CN118609814A - A blood glucose concentration prediction method, prediction system, terminal device, and storage medium based on machine learning - Google Patents

Abstract

The application provides a machine learning-based blood glucose concentration prediction method, a prediction system, terminal equipment and a storage medium, wherein the blood glucose concentration prediction method comprises the following steps: acquiring an odor characteristic data set of a person to be tested and corresponding measured original blood glucose concentration data; performing wavelet transformation denoising and normalization processing on the odor characteristic data set; classifying the processed odor characteristic data set through a pre-constructed classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set; searching the measured original blood sugar concentration data corresponding to the predicted low blood sugar data set and the predicted high blood sugar data set respectively; predicting a regression model; the blood glucose concentration of the person to be tested can be accurately predicted flexibly and efficiently by combining classification and regression, and the blood glucose prediction efficiency and generalization capability are improved; is applicable to the technical field of blood sugar concentration detection.

Description

A blood glucose concentration prediction method, prediction system, terminal device, and storage medium based on machine learning

Technical Field

The present application relates to the technical field of blood glucose concentration detection, and specifically to a blood glucose concentration prediction method, prediction system, terminal device, and storage medium based on machine learning.

Background Art

With the development of social economy, the increase of population aging and the change of people's lifestyle, the prevalence and incidence of diabetes are on the rise worldwide. Diabetes is a chronic disease characterized by hyperglycemia caused by absolute or relative insufficient secretion of insulin and utilization disorder. The disease is mainly divided into three types: type 1, type 2 and gestational diabetes. The cause is mainly attributed to the combined effects of genetic and environmental factors, including decreased insulin secretion caused by islet cell dysfunction, or the body's insensitivity to insulin, or both, so that glucose in the blood cannot be effectively utilized and stored. Some diabetic patients and families have a clustering phenomenon of the disease. Therefore, monitoring and accurate prediction of blood sugar is a technical problem that needs to be solved urgently.

In recent years, researchers at home and abroad have begun to explore the use of odor recognition technology as a tool for diabetes detection. For example, deep neural networks are used to process large-scale odor data and learn specific patterns of exhaled gas from it to assist in the diagnosis of diabetes; random forest algorithms are used to establish an association model between odor characteristics and diabetes to achieve the diagnosis of diabetes. However, the above algorithms have not been very effective in identifying blood sugar and have not obtained accurate blood sugar levels.

Summary of the invention

In order to solve one of the above-mentioned technical defects, the present application provides a blood glucose concentration prediction method, prediction system, terminal device, and storage medium based on machine learning.

According to a first aspect of the present application, a method for predicting blood glucose concentration based on machine learning is provided, comprising the following steps:

Acquire an odor characteristic data set of the subject to be tested and the corresponding measured original blood glucose concentration data, wherein the odor characteristic data set is acquired by collecting data through a gas sensor array;

Perform wavelet transform denoising and normalization processing on the odor feature data set to obtain a processed odor feature data set;

The processed odor feature data set is classified by a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set;

Searching for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia dataset and the predicted hyperglycemia dataset respectively;

The measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set are regressed through a pre-constructed regression model for predicting hypoglycemia concentration, and the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set are regressed through a pre-constructed regression model for predicting hyperglycemia concentration, to obtain the predicted result of the blood glucose concentration of the subject, and the predicted result of the blood glucose concentration is output.

Preferably, the method for constructing the classification model comprises the following steps:

Obtain the odor feature data set and corresponding blood glucose concentration data of the experimental samples;

Preprocess the data to handle missing values and outliers in the odor feature dataset and blood glucose concentration data;

Dividing the odor feature data set of the preprocessed experimental samples into a first training set and a first test set;

Initialize the parameters of the XGboost model, including the type of weak learner, the minimum loss reduction threshold required for leaf node splitting, the learning rate, and the maximum log depth;

The parameters of the XGboost model are optimized through the CFAWOA algorithm, the optimal parameters are output, and the optimal parameters are assigned to the XGboost model to obtain the optimized XGboost model;

Train the optimized XGboost model, substitute the first training set into the optimized XGboost model, and train the XGboost model by adding weak learners to update the prediction value of the XGboost model. When the maximum number of iterations of the XGboost model is met or the prediction performance of the XGboost model is no longer improved, stop training and use the trained XGboost model as the classification model; otherwise, continue to add weak learners to train the XGboost model after the previous round of training;

Evaluate the classification model, calculate the mean square error or root mean square error of the classification result obtained by the trained XGboost model for the first test set and the corresponding blood glucose concentration data, and then evaluate the classification model.

