Invasive systolic blood pressure prediction method and system during general anesthesia based on machine learning fusion

Abstract

The present invention is a method and system for predicting invasive systolic blood pressure under general anesthesia based on machine learning fusion. The method includes step S1, collecting the patient's physical feature data and vital sign data, preprocessing each data, and dividing the preprocessed data into a training set, a test set, and a validation set; slicing and sampling the training set, the test set, and the validation set, respectively, to obtain n training subsets, test subsets, and validation subsets; step S2, constructing a prediction model based on machine learning fusion, and using the prediction model to predict invasive systolic blood pressure under general anesthesia; the prediction model includes n primary learners and one secondary learner, and the output values ​​of all primary learners are used as inputs of the secondary learner. The system includes a database module, a data acquisition module, a prediction module, and an interactive module; a prediction program is stored in the prediction module. The present invention overcomes the defect of low prediction accuracy of a single algorithm, and optimizes the input of the secondary learner to further improve the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical surgical operation monitoring, and in particular relates to a method and system for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion. Background Art

[0002] During general anesthesia surgery, invasive systolic blood pressure is an important basis for doctors to judge the patient's vital signs, which will directly affect the smooth progress of the operation. Accurately predicting the changes in invasive systolic blood pressure during general anesthesia and making response strategies in advance can effectively reduce the risk of surgery. At present, life monitoring instruments are generally used to monitor invasive systolic blood pressure, blood oxygen, heart rate and other variables in real time during surgery. Doctors need to manually predict invasive systolic blood pressure based on clinical experience, but the results of manual prediction are not ideal. On the one hand, doctors have different clinical experience and professional levels, which leads to different judgments on the situation; on the other hand, the accuracy of manual prediction for continuous variables is limited. During surgery, if you want to quickly and accurately predict invasive systolic blood pressure after general anesthesia, you need to rely on a lot of professional knowledge, and you also need to consider the combined influence of multiple factors during the operation, which makes it difficult to accurately predict by manual methods. This manual prediction method not only has extremely high requirements for doctors' clinical experience, but also takes a lot of time and energy, affecting the progress of the operation.

[0003] With the development of machine learning technology, machine learning algorithms can not only learn a lot of knowledge through mathematical models and mechanism formulas, but also dig out deep-level connections or patterns between data, and can quickly and accurately calculate and process patients' physical characteristics, basic disease conditions, and vital signs. In addition, the amount of data required for machine learning is relatively small, and the prediction of invasive systolic blood pressure after general anesthesia can be achieved by establishing an offline model. Not only is the prediction speed fast and accurate, but it also saves a lot of time and energy, providing a guarantee for the safety of anesthesia and surgery.

[0004] TH Wu et al. (TH Wu, GK Pang and EW Kwong, "Predicting Systolic Blood Pressure Using Machine Learning" in 7th International Conference on Information and Automation for Sustainability, Colombo, Sri Lanka, 2014, pp. 1-6.) used machine learning technology to predict systolic blood pressure by modeling characteristic variables such as body mass index (BMI), age, exercise, drinking and smoking. They used back propagation neural networks and radial basis function networks to build and verify the prediction system. The prediction accuracy of the prediction system was as high as 69.9% when the absolute error was 15 mmHg, and the prediction precision was relatively low.

[0005] In summary, the present invention proposes a method and system for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion, and adopts the idea of ​​machine learning fusion to construct a prediction model, which significantly improves the prediction accuracy. Summary of the invention

[0006] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide a method and system for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion.

[0007] The technical solution adopted by the present invention to solve the technical problem is:

[0008] A method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion, characterized in that it comprises the following steps:

[0009] Step S1, collecting the patient's physical characteristic data and vital sign data, preprocessing each data, and dividing the preprocessed data into a training set, a test set, and a validation set; slicing and sampling the training set, the test set, and the validation set respectively to obtain n training subsets, test subsets, and validation subsets;

[0010] Step S2: construct a prediction model based on machine learning fusion, and use the prediction model to predict the invasive systolic blood pressure under general anesthesia; the prediction model includes n primary learners and one secondary learner, and the output values ​​of all primary learners are used as inputs of the secondary learner.

