A method for predicting central arterial pressure based on meta-learning neural network

Abstract

A method for predicting central arterial pressure based on meta-learning neural network, which relates to the field of artificial intelligence. It includes the following steps: S1. Obtain the central arterial pressure waveform and physiological indicators (gender, height, weight, body index, age, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, cardiac output) of the patient; S2. Perform waveform decomposition based on the Gaussian function superposition method to obtain Gaussian characteristic parameters (a i , b i , c i , i = 1, 2, 3, 4); S3. Explore the significant relationship between the patient's physiological indicators and Gaussian characteristic parameters, and select the physiological indicators with significant correlation as input data; S4. Establish a meta-learning data set based on physiological indicators and Gaussian characteristic parameters; S5. Build a neural network, predict Gaussian characteristic parameters based on the meta-learning algorithm, and then obtain the central arterial pressure, and compare and verify the predicted waveform with the clinically measured waveform.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and is a method for predicting central arterial pressure based on a meta-learning neural network. Background Art

[0002] When the heart contracts and relaxes rhythmically and periodically, the impact formed by the blood ejected from the ventricle into the aorta will propagate in the form of a wave from the root of the aorta along the arterial system, that is, a central arterial pressure wave is formed. Central arterial pressure refers to the blood pressure at the root of the ascending aorta. Research shows that central arterial pressure can reflect the state of the heart and blood vessels, contains rich information on the state of the cardiovascular system, can be used to reflect the state of the cardiovascular system, predict and diagnose cardiovascular diseases, and has extremely high clinical application value.

[0003] Currently, the catheter method is still the "gold standard" for measuring the central arterial pressure waveform. However, due to its invasive measurement, complex operation, many complications, and high cost, in order to make up for the deficiencies of invasive methods, non-invasive methods such as the direct substitution method, the generalized transfer function method, and the N-point moving average method have been proposed successively to estimate the central aortic pressure. Although it has low cost, low risk, and simple operation, there are still some disadvantages, such as low accuracy, poor generality, and high cost. Therefore, it is of great significance to improve the estimation accuracy and generalization ability of central arterial pressure.

[0004] Due to the influence of environmental, genetic, and behavioral factors, blood pressure signals often exhibit a high degree of specificity and complexity. Due to these complex blood pressure characteristics, traditional and manual blood pressure feature extraction methods are becoming increasingly challenging to meet the needs of modern medical care for blood pressure detection, diagnosis, and treatment. Machine learning, as a classic algorithm in artificial intelligence, has been increasingly valued in blood pressure measurement and the diagnosis and treatment of cardiovascular diseases. Huttunen et al. proposed a method for estimating aortic pulse wave velocity and blood pressure pulse transit time using machine learning and simulated data, and initially applied machine learning algorithms to aortic research. Xiao et al. first proposed a method for estimating aortic pressure from radial artery systolic and diastolic blood pressures using an artificial neural network, and demonstrated the feasibility of using deep learning methods to detect central arterial pressure. Although the above research has achieved non-invasive estimation of central arterial pressure, since machine learning is data-driven and requires a large number of training samples, it is still difficult to collect many physiological indicators and central arterial pressure data of a large number of patients.

[0005] Meta-learning originated from general artificial intelligence, aiming to enable machines to learn how to learn. Generally, meta-learning is regarded as a supplement to deep learning models to improve the generalization ability of the models. Meta-learning is currently an important approach to solving the few-shot problem. It can reuse the learned knowledge and enable the model to quickly learn new tasks. This method combines with meta-learning algorithms to propose a method for predicting central arterial pressure based on a meta-learning neural network, which can achieve non-invasive and rapid acquisition of central arterial pressure. Its predicted waveform can be applied to non-invasive numerical calculations of clinical indicators such as instantaneous wave-free ratio (iFR) and fractional flow reserve (FFR) to measure whether a patient has myocardial ischemia. Summary of the Invention

[0006] The present invention proposes a method for predicting central arterial pressure based on a meta-learning neural network. Compared with machine learning, meta-learning is applicable to the few-shot problem, can solve the problem of small data volume, and has a relatively fast prediction speed, avoiding invasive measurement. The method for non-invasively obtaining central arterial pressure includes the following steps: acquisition of central arterial pressure waveform and physiological indicators, decomposition of the waveform by the Gaussian superposition method, exploration of the significant relationship between physiological indicators and Gaussian characteristic parameters, establishment of a meta-learning data set, construction of a neural network, prediction, and verification.

