A blood glucose concentration prediction optimization method

Abstract

The present invention proposes a blood glucose concentration prediction optimization method, which relates to the technical field of data prediction; the method comprises: S1, constructing an original data set for blood glucose concentration prediction, performing a unit root test on the original data set, and if the original data is stable, no processing is performed, otherwise, first-order difference processing is performed on the non-stationary original data; a first data set is obtained; S2, normalizing the first data set, and smoothing the normalized eigenvalues ​​by correlation coefficients to obtain an input data set for blood glucose concentration prediction; S3, converting the input data set into a supervised input data set consisting of input prediction points and output prediction values ​​according to a time series; S4, inputting the supervised input data set into a blood glucose prediction model based on PECO-BP, and outputting a real-time prediction value of blood glucose concentration. The blood glucose prediction model based on PECO-BP proposed by the present invention can improve the prediction accuracy and prediction sensitivity of blood glucose concentration.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a method for optimizing blood sugar concentration prediction. Background Art

[0002] Currently in the market and scientific research, blood glucose prediction models are mainly divided into physiological characteristic blood glucose prediction models and data-driven blood glucose prediction models. Based on existing research, physiological models that require a large number of parameters are difficult to meet the needs in blood glucose prediction. With the development of blood glucose sensors, data-driven blood glucose prediction models have been further developed, and the prediction accuracy is relatively high. However, since blood glucose concentration is a dynamic regulation quantity and is easily affected by external factors, blood glucose prediction models need to be continuously improved.

[0003] Alia et al. proposed a blood glucose prediction method based on artificial neural network, which has better prediction effect than support vector regression, autoregression and extreme learning machine; Publication No. CN109935333A discloses a blood glucose prediction method based on OVMD-SE-PSO-BP, which uses a method combining optimal variational mode decomposition with sample entropy to decompose the acquired original blood glucose sequence into multiple subsequences, and establishes a particle swarm optimized BP neural network prediction model for each subsequence. This method can solve the problem that the back propagation (BP) neural network is prone to fall into the local minimum and the defect that the network structure is difficult to determine to a certain extent. However, due to the defects of the PSO particle swarm algorithm itself, the problem that the back propagation (BP) neural network is prone to fall into the local minimum cannot be completely solved.

[0004] The Educational Competitive Optimization (ECO) algorithm establishes a mathematical model for optimization through a gradual optimization mechanism in the primary school, junior high school, and high school stages, which can better balance the exploration and development processes. Compared with algorithms such as PSO, ECO dynamically adjusts the search strategy according to different stages: the primary school stage mainly focuses on global search to help find potential solutions; the junior high school stage strikes a balance between global search and local development; the high school stage strengthens local development and focuses on detailed optimization of the better solutions that have been found. This multi-stage dynamic search mechanism can enable ECO to avoid falling into local optimal solutions too early in the early stages, and at the same time make full use of development capabilities in the later stages to improve optimization accuracy. In contrast, the inertia weight and acceleration coefficient of PSO are relatively fixed throughout the search process and are difficult to adjust flexibly. However, this multi-stage dynamic search mechanism mainly relies on adaptive adjustment based on changes in the number of iterations, and cannot adjust the performance of the BP neural network prediction model according to dynamic changes in blood glucose concentration. In complex situations, it cannot avoid falling into local optimal solutions too early, affecting the prediction accuracy of the BP neural network prediction model. Summary of the invention

[0005] In view of the problems existing in the above-mentioned background technology, a blood glucose concentration prediction model is proposed. By improving the educational competitive optimization algorithm (ECO), the problem that the prediction model is prone to fall into the local minimum when the educational competitive optimization algorithm (ECO) optimizes the BP neural network prediction model for dynamic blood glucose concentration prediction cannot be solved. The improved educational competitive optimization algorithm (PECO) is combined with the BP neural network prediction model to obtain a blood glucose prediction model based on PECO-BP, thereby improving the prediction accuracy and prediction sensitivity of the present invention for blood glucose concentration.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a method for predicting and optimizing blood glucose concentration, which specifically comprises the following steps:

[0007] S1. constructing an original data set for predicting blood glucose concentration, performing a unit root test on the original data set, and if the original data is stable, no processing is performed; otherwise, performing first-order difference processing on the non-stationary original data; obtaining a first data set;

[0008] S2, normalizing the first data set, and smoothing the normalized first data set using a correlation coefficient to obtain an input data set for blood glucose concentration prediction;

[0009] S3, converting the input data set into a time series consisting of input prediction points And output predicted value A supervised input data set composed of a training set and a test set;

[0010] S4, inputting the supervised input data set into the blood glucose prediction model based on PECO-BP, and outputting the real-time prediction value of blood glucose concentration; the PECO-BP is to optimize the learning rate factor and regularization parameter of the BP neural network prediction model by using the improved educational competition optimization algorithm to obtain the optimal learning rate factor and regularization parameter, and the optimal learning rate factor and the optimal regularization parameter Used in BP neural network prediction model; among them, the improved educational competitive optimization algorithm (PECO) includes:

[0011] S41, according to the convergence rate and individual students With school individuals The difference between Improvement of the step size factor of the position update mathematical model in primary school ;

[0012] S42. According to each student The current fitness value and the population individual Distance to global optimal solution Dynamic improvement Endurance factor of educational competition optimization algorithm .

[0013] Specifically, the original data set for blood glucose concentration prediction contains the time series of blood glucose concentration (mg / dL) and N1 related variables, where the time series data of the related variables include: heart rate (bpm), body temperature (°C), food intake (calories), and blood pressure (mmHg). The original data set is a time series data set, in which each row represents a time point and each column represents a variable.

[0014] Specifically, by calculating the change between consecutive time points, the trend in the original data set can be eliminated, so that the time series data tends to be stable, and the first data set is obtained; the first data set is tested for stability, and the unit root test is used to determine the stability of the time series data and whether there is a unit root. The specific steps are: input the time series data in the first data set into the ADF test. If the time series data has a unit root, it is non-stationary; otherwise, the time series does not have a unit root, that is, it is stable; the non-stationary first data set is processed by first-order difference, and the mathematical model is:

[0015] (1);

[0016] In the formula, It is the time series value after the first-order difference, indicating the change at the t-th time point, reflecting the change between the current time point t and the previous time point t-1; is the value of the time series data in the first data set at time point t, indicating the value at the current moment; is the value of the time series data in the first data set at time point t-1, indicating the value at the previous moment.