More preferably, the method for constructing the regression model for predicting low blood sugar concentration and the method for constructing the regression model for predicting high blood sugar concentration both comprise the following steps:

The odor feature dataset of the preprocessed experimental samples is divided into a low-glucose training set and a high-glucose training set according to the blood glucose concentration; the low-glucose training set consists of S experimental samples: Dlow = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _S , y _S ), where g _s ∈ R ^m , m is the number of odor features in each experimental sample, y _s is the blood glucose concentration of the s-th low-glucose experimental sample, s = 1, 2, K, S, and S is the number of experimental samples in the low-glucose training set; the high-glucose training set consists of V samples, Dhigh = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _V , y _V ), where g _v ∈ R ^m , m is the number of odor features in each experimental sample, y _v is the actual blood glucose value of the v-th high-glucose experimental sample, v = 1, 2, K, V, and V is the number of experimental samples in the high-glucose training set;

A regression model for predicting low blood sugar concentration is constructed using the low blood sugar training set, and a regression model for predicting high blood sugar concentration is constructed using the high blood sugar training set.

Preferably, optimizing the parameters of the XGboost model by using the CFAWOA algorithm, outputting the optimal parameters, assigning the optimal parameters to the XGboost model, and obtaining the optimized XGboost model comprises the following steps:

S10, initialize the number of agents and the maximum number of iterations of the WOA algorithm, and randomly initialize the position values of all agents in each dimension within the value range;

S20, calculating the decision variable length of the WOA algorithm;

S30, setting parameters of the WOA algorithm, randomly generating initial whale individual position parameters, the parameters of the WOA algorithm include: population size N, dimension of search space, whale speed, and adaptive weight;

S40, taking the mean square error as the fitness value, calculating the fitness f(x) of each individual whale, and recording the current optimal individual and position;

S50, the sin chaos self-mapping model is applied to initialize the population and improve the population distribution; the sin chaos self-mapping model is: n＝0,1,2,...,N-1≤x _n ≤1,x _n ≠0;

S60, dynamically adjust the adaptive weight ω; Where f(x) is the fitness of individual whale x, u is the best fitness value in the whale population in the first iteration calculation, and iter represents the current number of iterations;

S70, updating the value of control parameter a, The value of a decreases linearly from 2 to 0; where t is the current iteration number and T _max is the maximum iteration number;

Update the current whale individual location X(t);

Update the values of A, C, l, and p, A = 2ar ₁ -a, C = 2r ₂ , A is a random number in [-a, a], l is a random number in (-1, 1); where r ₁ and r ₂ are random numbers in (0, 1);

Select the worst individual position vector X _worst and the best individual position vector X ^* (t) in the population;

S80, dynamically adjusting the b value according to the current search situation, and selecting the optimal constant b to define the shape of the spiral;

S90, the whale population selects a behavioral model to update the current position of individual whales;

S100, update the optimal whale position X ^* (t) according to the worst whale position X _worst , and calculate the fitness f (x) of the individuals in the group;

S110, determining whether the maximum number of iterations has been reached, if so, ending the loop and outputting the optimal parameters; otherwise, returning to step S30;

Or determine whether the fitness value of the current optimal solution is the same as the minimum fitness value in the previous iteration. If so, end the loop and output the optimal parameters; otherwise, return to step S30;

S120, assigning the optimal parameters to the XGboost model to obtain an optimized XGboost model.