[0007] Specifically, the present invention realizes the above object through the following scheme:

[0008] A method for predicting central arterial pressure based on a meta-learning neural network, characterized by including the following steps:

[0009] S1: Obtain the central arterial pressure waveform and physiological indicators (gender, height, weight, body mass index, age, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, cardiac output) of the patient;

[0010] S2: Based on the Gaussian function superposition method, perform waveform decomposition to obtain Gaussian characteristic parameters (a i , b i , c i , i = 1, 2, 3, 4);

[0011] S3: Explore the significant relationship between the patient's physiological indicators and Gaussian characteristic parameters, and select the physiological indicators with significant correlation as input data;

[0012] S4: Establish a meta-learning data set based on physiological indicators and Gaussian characteristic parameters;

[0013] S5: Construct a neural network, predict Gaussian characteristic parameters based on the meta-learning algorithm, and then obtain the central arterial pressure, and compare and verify the predicted waveform with the clinically measured waveform.

[0014] As a technical solution of the present invention, for the features described in S1, first, the central arterial pressure waveforms of 264 patients were measured by an invasive pressure guide wire, and at the same time, ten physiological indicators corresponding to each patient were recorded, including gender, height, weight, body index, age, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, and cardiac output.

[0015] As a technical solution of the present invention, for the features described in S2, based on the Gaussian function superposition method using MATLAB R2021a software, a fourth-order Gaussian function was used to decompose the single-cycle central arterial pressure waveform. The expression of the fourth-order Gaussian function is:

[0016] where a i represents the amplitude of each Gaussian decomposition wave, b i represents the position of each Gaussian decomposition wave, c i represents the half-wave width of each Gaussian decomposition wave, and t is time.

[0017] As a technical solution of the present invention, for the features described in S3, a thermal coefficient matrix analysis was performed on the patient physiological indicators and Gaussian characteristic parameters to explore the significant relationship between the patient physiological indicators and Gaussian characteristic parameters. Physiological indicators with significant correlations were selected as the input data for meta-learning. Through correlation coefficient matrix analysis, finally, six physiological indicators, namely age (AGE), mean arterial pressure (MAP), cardiac output (CO), systolic blood pressure (SBP), diastolic blood pressure (DBP), and heart rate (HR), were selected as the input variables of the neural network.

[0018] As a technical solution of the present invention, for the features described in S4, a meta-learning dataset was established based on physiological indicators and Gaussian characteristic parameters: the physiological indicators and Gaussian characteristic parameters of all patients were divided into a training set and a test set according to a ratio of 4:1.

[0019] As a technical solution of the present invention, for the features described in S5, a neural network was built. Based on the meta-learning neural network, Gaussian characteristic parameters were predicted, and then the central arterial pressure was obtained. The predicted waveform of the central arterial pressure was compared and verified with the clinically measured waveform. The specific implementation includes the following steps:

[0020] Step B1: Import the training set, validation set, and test set into the program through the pandas module;

[0021] Step B2: Establish a fully connected layer and a dropout layer, set the total number of layers (length) of the fully connected layer, the number of neurons (width) of a single fully connected layer, and the activation function;

[0022] Step B3: Select a suitable algorithm and optimizer;

[0023] Step B4: Set hyperparameters including the learning rate α of the base learner, the learning rate β of the meta-learner, and the number of iteration steps epochs;

[0024] Step B5: Improve the loss function through the analysis of waveform feature points and the pre-training results;

[0025] Step B6: Calculate the normalized mean square error (NMSE) between the predicted central arterial pressure waveform and the clinically measured central arterial pressure waveform. An NMSE < 5% is considered that the difference between the prediction result and the clinical result is small.

[0026] The central arterial pressure waveform obtained by using the method of the present invention is relatively close to the clinically measured waveform, with small errors, and non-invasive acquisition of the central arterial pressure can be realized. Description of the Drawings

[0027] Figure 1 : Flow chart of the method of the present invention

[0028] Figure 2 : Decomposition of waveform by Gaussian function superposition method

[0029] Figure 3 : Analysis of the thermal coefficient matrix of physiological indexes and Gaussian characteristic parameters

[0030] Figure 4 : Neural network structure

[0031] Figure 5 : Comparison between the predicted central arterial pressure waveform and the clinically measured waveform Detailed Embodiments

[0032] The following explains this specification in conjunction with specific embodiments, but the present invention is not limited to the following embodiments.