[0017] Specifically, the first data set is scaled to the interval of (0.1, 0.9), and after normalization, the correlation coefficient matrix between the related variables is calculated using the Pearson correlation coefficient formula, and the related variables with a correlation coefficient greater than 0.5 are smoothed using moving average. Among them, the related variables with a correlation coefficient greater than 0.5 have a high correlation with the blood glucose concentration, and the corresponding normalization processing method is:

[0018] (2);

[0019] In the formula, is the data value of the time series of the nth related variable at the tth time point after normalization, is the data value at the tth time point of the time series of the nth related variable before normalization, is the data value at the tth time point of the time series of the nth related variable before normalization, is the maximum data value of the time series of the nth related variable, is the minimum data value of the time series of the nth dependent variable.

[0020] Specifically, in the prediction of blood glucose concentration, the data of the input data set at n time points in the previous period are selected as the input prediction points, the blood glucose concentration at the current time point is used as the output prediction value, the input prediction points and the output prediction values ​​are combined to obtain a supervised input data set, and the supervised input data set is divided into a training set and a test set in a ratio of 6:4; the training set is used to train the blood glucose prediction model, and the test set is used to predict the blood glucose concentration, reflecting the blood glucose concentration prediction effect.

[0021] Specifically, the blood glucose prediction model based on PECO-BP includes: first, constructing a BP neural network prediction model, including setting the number of input layers, hidden layers, and output layers according to the number of features of the supervised input data set, the output layer represents the output predicted blood glucose concentration, and using the mean square error (MSE) as the loss function to measure the error between the predicted blood glucose concentration and the actual value; second, using the improved educational competitive optimization algorithm to optimize the learning rate factor and regularization parameter of the BP neural network prediction model to obtain the blood glucose prediction model based on PECO-BP.

[0022] Specifically, a supervised input data set is input into a PECO-BP-based blood glucose prediction model, and a real-time predicted value of blood glucose concentration is output, including: the supervised input data enters the PECO-BP blood glucose prediction model from the input layer, and is transmitted to the output layer through the hidden layer to obtain a real-time predicted value of blood glucose concentration; further, the real-time predicted value of blood glucose concentration is compared with the actual value, and the error is calculated through a loss function. If the error between the two meets the set accuracy requirement, the training process is terminated; if the requirement is not met, the error is transmitted from the output layer through the hidden layer to the input layer through back propagation, and the weights of the blood glucose prediction model are continuously adjusted to gradually reduce the sum of square errors of the blood glucose prediction model until the set accuracy requirement is met; finally, the test set data is used to predict blood glucose concentration using the trained PECO-BP-based blood glucose prediction model.

[0023] Specifically, the weight of the blood glucose prediction model is continuously adjusted after each training stage, specifically: the learning rate factor and the regularization parameter value are adjusted using the improved educational competition optimization algorithm, and after the optimal learning rate factor and the regularization parameter value are obtained, the optimal learning rate factor and the regularization parameter value are reused in the blood glucose prediction model based on PECO-BP, wherein the loss function is adjusted using the optimal regularization parameter value, the error value is calculated, and the weight of the blood glucose prediction model is adjusted using the optimal learning rate factor, and the above steps are repeated until the error value changes less than 0.0001; wherein the mathematical model of the optimal learning rate factor adjusting the weight of the blood glucose prediction model is:

[0024] (3);

[0025] In the formula, is the weight of the adjusted blood glucose prediction model, is the weight of the blood glucose prediction model before adjustment, is the optimal learning rate factor, is the gradient of the loss function with respect to the weight;

[0026] The mathematical model of adjusting the loss function using the optimal regularization parameter value is:

[0027] (4);

[0028] In the formula, is the loss function, M is the input data set size for blood glucose concentration prediction, y is the actual blood glucose concentration, is the blood glucose concentration at the tth time point, is the optimal regularization parameter value, Z is the weight scale of the blood glucose prediction model, is the square of the current z-th weight.

[0029] Specifically, the improved educational competitive optimization algorithm (PECO) includes the following improvements: and individual students With school individuals The difference between Improvement of the step size factor of the position update mathematical model in primary school ;

[0030] Among them, the convergence rate The convergence rate of the learning rate factor and regularization parameter optimization of the BP neural network prediction model is optimized by using the current educational competition optimization algorithm population diversity quantitative educational competition optimization algorithm (ECO). The mathematical model is:

[0031] (3);

[0032] Where iter is the current iteration number, R(iter) is the convergence rate value of the iter-th iteration, and H max is the maximum population diversity value of the iter-th iteration; H(iter) is the population diversity value of the iter-th iteration, and the mathematical model is:

[0033] (4);

[0034] In the formula, is the optimal fitness value of the i-th individual position in the population at the iter-th iteration; N is the maximum population size, where the individual positions of the population include the individual positions of students and the individual positions of schools;

[0035] Among them, the difference between the individual position of students and the individual position of schools in the primary school stage of ECO is calculated, and the mathematical model is:

[0036] (5);

[0037] Where D is the problem dimension, and its value is the number of learning rate factors and regularization parameter values. It is the iterth iteration The position value of each student on the jth dimension, is the position value of the k-th school individual on the j-th dimension in the iter-th iteration, j=1,2,3,..,D; is the iterth iteration The individual student and The difference between the individual schools; ; , The maximum size for individual students, The largest individual school size. + =N;

[0038] Among them, the improved step factor The mathematical model is:

[0039] (6);

[0040] In the formula, is the improved step factor value of the i-th individual in the iter-th iteration, wherein the individuals include student individuals and school individuals, and the maximum value is N; is the maximum value of the improved step size factor, which is 1, and μ is the amplitude value of the adjustment difference.