More preferably, the step S90, wherein the whale population selects a behavior model to update the position of the current individual whale, comprises the following steps:

The probability threshold is pre-set as p _i ;

If p＜ _pi , then perform the following steps:

If |A<1|, the position of the current whale individual is updated using the shrinking and encircling position update model; the shrinking and encircling position update model is: X(t+1)＝ωX ^* (t)-AD, where D＝|CX ^* (t)-X(t)|;

If |A≥1|, the search and foraging mode position update model is used to update the current individual whale position, and the current worst individual whale position is updated according to the feedback model; the search mode position update model is: X(t+1)＝X _rand -AD, where D＝|CX _rand -X _t |, X _rand is a randomly selected whale position vector; the feedback model is: X _worstnew ＝X _worst -r·(X _p -X _worst ), where r means that if X _worstnew is better than X _worst , then X _worstnew is accepted;

If _p≥pi , the spiral update position model is used to update the current individual whale position; the spiral update position model is: X(t+1) ^＝ ωX ^* (t)+ _Dpeblcos (2πl), where _Dp ＝|X ^* (t)-X(t)|.

Preferably, the regression model for predicting low blood sugar concentration is an adaptive boosting regression model:

where G(g) is the median of all _{ω'k hk} ( _gs ), k = 1, 2, K, K, K is the number of weak learners, s = 1, 2, K, S, S is the number of experimental samples in the hypoglycemia training set, _ω'k is the weight of the kth weak learner, and _hk ( _gs ) is the prediction result of the kth weak learner for the sth hypoglycemia experimental sample.

More preferably, the regression model for predicting high blood sugar concentration is a gradient boosting regression model:

Where g∈R _qj , L(y _v ,c) is the loss function, v=1,2,K,V, V is the number of experimental samples in the hyperglycemia training set, y _v is the actual blood glucose value of the vth hyperglycemia experimental sample, c is the cumulative model obtained in the q-1th round, Q is the number of iterations, c _qj is the best fitting value of the qth regression tree at the jth leaf node, j=1,2,KJ, and J is the number of leaf nodes.

According to a second aspect of the present application, a blood glucose concentration prediction system based on machine learning is provided, comprising:

A data acquisition unit, used to acquire a data set of odor characteristics of the subject to be tested and the corresponding original blood glucose concentration data, wherein the data set of odor characteristics is acquired by collecting data through a gas sensor array;

A preliminary processing unit, used for performing wavelet transform denoising and normalization processing on the odor feature data set to obtain a processed odor feature data set;

A classification unit, used to classify the processed odor feature data set of the test subject through a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set;

A search unit, used to search for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set and the predicted hyperglycemia data set respectively;

A blood glucose concentration prediction unit is used to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set through a pre-built regression model for predicting hypoglycemia concentration, and to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set through a pre-built regression model for predicting hyperglycemia concentration, so as to obtain a predicted result of the blood glucose concentration of the subject to be tested;

The prediction result output unit outputs the blood sugar concentration prediction result.

According to a third aspect of the present application, a terminal device is provided, including:

Memory;

Processor; and

Computer programs;

Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the blood glucose concentration prediction method as described in any one of the above contents.

According to a fourth aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement the blood glucose concentration prediction method as described in any of the above contents.

In this application, classification and regression are combined to flexibly and efficiently make accurate predictions of the blood glucose concentration of the subject. The classification model provides category outputs that can be easily interpreted and understood, simplifying subsequent work, and can also capture the complex relationship between features, which helps to reduce the feature dimension in the regression model. The classification model can determine the prediction category of blood glucose concentration; after classification, the corresponding regression model is used to regress the odor feature data of different categories, thereby achieving accurate numerical prediction of blood glucose concentration. Among them, the classification model is more resistant to outliers and data errors than the regression model. The method of classification first and then regression can simplify the problem, improve the efficiency and generalization ability of blood glucose prediction, improve the interpretability and robustness of the prediction process, and better capture the nonlinear relationship of odor feature data.

Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the contents indicated in the written description and the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

FIG1 is a flow chart of a blood glucose concentration prediction method based on machine learning provided in the present application;

FIG2 is a ROC curve diagram of prediction classification using classification method 1;

FIG3 is a ROC curve diagram of prediction classification using classification method 2 provided in the present application;

FIG4 is a graph showing the results of directly predicting blood glucose concentration using a gradient boosting regression model;

FIG5 is a graph showing the result of predicting blood glucose concentration using a blood glucose concentration prediction method based on machine learning provided in the present application;

FIG6 is a schematic diagram of the structure of a blood glucose concentration prediction system based on machine learning provided in the present application;

FIG7 is a schematic diagram of the structure of a classification unit provided in the present application;

In the figure: 100 is a data acquisition unit, 110 is a preliminary processing unit, 120 is a classification unit, 130 is a search unit, 140 is a blood glucose concentration prediction unit, 150 is a prediction result output unit, 1201 is a sample data acquisition module, 1202 is a data preprocessing module, 1203 is a partitioning module, 1204 is an initialization parameter module, 1205 is an optimization module, 1206 is a classification model training module, and 1207 is a classification model evaluation module.

DETAILED DESCRIPTION

In order to make the technical solutions and advantages in the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than an exhaustive list of all the embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

As shown in FIG1 , in response to the above problems, a method for predicting blood glucose concentration based on machine learning is provided in an embodiment of the present application, comprising the following steps:

Acquire an odor characteristic data set of the subject to be tested and the corresponding measured raw blood glucose concentration data, wherein the odor characteristic data set is acquired by collecting through a gas sensor array; wherein the gas sensor array includes the following sensors: WSP2110, MICS-5524, TGS2602, TGS822, SMD-1015, ENS160, TGS826, MICS-4514;

The odor feature data set is subjected to wavelet denoising and normalization processing to obtain a processed odor feature data set;

The processed odor feature data set is classified by a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set;

Searching for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia dataset and the predicted hyperglycemia dataset respectively;

The measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set are regressed through a pre-constructed regression model for predicting hypoglycemia concentration, and the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set are regressed through a pre-constructed regression model for predicting hyperglycemia concentration, to obtain the blood glucose concentration prediction result of the subject to be tested, and the blood glucose concentration prediction result is output.

In this application, the exhaled gas of different test subjects is collected multiple times through the odor recognition system of non-invasive detection to obtain odor characteristic data, and then classification and regression are combined to flexibly and efficiently make accurate predictions of the blood glucose concentration of the test subject. The classification model provides category outputs that can be easily interpreted and understood, simplifying subsequent work, and can also capture the complex relationship between features, which helps to reduce the feature dimensions in the regression model. The classification model can determine the prediction category of blood glucose concentration; after classification, the corresponding regression model is used to regress the odor characteristic data of different categories, thereby achieving accurate numerical prediction of blood glucose concentration. Among them, the classification model is more resistant to outliers and data errors than the regression model. The method of classification first and then regression can simplify the problem, improve the efficiency and generalization ability of blood glucose prediction, improve the interpretability and robustness of the prediction process, and better capture the nonlinear relationship of odor characteristic data.

Furthermore, the method for constructing the classification model comprises the following steps:

Obtain the odor feature data set and corresponding blood glucose concentration data of the experimental samples;

Preprocess the data to handle missing values and outliers in the odor feature dataset and blood glucose concentration data;

The odor feature data set of the preprocessed experimental samples is divided into a first training set and a first test set; wherein the _{first training set consists of Z experimental samples: Dfirst training set} = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _Z , y _Z ); wherein, g _z ∈ R ^m , z = 1, 2, K, Z, m is the number of odor features in each experimental sample, y _z is the blood glucose concentration of the zth hypoglycemia experimental sample, and Z is the number of experimental samples in the first training set;

Initialize the parameters of the distributed gradient boosting model (XGboost model), including the type of weak learners, the minimum loss reduction threshold required for leaf node splitting, the learning rate, and the maximum tree depth; select the maximum tree depth (max_depth), the learning rate that determines the weight update step size in each iteration (learning_rate), and the minimum loss reduction threshold required for leaf node splitting (gamma) as the main parameters;

The parameters of the XGboost model are optimized through the chaotic feedback adaptive whale optimization algorithm (CFAWOA algorithm), the optimal parameters are output, and the optimal parameters are assigned to the XGboost model to obtain the optimized XGboost model; in the optimized XGboost model, the max_depth value is 9, the learning_rate value is 0.15, and the gamma value is 0;