[0033] Embodiment 1

[0034] A method for predicting central arterial pressure based on meta-learning neural network includes the following steps:

[0035] S1: Obtain the central arterial pressure waveform and physiological indexes (gender, height, weight, body index, age, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, cardiac output) of the patient;

[0036] S2: Based on the Gaussian function superposition method, perform waveform decomposition to obtain Gaussian characteristic parameters (a i , b i , c i , i = 1, 2, 3, 4);

[0037] S3: Explore the significant relationship between the patient's physiological indexes and Gaussian characteristic parameters, and select the physiological indexes with significant correlation as input data;

[0038] S4: Establish a meta-learning data set based on physiological indicators and Gaussian feature parameters;

[0039] S5: Build a neural network, predict Gaussian feature parameters based on the meta-learning algorithm, thereby obtaining the central arterial pressure, and compare and verify the predicted waveform with the clinically measured waveform.

[0040] As a technical solution of the present invention, for the features in S1, first, extract the central arterial pressure waveform measured clinically through a pressure guide wire through RadiView2.2 software, then extract a complete cycle of the waveform respectively, and record ten physiological indicators of the patient, including gender, height, weight, body index, age, cardiac output, heart rate, mean arterial pressure, systolic blood pressure, and diastolic blood pressure;

[0041] As a technical solution of the present invention, for the features in S2, decompose the single-cycle central arterial pressure waveform using a fourth-order Gaussian function based on the Gaussian function superposition method through MATLAB R2021a software. The expression of the fourth-order Gaussian function is:

[0042]

[0043] In the formula, ai represents the amplitude of each Gaussian decomposition wave, bi represents the position of each Gaussian decomposition wave, ci represents the half-wave width of each Gaussian decomposition wave, and t is time.

[0044] As a technical solution of the present invention, for the features in S3, explore the significant relationship between the patient's physiological indicators and Gaussian feature parameters, perform a heat coefficient matrix analysis on the patient's physiological indicators and Gaussian feature parameters through python3.7 software, and select the physiological indicators with significant correlation as input data;

[0045] As a technical solution of the present invention, for the features in S4, establish a meta-learning data set based on physiological indicators and Gaussian feature parameters: divide the physiological indicators and Gaussian feature parameters of all patients into a training set and a test set according to 4:1;

[0046] As a technical solution of the present invention, for the features in S5, build a neural network for training to realize the prediction of Gaussian feature parameters, thereby obtaining the corresponding central arterial pressure of the patient, and compare and verify the central arterial pressure prediction waveform with the clinically measured waveform. The specific implementation includes the following steps:

[0047] Step A1: Select the input variables as six physiological indicators, namely age (AGE), mean arterial pressure (MAP), cardiac output (CO), systolic blood pressure (SBP), diastolic blood pressure (DBP), and heart rate (HR), and the output variables as 12 Gaussian characteristic parameters (ai, bi, ci, i = 1, 2, 3, 4). Use four fully connected layers and two dropout layers to prevent the model from overfitting during training. The activation function used is the ReLU function:

[0048]

[0049] Step A2: Select the MAML algorithm in meta-learning for task division. Set the hyperparameters including the base learner learning rate α = 0.005, the meta-learner learning rate β = 0.1, and the number of iteration steps epochs to be 1500 - 4000. Use the Adam optimizer for training, randomly initialize the model parameters θ, and randomly extract a part of the samples from task T i Calculate the loss through the constructed network, update the gradient, and calculate the optimal parameter θ', that is

[0050]

[0051] In the formula, θ i ′ is the optimal parameter of task T i θ is the initial parameter, α is the base learner learning rate, is the gradient of task T i ;

[0052] Update the randomly initialized model parameter θ by calculating the gradient with respect to the previous optimal parameter θ', that is

[0053]

[0054] In the formula, β is the meta-learner learning rate, is the gradient of each new task Ti with respect to the optimal parameter θ';