[0041] Specifically, in the primary school stage of the ECO algorithm, when the convergence of the learning rate factor and regularization parameter optimization of the BP neural network prediction model for blood glucose concentration prediction by the educational competitive optimization algorithm (ECO) is slow, that is, R(iter) is small, and the difference between the individual position of students and the school position is large, that is, When the value is larger, a larger step size should be used to speed up the exploration of the ECO algorithm. When the convergence is faster and the individual position is close to the optimal school position, the step size should be reduced to perform a more refined local search and improve the accuracy of the parameter solution.

[0042] Specifically, the improved educational competition optimization algorithm (PECO) includes the following improvements: dynamically improving the endurance factor of the educational competition optimization algorithm based on the distance between the current fitness value of each student in the junior high school stage of the educational competition optimization algorithm and the global optimal solution. , the improved endurance factor mathematical model is:

[0043] (7);

[0044] In the formula, is the iterth iteration The fitness value of each student is is the global optimal fitness value of the population individuals, is the iterth iteration The endurance factor of each student improved, To adjust the parameters, control the steepness of the endurance factor.

[0045] Specifically, the improved endurance factor is used to improve the position updating mathematical model of the educational competition optimization algorithm in the high school stage. When the individual position of the student is far from the global optimal solution, a higher patience value is maintained to enhance the global exploration speed of the educational competition optimization algorithm for the learning rate factor and regularization parameter optimization of the BP neural network prediction model; when the individual student is close to the global optimal solution, the patience value is reduced to enhance the development accuracy of the learning rate factor and regularization parameter of the BP neural network prediction model by the educational competition optimization algorithm, and improve the accuracy of the learning rate factor and regularization parameter; wherein, the global optimal solution is the individual position of the population corresponding to the global minimum fitness value.

[0046] Specifically, the improved educational competition optimization algorithm is used to optimize the learning rate factor and regularization parameter of the BP neural network prediction model, wherein the learning rate factor is a one-dimensional parameter, which represents the step size of the BP neural network prediction model at each weight update, and the regularization parameter is a one-dimensional parameter, which is used to control the regularization strength of the complexity of the BP neural network prediction model. Therefore, the problem dimension D of the improved educational competition optimization algorithm is 2. Further, the learning rate factor and the regularization parameter are encoded into a 2-dimensional space vector U. , where the space vector U establishes a real number mapping relationship with the individual positions of students and schools in the improved educational competition optimization algorithm. , where the total number of individual students and individual schools is N, and The position values ​​of the first dimension and the second dimension of the population individual positions of the improved educational competition optimization algorithm are respectively improved, and the learning rate factor and the regularization parameter value of the BP neural network prediction model are updated by updating the student individual and school individual positions of the improved educational competition optimization algorithm.

[0047] Specifically, the mathematical model of the improved educational competition optimization algorithm is:

[0048] S341. Calculate the improved step size factor using formula (6) Iter iteration ; Using the Improve the mathematical model of population individual position update in the primary school stage; when the i-th population individual is a school individual, execute i=k, otherwise execute , the mathematical model is:

[0049] (8);

[0050] In the formula, is the position value of the i-th population individual on the j-th dimension in the iter-th iteration, It is the iterth iteration The position value of each student on the jth dimension, is the position value of the individual in the kth school on the jth dimension in the iterth iteration, is the average position of the individual schools in the iterth iteration, Levy(D) is the Levy distribution of the problem dimension D, and d is the The nearest school location of each student, randn is a random variable that follows a normal distribution. It is the position value of the individual in the i-th population on the j-th dimension in the iter+1-th iteration;

[0051] S342. Calculate the improved endurance factor value using formula (7), and use the improved endurance factor value to improve the learning motivation factor E. The mathematical model of the learning motivation factor E is:

[0052] (9);

[0053] In the formula, is the learning motivation factor for the iter-th iteration, is the iterth iteration The endurance factor of each student's individual improvement, T max is the maximum number of iterations;

[0054] S343, based on S342, the mathematical model for updating the position of individuals in the population at the middle school stage is improved, and the mathematical model is:

[0055] (10);

[0056] In the formula, , , , , , Levy(D), d, T max、 , , Explanation is the same as S341 and S342; is the best individual position of the population in the iter iteration, R1 is a random number between 0 and 1, and H is the judgment threshold, which is set to 0.5;

[0057] S344. Based on the improved endurance factor, the mathematical model for updating the position of individuals in the population at the high school stage is improved, and the mathematical model is:

[0058] (11);

[0059] Wherein, the parameter interpretation is the same as S343, and R2 is a random number between 0 and 1.

[0060] Specifically, the improved educational competition optimization algorithm is used to optimize the learning rate factor and regularization parameter of the BP neural network prediction model. The method is:

[0061] S31, establishing a real number mapping relationship between the space vector U of the learning rate factor and the regularization parameter and the individual student positions and the individual school positions of the improved educational competition optimization algorithm;

[0062] S32, initialize the problem dimension D of the improved educational competition optimization algorithm, the maximum population size N, and the maximum number of iterations T max , improved step size factor , improved endurance factor , the minimum value of the fitness function, the position of the improved educational competition optimization algorithm initialized by chaotic mapping;

[0063] S33, using the fitness function to calculate the fitness values ​​of the population individuals in the initialization phase, arranging the fitness values ​​in descending order, and the population individuals corresponding to the first 50% of the fitness values ​​are used as school individuals, and the rest are used as student individuals;

[0064] S34, determine whether the current iteration number iter satisfies iter%3=1, if so, establish and execute the improved educational competition optimization algorithm for primary school stage position update mathematical model; otherwise, if iter%3=2, establish and execute the improved educational competition optimization algorithm for junior high school stage position update mathematical model, otherwise establish and execute the improved educational competition optimization algorithm for high school stage position update mathematical model;

[0065] S35, if the fitness value of the current i-th population individual belongs to the top 50%, the school position update mathematical model is executed; otherwise, the student position update mathematical model is executed to obtain the population individual position value of the improved education competition optimization algorithm;

[0066] S36, using the fitness function to calculate the current population individual fitness value, retaining the population individual position value corresponding to the minimum fitness value, which is the best population individual position of the current iteration , and the best individual position of the last iteration , retain the smaller position value of the two, which is the global optimal population individual position value ;

[0067] S37, the current iteration number executes iter=iter+1, and determines whether the current iteration iter is greater than the maximum iteration number T max If yes, the optimization process ends and the global optimal population individual position value is output , after parsing, assign it to the space vector U of the learning rate factor and regularization parameter; otherwise, return to execute S33.