[0011] Step S2 includes S2-1, constructing a primary learner: using the training set to train multiple machine learning models, and optimizing the trained models according to the test set to obtain an optimized model; using the validation set to make predictions according to the optimized model, selecting n machine learning models with high prediction accuracy to respectively construct primary learners, and obtaining n primary learners;

[0012] The n training subsets obtained in step S1 are used to train n primary learners respectively, and the parameters of the corresponding primary learners are optimized by using n test subsets to obtain optimized primary learners; and prediction is performed according to the n optimized primary learners by using n validation subsets;

[0013] S2-2. Construct a secondary learner: take the predicted values ​​of all primary learners as the input of the secondary learner, train the secondary learner, and obtain the trained secondary learner.

[0014] Step S2-2 also includes optimizing the input of the secondary learner; specifically, assigning a weight to each primary learner according to the prediction error of the primary learner, and in the process of assigning the weight, the larger the prediction error, the lower the weight;

[0015] Let the input matrix of the secondary learner be h1(x), h2(x), …, h n (x) are the predicted values ​​of each primary learner, ω1, ω2, …, ω n are the weights of the prediction values ​​of each primary learner, where ω1+ω2+…+ω n =1;

[0016] The weights are adjusted using the coefficient of variation method, and the coefficient of variation of each primary learner satisfies formula (2):

[0017]

[0018] In formula (2), V i is the coefficient of variation of the ith primary learner, σ i is the standard deviation between the predicted value and the true value of the ith primary learner, h i '(x) is the true value corresponding to the i-th primary learner;

[0019]

[0020] The weight of the predicted value of the i-th primary learner satisfies formula (3).

[0021] Preprocessing includes missing value filling, attribute identification, standardization and dimensionality reduction. The specific process of slice sampling is: using the gradient boosting decision tree to train the sample data set after dimensionality reduction to obtain the probability of each feature, that is, the feature importance list;

[0022] The sample data set after dimensionality reduction is divided into training set, test set and validation set, and the training set, test set and validation set are copied n times respectively, where n is a positive integer; according to the sample sampling ratio, the copied training set, test set and validation set are sliced ​​and sampled n times respectively, and then feature selection is performed on the feature importance list according to the feature sampling ratio to obtain n training subsets, test subsets and validation subsets.

[0023] There are three primary learners, which are constructed using support vector regression model, polynomial regression model and linear regression model respectively; the secondary learner is constructed using ridge regression model.

[0024] A general anesthesia invasive systolic blood pressure prediction system based on machine learning fusion, characterized in that the system includes a database module, a data acquisition module, a prediction module and an interaction module; the prediction module stores a prediction program, and the prediction program predicts the general anesthesia invasive systolic blood pressure according to the above method.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The prediction model of the present invention is constructed based on the idea of ​​machine learning fusion. The prediction model includes multiple primary learners and one secondary learner. The multiple primary learners are trained based on different machine learning algorithms. The prediction results of all primary learners are used as the input of the secondary learner, which effectively overcomes the defect of low prediction accuracy of a single algorithm. In order to further improve the accuracy of the prediction model, a weight is given to the prediction result of each primary learner according to the prediction error of the primary learner. The larger the prediction error, the lower the weight, which optimizes the input of the secondary learner and further improves the prediction accuracy of the prediction model.

[0027] 2. Experimental results show that the prediction accuracy of the method of the present invention for invasive systolic blood pressure under general anesthesia is as high as 91.5%. Compared with the study of THWu et al., the accuracy is improved by about 20% when the absolute error is the same. Even when the research background and characteristic variables are different, it shows that the prediction accuracy of the present invention is high, and it can achieve rapid and accurate prediction of the invasive systolic blood pressure of patients after general anesthesia, giving doctors sufficient time to formulate response strategies, ensuring the safety of anesthesia, and greatly reducing the risk of surgery.

[0028] 3. The features collected by the present invention include the patient's physical characteristics data and vital signs data, and take into account the combined impact of multiple factors on the invasive systolic blood pressure of general anesthesia to the greatest extent. From the perspective of medical expertise and the actual situation on the operating room site, most of the factors that affect the results are taken into account, and the data are deeply mined using machine learning algorithms, which makes up for the deficiency of medical staff in judging the invasive systolic blood pressure of general anesthesia based on vital signs data. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the overall flow chart of the present invention;

[0030] Figure 2 is a flow chart of constructing a prediction model according to the present invention;

[0031] Figure 3 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0032] The technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments and drawings, but the protection scope of the present application is not limited thereto.