[0055] Step A4: Combine the main characteristic points of the central arterial pressure waveform (systolic blood pressure and its corresponding time t1, diastolic blood pressure and its corresponding time t2). Use SPSS 26.0 software to perform a correlation analysis on the time t1 corresponding to the systolic blood pressure in the waveform with the six physiological indicators of age (AGE), mean arterial pressure (MAP), heart rate (HR), cardiac output (CO), systolic blood pressure (SBP), and diastolic blood pressure (DBP). The correlation results are shown in Table 1:

[0056] Table 1 Correlation relationship between the time point t1 corresponding to systolic blood pressure and physiological indicators

[0057]

[0058] * indicates P < 0.05, that is, t1 has a significant correlation with the physiological indicators;

[0059] According to the results of correlation analysis, age (AGE), cardiac output (CO), heart rate (HR), and the time t1 corresponding to the systolic blood pressure in one cardiac cycle are selected to establish a relationship. t1 and t2 are respectively expressed as:

[0060] t1 = 0.001 * AGE - 0.002 * CO - 0.002 * HR + 0.306 (6)

[0061] t2 = tc = 60 / HR (7)

[0062] In the formula, AGE is age, CO is cardiac output, HR is heart rate, t1 represents the time corresponding to the systolic blood pressure, t2 represents the time corresponding to the diastolic blood pressure, and tc is the cardiac cycle;

[0063] Step A5: The commonly used loss function in the field of machine learning is the mean squared error (MSE), and its expression is:

[0064]

[0065] In the formula, y i represents the true value, represents the predicted value, and n is the sample size;

[0066] To solve the problem of a small dataset, we imposed physical constraint conditions on the main feature points of the waveform based on MSE. The expression of the improved loss function is:

[0067] loss = MSE + 0.1 * [SBP – Gauss_function(1)] + [–Gauss_tion()] (9)

[0068] In the formula, MSE represents the mean squared error, SBP is the systolic blood pressure, DBP is the diastolic blood pressure, and Gauss_function(t) represents the Gaussian formula;

[0069] Step A6: Import the training set and test set into the program through the pandas module; train through the neural network model to realize the prediction of Gaussian characteristic parameters, and then realize the non-invasive acquisition of central arterial pressure.

[0070] Step A7: Calculate the normalized mean squared error (NMSE) between the predicted waveform data of central arterial pressure and the waveform data measured clinically to verify the accuracy of the prediction results. The final NMSE result is 0.0342, which meets the application standard of less than 5%.

Claims

1. A method for predicting central arterial pressure based on a meta-learning neural network, characterized in that, It includes the following steps: S1: Obtain the central arterial pressure waveform and physiological indicators of the patient. The physiological indicators include gender, height, weight, body index, age, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, and cardiac output; S2: Perform waveform decomposition based on the Gaussian function superposition method to obtain the Gaussian characteristic parameters a i , b i , c i , where i = 1, 2, 3, 4; S3: Explore the significant relationship between the patient's physiological indicators and Gaussian characteristic parameters, and select six physiological indicators with significant correlations as input data; The six physiological indicators are age AGE, mean arterial pressure MAP, cardiac output CO, systolic blood pressure SBP, diastolic blood pressure DBP, and heart rate HR, S4: Establish a meta-learning data set based on physiological indicators and Gaussian characteristic parameters; S5: Build a neural network, predict Gaussian characteristic parameters based on the meta-learning algorithm, and then obtain the central arterial pressure, and compare and verify the predicted waveform with the clinically measured waveform; The features described in step S2 are as follows: Extract the central arterial pressure waveform of a complete cardiac cycle for each patient, and perform waveform decomposition using a fourth-order Gaussian function based on the Gaussian function superposition method through MATLAB R2021a software to obtain the corresponding 12 Gaussian characteristic parameters. Each decomposed wave respectively characterizes the four periods of the main wave, tidal wave, dicrotic wave, and pre-systolic wave. The formula of the fourth-order Gaussian function can be expressed as: In the formula, a i represents the amplitude of each Gaussian decomposition wave, b i represents the position of each Gaussian decomposition wave, c i represents the width of each Gaussian decomposition wave, and t is time.

2. The method for predicting central arterial pressure based on a meta-learning neural network according to claim 1, characterized in that, The features described in step S4 are as follows: Establish a meta-learning data set based on physiological indicators and Gaussian characteristic parameters: Divide the physiological indicators and Gaussian characteristic parameters of all patients into a training set and a meta-test set according to 4:

1.