[0068] Specifically, the fitness function value reflects the robustness of the current learning rate factor and regularization parameter solution. The robustness is the effect of the BP neural network prediction model network on the adjustment of the loss function and weights under the current learning rate factor and regularization parameter solution, as well as the accuracy of the prediction of blood glucose solubility. The smaller the fitness value, the more accurate the blood glucose prediction model's prediction of blood glucose solubility.

[0069] Compared with the prior art, the blood glucose concentration prediction optimization method proposed in the present invention has the following technical effects: the present invention improves the educational competitive optimization algorithm (ECO) to obtain an improved educational competitive optimization algorithm (PECO), and combines it with the BP neural network prediction model to form a blood glucose prediction model based on PECO-BP, thereby solving the problem that the BP neural network prediction model is prone to fall into the local minimum when predicting blood glucose, resulting in large prediction deviations, and has better prediction effects. Secondly, through the dynamic adjustment mechanism of the PECO algorithm, the present invention optimizes the search strategy in different stages of ECO, so that the ECO algorithm achieves a better balance between global search and local development. This multi-stage dynamic search mechanism not only avoids the algorithm from falling into the local optimal solution too early, but also improves the prediction sensitivity of the PECO-BP prediction model under complex conditions. In summary, the method of the present invention makes the entire blood glucose concentration prediction process more efficient and stable. This efficient and stable prediction method has important value in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart of the blood glucose concentration prediction optimization method.

[0071] Figure 2 It is the fitness graph for optimizing the learning rate factor and regularization parameter of the blood glucose concentration prediction model.

[0072] Figure 3 This is a graph showing the prediction results of the PECO-BP prediction model and the ECO-BP prediction model for the blood glucose prediction data of the training set.

[0073] Figure 4This is a graph showing the prediction results of the PECO-BP prediction model and the ECO-BP prediction model for the blood glucose prediction data of the test set.

[0074] Figure 5 This is a comparison chart of the errors in predicting blood glucose concentration between the PECO-BP prediction model and the ECO-BP prediction model. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0076] The present invention provides a technical solution: by improving the educational competitive optimization algorithm (ECO), the problem that the prediction model of the BP neural network prediction model for dynamic blood glucose concentration prediction cannot be solved when the educational competitive optimization algorithm (ECO) optimizes the prediction model of the dynamic blood glucose concentration prediction, the improved educational competitive optimization algorithm (PECO) is combined with the BP neural network prediction model to obtain a blood glucose prediction model based on PECO-BP, and the prediction accuracy of blood glucose concentration is improved. The specific steps are as follows: Figure 1 As shown, it includes S1 to S4.

[0077] S1. Construct the original data set for blood glucose concentration prediction, such as Figure 2 As shown, a unit root test is performed on the original data set. If the original data is stable, no processing is performed. Otherwise, first-order difference processing is performed on the non-stationary original data; and a first data set is obtained.

[0078] Furthermore, for the original data set for blood glucose concentration prediction: blood glucose concentration (mg / dL) and the time series of N1 related variables, the time series data of the related variables include: heart rate (bpm), body temperature (℃), food intake (calories), blood pressure (mmHg); calculate the change between consecutive time points of the original data set, use the average value of the related variables to supplement the missing values, and obtain a new data set containing the change between consecutive time points of the original data set, recorded as the first data set; perform a stationarity test on the first data set, the specific steps are: input the time series data in the first data set into the ADF test, if the time series data has a unit root, it is non-stationary; otherwise, the time series does not have a unit root, it is stationary; perform first-order difference processing on the non-stationary first data set, and the mathematical model is:

[0079] (1);

[0080] In the formula, It is the time series value after the first-order difference, indicating the change at the t-th time point, reflecting the change between the current time point t and the previous time point t-1; is the value of the time series data in the first data at time point t, indicating the value at the current moment; It is the value of the time series data in the first data at time point t-1, indicating the value at the previous moment.

[0081] S2. Normalize the first data set, and smooth the normalized first data set using a correlation coefficient to obtain an input data set for blood glucose concentration prediction.

[0082] Furthermore, the first data set is scaled to the interval of (0.1, 0.9), and after normalization, the correlation coefficient matrix between the related variables is calculated using the Pearson correlation coefficient formula, and the related variables with correlation coefficients greater than 0.5 are smoothed using moving average. Among them, the related variables with correlation coefficients greater than 0.5 have a high correlation with blood glucose concentration, and the corresponding normalization processing method is:

[0083] (2);

[0084] In the formula, is the data value of the time series of the nth related variable at the tth time point after normalization, is the data value at the tth time point of the time series of the nth related variable before normalization, is the data value at the tth time point of the time series of the nth related variable before normalization, is the maximum data value of the time series of the nth related variable, is the minimum data value of the time series of the nth dependent variable.

[0085] Furthermore, the normalization code is completed in MATLAB as follows:

[0086] [p_train, ps_input] = mapminmax(P_train, 0.1,0.9);

[0087] p_test = mapminmax('apply', P_test, ps_input);

[0088] [t_train, ps_output] = mapminmax(T_train, 0.1, 0.9);

[0089] t_test = mapminmax('apply', T_test, ps_output).

[0090] S3, converting the input data set into a time series consisting of input prediction points And output predicted value The supervised input dataset comprises a training set and a test set.

[0091] Furthermore, data from the input data set at n time points in the previous period are selected as input prediction points, and the blood glucose concentration at the current time point is used as the output prediction value. The input prediction points and the output prediction values ​​are combined to obtain a supervised input data set, and the supervised input data set is divided into a training set and a test set in a ratio of 6:4; the training set is used to train the blood glucose prediction model, and the test set is used to predict the blood glucose concentration, reflecting the blood glucose concentration prediction effect.

[0092] Furthermore, we create input-output pairs in MATLAB to generate supervised input data sets. The code is as follows:

[0093] X = [];

[0094] Y = [];

[0095] for i = 1:(data_length - window_size - horizon + 1)

[0096] X = [X; data(i:i+window_size-1)];

[0097] Y = [Y; data(i+window_size:i+window_size+horizon-1)];

[0098] end .