[0033] The present invention is a method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion (hereinafter referred to as method, see Figure 1-2 ), including the following steps:

[0034] Step S1, collecting the patient's physical characteristic data and vital sign data, preprocessing each data, and dividing the preprocessed data into a training set, a test set, and a validation set; slicing and sampling the training set, the test set, and the validation set respectively to obtain n training subsets, test subsets, and validation subsets;

[0035] Step S1-1, the physical characteristic data include the patient's age, gender, height, weight, ASA grade and whether he suffers from hypertension; the physical characteristic data and basic medical conditions should be entered into the database in time before the operation, so that the data can be extracted from the database in time when the patient needs to be predicted; the prediction of invasive systolic blood pressure under general anesthesia is affected by many factors, and only the above-mentioned patient's physical characteristic data and basic medical conditions are not enough to accurately predict the patient's invasive systolic blood pressure under general anesthesia;

[0036] Vital sign data include heart rate, invasive systolic blood pressure and blood oxygen saturation (SaO2). The vital sign data collection period is the surgical preparation stage when the patient just enters the operating room. The data collection device is the life monitor equipped in the operating room, and the collection time interval is 60s. During the surgical preparation stage, the patient has no other interference factors except arterial needle and anesthesia operation, so the authenticity and accuracy of the prediction are guaranteed.

[0037] Step S1-2, preprocessing includes missing value filling, attribute identification, standardization and dimensionality reduction; if the physical feature data is missing, the sample is deleted; for vital sign data, the nearest neighbor interpolation method is used to fill individual missing values; the attributes of the features after filling the missing values ​​are identified, that is, gender and whether or not to have hypertension are assigned 0-1 respectively, and the ASA level is treated as a discrete variable. The ASA level is divided into five levels, and the corresponding values ​​are 1-5; the remaining features are treated as continuous variables;

[0038] The patient's physical characteristics data and vital signs data should be complete and in the same unit. However, due to operational errors, instrument failures or other reasons, the data may be abnormal. If one or more of the above data is missing, this data is defined as missing data. If the patient's physical characteristics data is missing, the sample is deleted. If the patient's vital signs data is missing, the missing data is filled in by the adjacent data.

[0039] The features after attribute recognition are standardized by formula (1). Standardization is to scale the feature values ​​so that each feature value falls into a small specific interval.

[0040]

[0041] Among them, X is the original eigenvalue, is the average value of the original feature sequence, S 2 is the variance of the feature sequence;

[0042] All the standardized features were processed by principal component analysis (PCA) to reduce the dimension, and the features with a high correlation with invasive systolic blood pressure during general anesthesia were screened out to obtain the sample data set after dimensionality reduction;

[0043] Step S1-3, using the Gradient Boosting Decision Tree (GBDT) to train the sample data set after dimensionality reduction, to obtain the probability of each feature. The more important the feature, the greater the probability. Thus, a probability list reflecting the importance of each feature, i.e., a feature importance list, is obtained. By selecting features according to the feature importance list, redundant features can be filtered out, and a prediction model with higher efficiency and lower consumption can be constructed.

[0044] The sample data set after dimensionality reduction is divided into training set, test set and validation set, and the training set, test set and validation set are copied n times, where n is a positive integer; according to the sample sampling ratio, the copied training set, test set and validation set are sliced ​​and sampled n times, and then feature selection is performed on the feature importance list according to the feature sampling ratio to obtain n training subsets, test subsets and validation subsets;

[0045] Step S2: construct a prediction model based on machine learning fusion, where the prediction model includes n primary learners and one secondary learner;

[0046] Step S2-1, constructing a primary learner;

[0047] Use the training set to train a variety of machine learning models, and optimize the trained models based on the test set. Use grid search to optimize the parameters of various algorithms to obtain the optimized model. The principle of grid search is to arrange and combine all the parameters or hyperparameters of the model. All the arrangements and combinations form a two-dimensional grid. Multiple hyperparameters can be regarded as a grid in a higher-dimensional space. All nodes in the grid are traversed, and the optimal solution in the grid is used as the final parameter of the model.