3. A method for predicting central arterial pressure based on a meta-learning neural network according to claim 1, characterized in that, The features described in step S5 are as follows: Build a neural network, predict Gaussian characteristic parameters based on the meta-learning algorithm. The specific implementation includes the following steps: Step A1: Establish a fully connected layer and a dropout layer, set the total number of layers of the fully connected layer, the number of neurons in a single fully connected layer, and the activation function; use three fully connected layers and two dropout layers; the input is six variables: age AGE, mean arterial pressure MAP, cardiac output CO, systolic blood pressure SBP, diastolic blood pressure DBP, and heart rate HR, and the output is 12 Gaussian feature parameters (a i , b i , c i , i = 1, 2, 3, 4), Step A2: Select the Adam optimizer and Loss function; The features described in step S5 are as follows: Perform task division according to the MAML algorithm framework. The specific implementation includes the following steps: Step B1: Set hyperparameters including the learning rate α of the base learner, the learning rate β of the meta-learner, and the number of iteration steps epochs; Step B2: Randomly initialize the model parameters θ, and randomly select a part of the samples from task T i to calculate the loss through the constructed network, update the gradient, and calculate the optimal parameter θ', that is In the formula, θ′ i is the optimal parameter of task T i , θ is the initial parameter, α is the learning rate of the base learner, is the gradient of task T i ; By calculating the gradient with respect to the optimal parameter θ' of the previous step, update the randomly initialized model parameter θ, that is In the formula, β is the learning rate of the meta-learner, is the gradient of each new task Ti with respect to the optimal parameter θ'.

4. A method for predicting central arterial pressure based on a meta-learning neural network according to claim 1, characterized in that, Step S5 includes the following steps: Step C1: Use the six significant physiological parameters selected through correlation coefficient matrix analysis as the input data of the model; Step C2: Import the training set, validation set, and test set into the program through the pandas module; Step C3: The inputs are six variables, namely age AGE, mean arterial pressure MAP, cardiac output CO, systolic blood pressure SBP, diastolic blood pressure DBP, and heart rate HR, and the outputs are 12 Gaussian feature parameters (a i , b i , c i , i = 1, 2, 3, 4). Four fully connected layers are used, and two dropout layers are used to prevent the model from overfitting during the training process. The ReLU function is used as the activation function Step C4: Select the MAML algorithm to build the model framework, set the hyperparameters meta-learning learner rate α = 0.005, base learner learning rate β = 0.1, and the iteration step epochs to 1500 - 4000, and use the Adam optimizer for training and optimization; Step C5: Combine the main characteristic points of the central arterial pressure waveform, namely systolic blood pressure and its corresponding time t1, diastolic blood pressure and its corresponding time t2, and the pre-training results. Use SPSS 26.0 software to perform a correlation analysis between systolic blood pressure and its corresponding time t1 and physiological indicators. The physiological indicators are age AGE, mean arterial pressure MAP, heart rate HR, cardiac output CO, systolic blood pressure SBP, and diastolic blood pressure DBP; Select the significantly correlated physiological indicators AGE, CO, and HR to establish a relationship with t1. Express the time t1 corresponding to the systolic blood pressure and the time t2 corresponding to the diastolic blood pressure in one cardiac cycle as follows: t1 = 0.001 * AGE - 0.002 * CO - 0.002 * HR + 0.306 (5) t2 = tc = 60 / HR (6) In the formula, AGE is age, CO is cardiac output, HR is heart rate, t1 represents the time corresponding to systolic blood pressure, t2 represents the time corresponding to diastolic blood pressure, and tc is the cardiac cycle; The commonly used loss function in the field of machine learning is the mean square error MSE, and its expression is: In the formula, y i represents the true value, represents the predicted value, and n is the sample size; Based on MSE, physical constraint conditions are imposed on the main characteristic points of the waveform. The expression of the improved loss function is: loss = MSE + 0.1 * [SBP – Gauss_function(t1)] + [DBP – Gauss_function(t2)] (8) In the formula, MSE represents the mean square error, SBP is systolic blood pressure, DBP is diastolic blood pressure, and Gauss_function(t) represents the Gaussian formula; Step B6: Calculate the normalized mean square error (NMSE) between the predicted central arterial pressure waveform and the clinically measured central arterial pressure waveform. NMSE < 5% is regarded as a small difference between the predicted result and the clinical result.

Images