[0099] Furthermore, in MATLAB, the code for dividing the supervised input data set into training set and test set in a ratio of 6:4 is as follows:

[0100] First, determine the size of the training and test sets.

[0101] num_train = round(train_ratio * size(X, 1));

[0102] num_test = size(X, 1) - num_train;

[0103] Secondly, divide the training set and test set

[0104] P_train = res(1: num_train, 1: f_)';

[0105] T_train = res(1: num_train, f_ + 1: end)';

[0106] M = size(P_train, 2);

[0107] P_test = res(num_train + 1: end, 1: f_)';

[0108] T_test = res(num_train + 1: end, f_ + 1: end)';

[0109] N = size(P_test, 2).

[0110] S4, inputting the supervised input data set into the blood glucose prediction model based on PECO-BP, and outputting the real-time prediction value of blood glucose concentration; the PECO-BP is to optimize the learning rate factor and regularization parameter of the BP neural network prediction model by using the improved educational competition optimization algorithm to obtain the optimal learning rate factor and regularization parameter, and the optimal learning rate factor and the optimal regularization parameter Used in BP neural network prediction model.

[0111] Furthermore, the supervised input data set is input into the PECO-BP-based blood glucose prediction model to output a real-time predicted value of blood glucose concentration, including: the supervised input data enters the PECO-BP blood glucose prediction model from the input layer, is transmitted to the output layer through the hidden layer, and obtains the real-time predicted value of blood glucose concentration; further, the real-time predicted value of blood glucose concentration is compared with the actual value, and the error is calculated through the loss function. If the error between the two meets the set accuracy requirement, the training process is terminated; if the requirement is not met, the error is transmitted from the output layer through the hidden layer to the input layer through back propagation, and the weights of the blood glucose prediction model are continuously adjusted to gradually reduce the sum of square errors of the blood glucose prediction model until the set accuracy requirement is met; finally, the test set data is used to predict blood glucose concentration using the trained PECO-BP-based blood glucose prediction model.

[0112] Furthermore, the weights of the blood glucose prediction model are continuously adjusted after each training stage, specifically: the learning rate factor and the regularization parameter value are adjusted using the improved educational competition optimization algorithm, and after the optimal learning rate factor and the regularization parameter value are obtained, the optimal learning rate factor and the regularization parameter value are reused in the blood glucose prediction model based on PECO-BP, wherein the loss function is adjusted using the optimal regularization parameter value, the error value is calculated, and the weights of the blood glucose prediction model are adjusted using the optimal learning rate factor, and the above steps are repeated until the error value changes less than 0.0001; wherein the mathematical model of the optimal learning rate factor adjusting the weights of the blood glucose prediction model is:

[0113] (3);

[0114] In the formula, is the weight of the adjusted blood glucose prediction model, is the weight of the blood glucose prediction model before adjustment, is the optimal learning rate factor, is the gradient of the loss function with respect to the weight;

[0115] The mathematical model of adjusting the loss function using the optimal regularization parameter value is:

[0116] (4);

[0117] In the formula, is the loss function, M is the input data set size for blood glucose concentration prediction, y is the actual blood glucose concentration, is the blood glucose concentration at the tth time point, is the optimal regularization parameter value, Z is the weight scale of the blood glucose prediction model, is the square of the current z-th weight.

[0118] Furthermore, according to the convergence rate and individual students With school individuals The difference between Improvement of the step size factor of the position update mathematical model in primary school ,Improved Educational Competitive Optimization Algorithm (PECO).

[0119] Furthermore, the convergence rate The convergence rate of the learning rate factor and regularization parameter optimization of the BP neural network prediction model is optimized by using the current educational competition optimization algorithm population diversity quantitative educational competition optimization algorithm (ECO). The mathematical model is:

[0120] (3);

[0121] Where iter is the current iteration number, R(iter) is the convergence rate value of the iter-th iteration, and H max is the maximum population diversity value of the iter-th iteration; H(iter) is the population diversity value of the iter-th iteration, and the mathematical model is:

[0122] (4);

[0123] In the formula, is the optimal fitness value of the i-th individual position of the population at the iter-th iteration; N is the maximum population size, where the individual positions of the population include the individual positions of students and the individual positions of schools;

[0124] Among them, the difference between the individual position of students and the individual position of schools in the primary school stage of ECO is calculated, and the mathematical model is:

[0125] (5);

[0126] Where D is the problem dimension, and its value is the number of learning rate factors and regularization parameter values. It is the iterth iteration The position value of each student on the jth dimension, is the position value of the k-th school individual on the j-th dimension in the iter-th iteration, j=1,2,3,..,D; is the iterth iteration The individual student and The difference between the individual schools; ; , The maximum size for individual students, The largest individual school size. + =N;

[0127] Among them, the improved step factor The mathematical model is:

[0128] (6);

[0129] In the formula, is the improved step factor value of the i-th individual in the iter-th iteration, wherein the individuals include student individuals and school individuals, and the maximum value is N; is the maximum value of the improved step size factor, which is 1, and μ is the amplitude value of the adjustment difference.

[0130] Furthermore, according to each student The current fitness value and the population individual Distance to global optimal solution Dynamic improvement Endurance factor of educational competition optimization algorithm ; The improved mathematical model of endurance factor is:

[0131] (7);

[0132] In the formula, is the iterth iteration The fitness value of each student is is the global optimal fitness value of the population individuals, is the iterth iteration The endurance factor of each student improved, To adjust the parameters, control the steepness of the endurance factor.

[0133] Furthermore, the improved educational competition optimization algorithm is used to optimize the learning rate factor and regularization parameter of the BP neural network prediction model, wherein the learning rate factor is a one-dimensional parameter, which represents the step size of the BP neural network prediction model at each weight update, and the regularization parameter is a one-dimensional parameter, which is used to control the regularization strength of the complexity of the BP neural network prediction model. Therefore, the problem dimension D of the improved educational competition optimization algorithm is 2. Furthermore, the learning rate factor and the regularization parameter are encoded into a 2-dimensional space vector U. , where the space vector U establishes a real number mapping relationship with the individual positions of students and schools in the improved educational competition optimization algorithm. , where the total number of individual students and individual schools is N, and The position values ​​of the first dimension and the second dimension of the population individual positions of the improved educational competition optimization algorithm are respectively improved, and the learning rate factor and the regularization parameter value of the BP neural network prediction model are updated by updating the student individual and school individual positions of the improved educational competition optimization algorithm.