The optimized XGboost model is trained, the first training set is substituted into the optimized XGboost model, and the prediction value of the XGboost model is updated by adding weak learners to train the XGboost model. When the maximum number of iterations of the XGboost model is met or the prediction performance of the XGboost model is no longer improved, the training is stopped and the trained XGboost model is used as the classification model; otherwise, the weak learners are continued to be added to train the XGboost model after the previous round of training; the classification model F is:

F = { _fe (z) = WQ _(z) };

Where: F is the classification model, is the final trained XGboost model, i.e., the tree structure set, e=1,2,K,E, E is the number of decision trees, fe _(z) is the predicted value of the blood glucose concentration of the odor feature data set of the zth experimental sample by the eth decision tree, Q(z) is the tree structure of the zth experimental sample mapped to the leaf node, and W is the real number score of the leaf node (i.e., the leaf node weight);

Evaluate the classification model, calculate the mean square error or root mean square error of the classification result obtained by the trained XGboost model for the first test set and the corresponding blood glucose concentration data, and then evaluate the classification model.

This application classifies the odor feature data set of the subject to be tested through a pre-built classification model, and optimizes the parameters of the XGboost model through the CFAWOA algorithm, avoiding the low optimization accuracy and easy to fall into local minima problems that occur when the standard whale optimization algorithm handles complex function optimization problems. The CFAWOA algorithm has good generalization ability and can help the XGboost model better fit the data and improve prediction performance. In this application, the classification model is obtained by training the first training set, and then the classification model is evaluated by the first test set. Ensure the classification accuracy of the obtained classification model.

Furthermore, optimizing the parameters of the XGboost model by the CFAWOA algorithm, outputting the optimal parameters, assigning the optimal parameters to the XGboost model, and obtaining the optimized XGboost model includes the following steps:

S10, initialize the number of agents and the maximum number of iterations of the WOA algorithm, and randomly initialize the position values of all agents in each dimension within the value range; specifically, initialize the maximum number of iterations to 500;

S20, calculating the decision variable length of the WOA algorithm;

S30, setting parameters of the WOA algorithm, randomly generating initial whale individual position parameters, the parameters of the WOA algorithm include: population size N, dimension of search space, whale speed, and adaptive weight;

Specifically, the population size is set to 30, the dimension of the search space is set to 4, the whale speed is set to 0, and the adaptive weight is set to 0.5;

S40, taking the mean square error as the fitness value, calculating the fitness f(x) of each individual whale, and recording the current optimal individual and position;

S50, the sin chaos self-mapping model is applied to initialize the population and improve the population distribution; the sin chaos self-mapping model is: n＝0,1,2,...,N-1≤x _n ≤1,x _n ≠0;

S60, dynamically adjust the adaptive weight ω; Where f(x) is the fitness of individual whale x, u is the best fitness value in the whale population in the first iteration calculation, and iter represents the current number of iterations;

S70, updating the value of control parameter a, The value of a decreases linearly from 2 to 0; where t is the current iteration number and T _max is the maximum iteration number;

Update the current whale individual location X(t);

Update the values of A, C, l, and p, A = 2ar ₁ -a, C = 2r ₂ , A is a random number in [-a, a], l is a random number in (-1, 1); where r ₁ and r ₂ are random numbers in (0, 1);

Select the worst individual position vector X _worst and the best individual position vector X ^* (t) in the population;

S80, dynamically adjusting the b value according to the current search situation, and selecting the optimal constant b to define the shape of the spiral;

S90, the whale population selects a behavior model to update the position of the current individual whale, including the following steps:

The probability threshold is pre-set as p _i ;

If p＜ _pi , then perform the following steps:

If |A<1|, the position of the current whale individual is updated using the shrinking and encircling position update model; the shrinking and encircling position update model is: X(t+1)＝ωX ^* (t)-AD, where D＝|CX ^* (t)-X(t)|;

