure threshold, R(x) represents the function of classifying BIS threshold, s(x) represents the value function, R n ,S n ,U n ,O n ,B n Respectively represent the heart rate, systolic blood pressure, diastolic blood pressure, blood oxygen and BIS value at the nth time node, pre r Represents the predicted value of heart rate, pre s Represents the predicted value of systolic blood pressure, pre u represents the predicted value of diastolic blood pressure, o Indicates the predicted value of blood oxygen, pre b Indicates the predicted value of BIS value.

7. The personalized anesthesia plan prediction system based on multimodal deep learning according to any one of claims 1 to 6, characterized in that: Also includes: The data input module is used to uniformly input the acquired physiological data to the upper terminal through USB, read the data from the USB interface through the python pyUSB library, and transmit it back to python for input into the neural network prediction module.

8. A personalized anesthesia plan prediction method based on multimodal deep learning, characterized in that: include: Periodically obtain physiological data during surgery; Providing a neural network model, iteratively updating the hyperparameters of the neural network model to obtain optimal network hyperparameters; The physiological data and the network hyperparameters are used as inputs of the neural network model, and the predicted physiological index trend or specific parameter value is output to obtain a continuous numerical prediction result; Binarizing the continuous numerical prediction result according to the set threshold value and classification rule to generate a corresponding binary classification result; The binary classification result is decoded to obtain a personalized use plan of the anesthetic drug.

9. A computer terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When executing the computer program, the processor can be used to run the system described in any one of claims 1 to 7, or to execute the method described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it can be used to run the system described in any one of claims 1 to 7, or to execute the method described in claim 8.

Images