[0134] Furthermore, the mathematical model of the improved educational competition optimization algorithm is established as follows:

[0135] S341. Calculate the improved step size factor using formula (6) Iter iteration ; Using the Improve the mathematical model of population individual position update in the primary school stage; when the i-th population individual is a school individual, execute i=k, otherwise execute , the mathematical model is:

[0136] (8);

[0137] In the formula, is the position value of the i-th population individual on the j-th dimension in the iter-th iteration, It is the iterth iteration The position value of each student on the jth dimension, is the position value of the individual in the kth school on the jth dimension in the iterth iteration, is the average position of the individual schools in the iterth iteration, Levy(D) is the Levy distribution of the problem dimension D, and d is the The nearest school location of each student, randn is a random variable that follows a normal distribution. It is the position value of the individual in the i-th population on the j-th dimension in the iter+1-th iteration;

[0138] S342. Calculate the improved endurance factor value using formula (7), and use the improved endurance factor value to improve the learning motivation factor E. The mathematical model of the learning motivation factor E is:

[0139] (9);

[0140] In the formula, is the learning motivation factor for the iter-th iteration, is the iterth iteration The endurance factor of each student's individual improvement, T max is the maximum number of iterations;

[0141] S343, based on S342, the mathematical model for updating the position of individuals in the population at the middle school stage is improved, and the mathematical model is:

[0142] (10);

[0143] In the formula, , , , , , Levy(D), d, T max、 , , Explanation is the same as S341 and S342; is the best individual position of the population in the iter iteration, R1 is a random number between 0 and 1, and H is the judgment threshold, which is set to 0.5;

[0144] S344. Based on the improved endurance factor, the mathematical model for updating the position of individuals in the population at the high school stage is improved, and the mathematical model is:

[0145] (11);

[0146] Wherein, the parameter interpretation is the same as S343, and R2 is a random number between 0 and 1.

[0147] Furthermore, the improved educational competition optimization algorithm is used to optimize the learning rate factor and regularization parameter of the BP neural network prediction model. The method is as follows:

[0148] S31, establishing a real number mapping relationship between the space vector U of the learning rate factor and the regularization parameter and the individual student positions and the individual school positions of the improved educational competition optimization algorithm;

[0149] S32, initialize the problem dimension D of the improved educational competition optimization algorithm, the upper bound ub and lower bound lb of the individual positions of the population, the maximum population size N, and the maximum number of iterations T max , improved step size factor Maximum, improved endurance factor Initial value, minimum value of fitness function value, position of initializing improved educational competition optimization algorithm through chaotic mapping;

[0150] S33, using the fitness function to calculate the fitness values ​​of the population individuals in the initialization phase, arranging the fitness values ​​in descending order, and the population individuals corresponding to the first 50% of the fitness values ​​are used as school individuals, and the rest are used as student individuals;

[0151] S34, determine whether the current iteration number iter satisfies iter%3=1, if so, establish and execute the improved educational competition optimization algorithm for primary school stage position update mathematical model; otherwise, if iter%3=2, establish and execute the improved educational competition optimization algorithm for junior high school stage position update mathematical model, otherwise establish and execute the improved educational competition optimization algorithm for high school stage position update mathematical model;

[0152] S35, if the fitness value of the current i-th population individual belongs to the top 50%, the school position update mathematical model is executed; otherwise, the student position update mathematical model is executed to obtain the population individual position value of the improved education competition optimization algorithm;

[0153] S36, using the fitness function to calculate the current population individual fitness value, retaining the population individual position value corresponding to the minimum fitness value, which is the best population individual position of the current iteration , and the best individual position of the last iteration , retain the smaller position value of the two, which is the global optimal population individual position value ;

[0154] S37, the current iteration number executes iter=iter+1, and determines whether the current iteration iter is greater than the maximum iteration number T max If yes, the optimization process ends and the global optimal population individual position value is output , after parsing, assign it to the space vector U of the learning rate factor and regularization parameter; otherwise, return to execute S33.

[0155] Furthermore, the above mathematical model is implemented in MATLAB code, the fitness function of the improved educational competition optimization algorithm is initialized, and the problem dimension D=2, the maximum population size N=100, and the maximum number of iterations T are set. max =80, the lower bound of the individual position of the population is lb=[6 1e-4] and the upper bound is ub=[32 1e-3]; =1, =0.1, the code is as follows:

[0156] fun = @fobj;

[0157] lb = [6 1e-4];

[0158] ub = [32 1e-3];

[0159] D = length(ub);

[0160] SearchAgents_no = 100;

[0161] Max_iteration = 80;

[0162] smax=1;

[0163] p=0.1.

[0164] Furthermore, a BP neural network prediction model is constructed. The code is as follows:

[0165] net = newff(p_train, t_train, best_hd);

[0166] Set the training parameters, the code is as follows

[0167] net.trainParam.epochs = 1000; Number of training times;

[0168] net.trainParam.goal = 1e-6; target error value;

[0169] net.trainParam.lr = best_lr; learning rate;

[0170] net.trainParam.hd=best_hd ; Regularization parameter.

[0171] Furthermore, complete formula (8) to formula (11) in MATLAB, call the improved educational competitive optimization algorithm (PECO) and the standard educational competitive optimization algorithm (ECO), map the population individual positions to the space vectors of the learning rate factor and the regularization parameter, and iteratively optimize the learning rate factor and the regularization parameter. The code is as follows:

[0172] [Best_score, Best_pos, curve] = ECO(SearchAgents_no, Max_iteration,lb, ub, dim, fun);

[0173] [Best_score1, Best_pos1, curve1] = PECO(SearchAgents_no, Max_iteration, lb, ub, dim, fun);

[0174] best_hd = round(Best_pos1(1,1));

[0175] best_lr = Best_pos1(1,2).