[0048] A variety of machine learning models including polynomial regression (Lasso), linear regression (LR), support vector regression (SVR), K-nearest neighbor algorithm (KNN), extreme gradient boosting tree (XGBoost), multi-layer perceptron (MLP), decision tree (DecisionTree), random forest (RF), Adaboost, gradient boosting decision tree (GBDT) and ridge regression (Ridge); the validation set was used to obtain the prediction results according to the optimized model, and the prediction accuracy of various models was calculated. The absolute error of 15 mmHg was used as the prediction accuracy evaluation index to calculate the initial accuracy and optimized accuracy of various models, and the results shown in Table 1 were obtained;

[0049] Table 1 Prediction results of various machine learning models

[0050]

[0051] Select n machine learning models with higher prediction accuracy to construct primary learners respectively, and obtain n primary learners; as shown in Table 1, within the absolute error of ±15 mmHg, the prediction accuracy of the three machine learning models of SVR, Lasso and LR is higher, so this embodiment selects these three models to construct three primary learners;

[0052] The n training subsets obtained in step S1-3 are used to train n primary learners respectively, and the parameters of the corresponding primary learners are optimized through n test subsets to obtain optimized primary learners; the invasive systolic blood pressure under general anesthesia is predicted according to the n optimized primary learners using n validation subsets;

[0053] Step S2-2, constructing a secondary learner;

[0054] The predicted values ​​of all primary learners are used as the input of the secondary learner, and the secondary learner is trained to obtain the trained secondary learner;

[0055] Theoretically, the secondary learner of machine learning fusion can use multiple models. In order to prevent overfitting, the complexity of the prediction model needs to be limited, and it is not advisable to select a machine learning model with too high complexity. In this embodiment, four common machine learning models are selected to predict invasive systolic blood pressure under general anesthesia. The prediction accuracy of various algorithms is shown in Table 2. The model with the highest prediction accuracy is selected as the secondary learner. It can be seen from Table 2 that within the absolute error of ±15 mmHg, the accuracy of ridge regression is the highest, that is, the prediction accuracy is the highest. Therefore, this embodiment uses ridge regression (Ridge) as the secondary learner.

[0056] Table 2 Prediction accuracy of secondary learners for different algorithms

[0057]

[0058] Optimize the input of the secondary learner; predict n test subsets through n optimized primary learners to obtain the prediction value of the primary learner; use the prediction value of the primary learner as the input of the secondary learner, and the prediction value of the secondary learner is the prediction value of the prediction model; assign weights to each primary learner according to the prediction error of the primary learner, so as to optimize the input of the secondary learner and complete the construction of the secondary learner; in the process of assigning weights, the larger the prediction error, the lower the weight;

[0059] Let the input matrix of the secondary learner be h1(x), h2(x), …, h n (x) are the predicted values ​​of each primary learner, ω1, ω2, …, ω n are the weights of the prediction values ​​of each primary learner, where ω1+ω2+…+ω n =1;

[0060] The coefficient of variation method is used to adjust the weights, thereby improving the accuracy of the prediction model and further reducing the prediction error; the coefficient of variation of each primary learner satisfies formula (2):

[0061]

[0062] In formula (2), V i is the coefficient of variation of the ith primary learner, σ i is the standard deviation between the predicted value and the true value of the ith primary learner, h i '(x) is the true value corresponding to the i-th primary learner;

[0063] The weight of the prediction value of the i-th primary learner is:

[0064]

[0065] In summary, the prediction model includes n primary learners and one secondary learner, and the prediction model is used for real-time prediction of invasive systolic blood pressure during general anesthesia.