If |A≥1|, the search and foraging mode position update model is used to update the current individual whale position, and the current worst individual whale position is updated according to the feedback model; the search mode position update model is: X(t+1)＝X _rand -AD, where D＝|CX _rand -X _t |, X _rand is a randomly selected whale position vector; the feedback model is: X _worstnew ＝X _worst -r·(X _p -X _worst ), where r means that if X _worstnew is better than X _worst , then X _worstnew is accepted;

If p ≥ p _i , the spiral update position model is used to update the current individual whale position; the spiral update position model is: X(t+1)＝ωX ^* (t)+D _p e ^bl cos(2πl), where D _p ＝|X ^* (t)-X(t)|;

S100, update the optimal whale position X ^* (t) according to the worst whale position X _worst , and calculate the fitness f (x) of the individuals in the group;

S110, determining whether the maximum number of iterations has been reached, if so, ending the loop and outputting the optimal parameters; otherwise, returning to step S30;

Or determine whether the fitness value of the current optimal solution is the same as the minimum fitness value in the previous iteration. If so, end the loop and output the optimal parameters; otherwise, return to step S30;

S120, assigning the optimal parameters to the XGboost model to obtain an optimized XGboost model.

In this application, chaos theory is introduced to generate the initial population based on the whale optimization algorithm, which increases the population diversity and lays the foundation for the global search of the algorithm. In addition, a feedback model is added in the later stage of whale position update, and the worst whale is quickly brought closer to the optimal whale through communication learning, which further improves the global search ability of the algorithm. At the same time, the adaptive inertia weight ω is introduced into the whale individual position update formula, and the optimization performance of the algorithm is further improved by balancing the development and exploration capabilities of the algorithm.

Furthermore, the method for constructing the regression model for predicting low blood sugar concentration and the regression model for predicting high blood sugar concentration both include the following steps:

The odor feature dataset of the preprocessed experimental samples is divided into a low-glucose training set and a high-glucose training set according to the blood glucose concentration; the low-glucose training set consists of S experimental samples: _Dlow = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _S , y _S ), where g _s ∈ R ^m , m is the number of odor features in each experimental sample, y _s is the blood glucose concentration of the s-th low-glucose experimental sample, s = 1, 2, K, S, and S is the number of experimental samples in the low-glucose training set; the high-glucose training set consists of V samples, _Dhigh = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _V , y _V ), where g _v ∈ R ^m , m is the number of odor features in each experimental sample, y _v is the actual blood glucose value of the v-th high-glucose experimental sample, v = 1, 2, K, V, and V is the number of experimental samples in the high-glucose training set;

A regression model for predicting low blood sugar concentration is constructed using the low blood sugar training set, and a regression model for predicting high blood sugar concentration is constructed using the high blood sugar training set.

Specifically, the regression model for predicting low blood sugar concentration is an adaptive boosting regression model, and its construction method includes the following steps:

1. Initialize the weights, record the initial distribution of the hypoglycemia training set as Dist ₁ distribution, and initialize the weight of each hypoglycemia experimental sample to 1/n. The Dist ₁ distribution is used for the training of the first learner h ₁ , the Dist _k distribution is used for the training of the kth weak learner h _k , and so on;

2. Loop iteration:

The following steps are iterated one by one according to the number of the learner. The number of the learner is defined as k, and k∈1,2,3,K,K.

1) Calculate the maximum error E _k of the weak learner h _k on the hypoglycemia training set;

E _k =max|y _s -h _k (g _s )|;

Where h _k (g _s ) is the prediction result of the kth weak learner for the sth hypoglycemia experimental sample;

2) Calculate the relative error e _ks of h _k for each hypoglycemia experimental sample;

3) Calculate the error rate e _k of the current weak learner;

4) Update the weight ω' _k of the current weak learner;

5) Update the weight distribution of the hypoglycemia training set;

Where Z _k is the normalization factor,

3. After the loop iteration is completed, the weighted average of the prediction results of multiple weak learners is calculated to obtain a strong regressor, that is, the adaptive boosting regression model H(g);

Where G(g) is the median of all ω' _k h _k ( _gi ).