[0176] Furthermore, the training set is used to train the PECO-BP neural network prediction model, and the trained PECO-BP neural network prediction model is used to predict blood glucose solubility and calculate the loss function value. The code is as follows:

[0177] Network training:

[0178] net.trainParam.showWindow = 1;

[0179] net = train(net, p_train, t_train);

[0180] PECO-BP neural network prediction model predicts:

[0181] t_sim1 = sim(net, p_train);

[0182] t_sim2 = sim(net, p_test);

[0183] Adjust the loss function value using the optimal regularization parameter value:

[0184] error2 = sqrt(sum((T_sim2 - T_test ).^2) . / N)+ Q_best * sum(z * w^2).

[0185] Furthermore, run the code and output the result graph, such as Figures 3 to 5 .

[0186] Further, such as Figure 3 ,From the fitness graph, we can see that the fitness value of the PECO-BP based blood ,glucose prediction model is significantly smaller than that of ECO-BP, which indicates that the PECO-BP blood ,glucose solubility prediction model is accurate in terms of prediction. Figure 3 The PECO-BP curve in the figure can effectively avoid falling into the local minimum during the iteration process, and the fitness value continues to decrease and eventually stabilizes, indicating that the improved PECO algorithm has successfully improved the accuracy of blood glucose concentration prediction; in the early stage of iteration, the fitness value of the PECO-BP model decreases rapidly, showing a faster convergence speed; after about the 10th iteration, the fitness value of PECO-BP gradually stabilizes, and finally reaches the optimal value of 0.114909 at the 75th iteration, while ECO-BP reaches 0.187902 at the same number of iterations, which shows that PECO-BP can more effectively reduce the fitness value and avoid the problem that the BP prediction model is prone to falling into the local minimum.

[0187] Further, from Figure 4 It can be seen from the figure that the comparison between the true value of blood glucose concentration (solid line part) and the predicted value of the blood glucose concentration prediction model based on ECO-BP and PECO-BP shows that the predicted value of the PECO-BP blood glucose concentration prediction model is closer to the trend of the true value, and the fluctuation range is more consistent with the actual blood glucose concentration, indicating that its prediction accuracy is higher. The improvement of PECO-BP enables it to better fit the changes in blood glucose concentration, especially in areas with large fluctuations in the data samples, and can more accurately reflect the high and low changes in blood glucose concentration; although the PECO-BP blood glucose concentration prediction model can also capture the overall trend of blood glucose concentration, the difference between the predicted value and the true value is greater than that of PECO-BP, especially at points with large fluctuations, the prediction result deviation is more obvious, indicating that the prediction ability of ECO-BP is insufficient and is easily affected by local optimality.

[0188] Further, such as Figure 5As shown in the figure, from the data chart of the test set, it can be seen that the performance of the PECO-BP model on the test set is still better than that of the ECO-BP model, and the prediction error is smaller. Although there are fluctuations in the error graph of both models, the error of the PECO-BP model is closer to 0, indicating that its generalization ability is strong and it can predict new data more accurately. The performance of the PECO-BP model on both the prediction set and the test set is better than that of the ECO-BP model, demonstrating its higher prediction accuracy and generalization ability. The improved educational competition optimization algorithm makes the PECO-BP model have obvious advantages in solving local optimal problems, reducing error fluctuations and enhancing model stability, and therefore it is more suitable for dynamic prediction of blood glucose concentration.

[0189] Further, such as Figure 5 As shown in the figure, the error distribution of the prediction models of PECO-BP and ECO-BP on different sample numbers. The fluctuation range of the blood glucose prediction error based on ECO-BP is relatively large, and the error values ​​of multiple data points are above 20, even close to 30, indicating that the blood glucose prediction based on ECO-BP has a large error and is unstable when predicting blood glucose concentration, and the prediction accuracy is relatively low. The fluctuation of the blood glucose prediction error based on PECO-BP is relatively small, most of the error values ​​are within 10, and the error of most data points is close to 0, which shows that the prediction error of blood glucose prediction based on PECO-BP is small, the prediction accuracy is high, and the prediction results are more stable. In summary, the prediction performance of blood glucose prediction based on PECO-BP on the test set is better than that of the ECO-BP model, with lower prediction error and higher accuracy. The improved optimization algorithm of PECO-BP can effectively reduce the prediction error and improve the prediction stability of the model.

Claims

1. A method for optimizing blood glucose concentration prediction, characterized in that: The following steps are involved: S1. constructing an original data set for predicting blood glucose concentration, and performing a unit root test on the original data set. If the original data is stable, no processing is performed; otherwise, first-order difference processing is performed on the non-stationary original data; Obtain a first data set; S2, normalizing the first data set, and smoothing the normalized first data set using a correlation coefficient to obtain an input data set for blood glucose concentration prediction; S3, converting the input data set into a time series consisting of input prediction points And output predicted value A supervised input data set composed of a training set and a test set; S4, inputting the supervised input data set into the blood glucose prediction model based on PECO-BP, and outputting the real-time prediction value of blood glucose concentration; the PECO-BP is to optimize the learning rate factor and regularization parameter of the BP neural network prediction model by using the improved educational competition optimization algorithm to obtain the optimal learning rate factor and regularization parameter, and the optimal learning rate factor and the optimal regularization parameter Used in BP neural network prediction model; among them, the improved educational competitive optimization algorithm (PECO) includes: S41, according to the convergence rate and individual students With school individuals The difference between Improvement of the step size factor of the position update mathematical model in primary school ; S42. According to each student The current fitness value and the population individual Distance to global optimal solution Dynamic improvement Endurance factor of educational competition optimization algorithm .

2. A method for predicting and optimizing blood glucose concentration according to claim 1, characterized in that: The original data set for constructing blood glucose concentration prediction includes blood glucose concentration (mg / dL) and time series of N1 related variables, wherein the time series data of the related variables include: heart rate (bpm), body temperature (℃), food intake (calories), and blood pressure (mmHg).