[0066] The present invention also provides a general anesthesia invasive systolic blood pressure prediction system based on machine learning fusion (hereinafter referred to as system, see Figure 3 ), including a database module, a data acquisition module, a prediction module and an interaction module; the interaction module is used to manually enter the patient's physical characteristic data and display the prediction results. The physical characteristic data include the patient's age, gender, height, weight, ASA level and whether he suffers from hypertension; the database module is used to store the patient's physical characteristic data; the data acquisition module is used to collect the patient's vital signs data. The vital signs data are collected by the life monitoring instruments equipped in the hospital operating room and transmitted to the data acquisition module. The vital signs data include heart rate, invasive systolic blood pressure and blood oxygen saturation (SaO2); the prediction module stores a prediction program, and the prediction program predicts the invasive systolic blood pressure under general anesthesia according to the above-mentioned general anesthesia invasive systolic blood pressure prediction method based on machine learning fusion.

[0067] The prediction system will periodically update the prediction model during use. For example, when the predicted number of patients reaches a certain number, the system will add these new samples to the original training set, expand the original training set, and use the expanded training set to retrain the prediction model, thereby iteratively updating the prediction model. Iterative update is a basic method for optimizing the prediction model. It is through continuous iteration of the model that the error between the predicted value and the true value is gradually reduced, thereby solving the optimal model parameters. As the samples are updated, the entire iterative process continues, and the prediction model will continue to fit the newly input sample data to reduce the prediction error, thereby realizing the periodic update of the prediction model.

[0068] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion, characterized in that: The following steps are involved: Step S1, collecting the patient's physical characteristic data and vital sign data, preprocessing each data, and dividing the preprocessed data into a training set, a test set, and a validation set; slicing and sampling the training set, the test set, and the validation set respectively to obtain n training subsets, test subsets, and validation subsets; Step S2: construct a prediction model based on machine learning fusion, and use the prediction model to predict the invasive systolic blood pressure during general anesthesia; the prediction model includes n primary learners and one secondary learner, and the output values ​​of all primary learners are used as inputs of the secondary learner; Assign a weight to each primary learner according to its prediction error. In the process of assigning weights, the larger the prediction error, the lower the weight. Let the input matrix of the secondary learner be , are the predicted values ​​of each primary learner, are the weights of the prediction values ​​of each primary learner, where ; The weights are adjusted using the coefficient of variation method, and the coefficient of variation of each primary learner satisfies formula (2): （2） In formula (2), For the The coefficient of variation of the primary learner, For the The standard deviation between the predicted value and the true value of the primary learner, For the The true value corresponding to the primary learner; （3） No. The weights of the prediction values ​​of the primary learners satisfy formula (3).

2. The method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion according to claim 1, characterized in that: Step S2 includes: S2-1. Constructing primary learners: Use the training set to train multiple machine learning models, and optimize the trained models according to the test set to obtain the optimized models; use the validation set to make predictions based on the optimized models, select n machine learning models with high prediction accuracy to construct primary learners respectively, and obtain n primary learners; The n training subsets obtained in step S1 are used to train n primary learners respectively, and the parameters of the corresponding primary learners are optimized by using n test subsets to obtain optimized primary learners; and prediction is performed according to the n optimized primary learners by using n validation subsets; S2-2. Construct a secondary learner: take the predicted values ​​of all primary learners as the input of the secondary learner, train the secondary learner, and obtain the trained secondary learner.

3. The method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion according to claim 1, characterized in that: Preprocessing includes missing value filling, attribute identification, standardization and dimensionality reduction. The specific process of slice sampling is: using the gradient boosting decision tree to train the sample data set after dimensionality reduction to obtain the probability of each feature, that is, the feature importance list; The sample data set after dimensionality reduction is divided into training set, test set and validation set, and the training set, test set and validation set are copied n times respectively, where n is a positive integer; According to the sample sampling ratio, the copied training set, test set and validation set are sliced ​​and sampled n times respectively, and then feature selection is performed on the feature importance list according to the feature sampling ratio to obtain n training subsets, test subsets and validation subsets.

4. The method for predicting invasive systolic blood pressure during general anesthesia based on machine learning fusion according to claim 1, characterized in that: There are three primary learners, which are constructed using support vector regression model, polynomial regression model and linear regression model respectively; the secondary learner is constructed using ridge regression model.

5. A general anesthesia invasive systolic blood pressure prediction system based on machine learning fusion, characterized in that: The system comprises a database module, a data acquisition module, a prediction module and an interaction module; the prediction module stores a prediction program, and the prediction program predicts the invasive systolic blood pressure during general anesthesia according to any one of the methods described in claims 1-4.

Images