3. A blood glucose concentration prediction and optimization method according to claim 2, characterized in that: The supervised input data set is input into the PECO-BP-based blood glucose prediction model, and the real-time predicted value of blood glucose concentration is output, including: the supervised input data enters the PECO-BP blood glucose prediction model from the input layer, and is transmitted to the output layer through the hidden layer to obtain the real-time predicted value of blood glucose concentration; the real-time predicted value of blood glucose concentration is compared with the actual value, and the error is calculated through the loss function. If the error between the two meets the set accuracy requirement, the training process is terminated; if the requirement is not met, the error is transmitted from the output layer through the hidden layer to the input layer by back propagation, and the weight of the blood glucose prediction model is continuously adjusted to gradually reduce the sum of square errors of the blood glucose prediction model until the set accuracy requirement is met; finally, the test set data is used to predict the blood glucose concentration using the trained PECO-BP-based blood glucose prediction model, and the weight of the blood glucose prediction model is continuously adjusted, specifically: the learning rate factor and the regularization parameter value are adjusted using the improved educational competition optimization algorithm to obtain the optimal learning rate factor and the regularization parameter value, and the weight and loss function of the blood glucose prediction model are adjusted using the optimal learning rate factor and the regularization parameter value, and the above steps are repeated until the error value changes less than 0.0001.

4. A method for predicting and optimizing blood glucose concentration according to claim 3, characterized in that: The improved educational competition optimization algorithm includes the following specific improvements: and individual students With school individuals The difference between Improvement of the step size factor of the position update mathematical model in primary school ; Among them, the convergence rate The convergence rate of the learning rate factor and regularization parameter optimization of the BP neural network prediction model is optimized by using the current educational competition optimization algorithm population diversity quantitative educational competition optimization algorithm. The mathematical model is: （3）； Where iter is the current iteration number, R(iter) is the convergence rate value of the iter-th iteration, and H max is the maximum population diversity value of the iter-th iteration; H(iter) is the population diversity value of the iter-th iteration, and the mathematical model is: （4）； In the formula, is the optimal fitness value of the i-th individual position of the population at the iter-th iteration; N is the maximum population size, where the individual positions of the population include the individual positions of students and the individual positions of schools; Among them, the difference between the individual positions of students and the individual positions of schools in the primary school stage of the educational competition optimization algorithm is calculated, and the mathematical model is: （5）； Where D is the problem dimension, and its value is the number of learning rate factors and regularization parameter values. It is the iterth iteration The position value of each student on the jth dimension, is the position value of the k-th school individual on the j-th dimension in the iter-th iteration, j=1,2,3,..,D; is the iterth iteration The individual student and The difference between the individual schools; ; , The maximum size for individual students, The largest individual school size. + =N; Among them, the improved step factor The mathematical model is: （6）； In the formula, is the improved step factor value of the i-th individual in the iter-th iteration, wherein the individuals include student individuals and school individuals, and the maximum value is N; is the maximum value of the improved step size factor, which is 1, and μ is the amplitude value of the adjustment difference.

5. A method for predicting and optimizing blood glucose concentration according to claim 4, characterized in that: The improved educational competition optimization algorithm includes: dynamically improving the endurance factor of the educational competition optimization algorithm according to the distance between the current fitness value of each individual student in the junior high school stage of the educational competition optimization algorithm and the global optimal solution , the improved endurance factor mathematical model is: （7）； In the formula, is the iterth iteration The fitness value of each student is is the global optimal fitness value of the population individuals, is the iterth iteration The endurance factor of each student improved, To adjust the parameters, control the steepness of the endurance factor.

6. A method for predicting and optimizing blood glucose concentration according to claim 5, characterized in that: The improved educational competition optimization algorithm is used to optimize the learning rate factor and regularization parameter of the BP neural network prediction model. The method is as follows: S31, establishing a real number mapping relationship between the space vector U of the learning rate factor and the regularization parameter and the individual student positions and the individual school positions of the improved educational competition optimization algorithm; S32, initialize the problem dimension D of the improved educational competition optimization algorithm, the maximum population size N, and the maximum number of iterations T max , improved step size factor , improved endurance factor , the minimum value of the fitness function, the position of the improved educational competition optimization algorithm initialized by chaotic mapping; S33, using the fitness function to calculate the fitness values ​​of the population individuals in the initialization phase, arranging the fitness values ​​in descending order, and the population individuals corresponding to the first 50% of the fitness values ​​are used as school individuals, and the rest are used as student individuals; S34, determine whether the current iteration number iter satisfies iter%3=1, if so, establish and execute the improved educational competition optimization algorithm for primary school stage position update mathematical model; otherwise, if iter%3=2, establish and execute the improved educational competition optimization algorithm for junior high school stage position update mathematical model, otherwise establish and execute the improved educational competition optimization algorithm for high school stage position update mathematical model; S35, if the fitness value of the current i-th population individual belongs to the top 50%, the school position update mathematical model is executed; otherwise, the student position update mathematical model is executed to obtain the population individual position value of the improved education competition optimization algorithm; S36, using the fitness function to calculate the current population individual fitness value, retaining the population individual position value corresponding to the minimum fitness value, which is the best population individual position of the current iteration , and the best individual position of the last iteration , retain the smaller position value of the two, which is the global optimal population individual position value ; S37, the current iteration number executes iter=iter+1, and determines whether the current iteration iter is greater than the maximum iteration number T max If yes, the optimization process ends and the global optimal population individual position value is output , after parsing, assign it to the space vector U of the learning rate factor and the regularization parameter to obtain the optimal learning rate factor and the optimal regularization parameter; otherwise, return to execute S33.

7. A method for predicting and optimizing blood glucose concentration according to claim 6, characterized in that: In S37, the optimal learning rate factor adjusts the weight of the blood glucose prediction model, and the mathematical model is: （3）； In the formula, is the weight of the adjusted blood glucose prediction model, is the weight of the blood glucose prediction model before adjustment, is the optimal learning rate factor, is the gradient of the loss function with respect to the weight; The optimal regularization parameter adjusts the loss function of the blood glucose prediction model, and the mathematical model is: （4）； In the formula, is the loss function, M is the input data set size for blood glucose concentration prediction, y is the actual blood glucose concentration, is the blood glucose concentration at the tth time point, is the optimal regularization parameter value, Z is the weight scale of the blood glucose prediction model, is the square of the current z-th weight.

Images