In the adaptive boosting regression model provided in the present application, based on the distribution of the hypoglycemia training set, a weak learner h _k is trained on the hypoglycemia training set, and the weighted error rate e _k is calculated according to the prediction of the weak learner h _k and the blood glucose concentration data corresponding to the blood glucose training set. The weight ω' _k of the weak learner is calculated according to the weighted error rate e _k , the weight of the samples misclassified by the weak learner is increased, and the weight of the samples correctly classified is reduced to obtain a new weight distribution. Finally, K weak learners are weighted and combined into a strong learner to obtain an adaptive boosting regression model, which can accurately predict the blood glucose concentration of the predicted hypoglycemia data set.

Furthermore, the regression model for predicting high blood sugar concentration is a gradient boosting regression model, and its construction method includes the following steps:

Loop iteration,

1. Initialize the weak learner H' ₀ (g), and find a weak learner in the function space so that the cumulative model loss after adding the weak learner is minimized;

Where L(y _v ,c) is the loss function, c is the cumulative model obtained in the q-1th round;

2. For the number of iterations q = 1, 2, K, Q, where Q is the number of iterations, perform the following steps one by one:

1) Calculate the negative gradient for the hyperglycemic experimental sample;

Negative gradients ensure that when the best fit value is subsequently calculated, the new weak learner can be improved in the direction of reducing the cumulative model loss;

2) Fit a CART regression tree and obtain the qth regression tree, with a leaf node region R _qj , where j = 1, 2, KJ, and J is the number of corresponding leaf nodes;

3) Update and calculate the best fitting value c _qj of the qth regression tree at the jth leaf node;

4) Obtain a strong learner, namely the gradient boosting regression model H'(g);

The gradient boosting regression model provided in the present application is a technology for learning from its erroneous data, which generates multiple weak learners in series. The goal of each weak learner is to fit the negative gradient of the loss function of the previously accumulated model, so that the cumulative model loss after adding the weak learner is reduced in the direction of the negative gradient. The obtained strong learner can accurately predict the blood glucose concentration of the predicted hyperglycemia data set.

In order to further illustrate the classification effect of the classification model provided in this application on the preprocessed odor feature data set of the test subjects, 160 experimental samples were selected and predicted and classified by different methods, and divided into a training set and a test set according to the ratio of 70% and 30%.

Classification method 1: Use XGboost model for prediction and classification;

Classification method 2: Use the classification model provided in this application for prediction and classification;

As shown in Figures 2 and 3, the AUC value of the ROC curve after prediction and classification using classification method 1 is 0.89, and the AUC value of the ROC curve after prediction and classification using classification method 2 is 0.94. It can be seen that the prediction and classification effect of the classification model provided by the present application is better, which provides a reliable basis for further accurate detection of blood glucose concentration through regression models.

In order to further illustrate the blood glucose concentration prediction effects of the regression model for predicting low blood glucose concentration and the regression model for predicting high blood glucose concentration provided in the present application, comparative examples and experimental examples are set.

Comparative example: Directly use the gradient boosting regression model to predict the blood glucose concentration of the experimental samples;

Experimental example: The predicted hypoglycemia dataset and predicted hyperglycemia dataset classified by the above classification method 2 are respectively subjected to regression prediction using the adaptive boosting regression model H(g) and the gradient boosting regression model H'(g) provided in this application to obtain blood glucose concentration prediction results.

As shown in Figure 4, when the gradient boosting regression model is directly used to predict the blood glucose concentration of the experimental samples, the mean absolute error between the true blood glucose concentration and the predicted blood glucose concentration is 0.84 mmol/L, the root mean square error (RMSE) of the blood glucose concentration prediction is 0.32, ^R2 is 0.76, and the training time is 370 ms.

As shown in Figure 5, when the blood glucose concentration prediction method based on machine learning provided in the present application is used to predict the blood glucose concentration: when the blood glucose concentration is less than 0.7, the mean absolute error is 0.53 (the root mean square error is 0.13, and ^R2 is 0.93); when the blood glucose concentration is greater than or equal to 0.7, the mean absolute error is 0.84 (the root mean square error is 0.24, and ^R2 is 0.82). It can be concluded that the mean absolute error of the blood glucose concentration prediction method based on machine learning provided in the present application is approximately 0.60 mmol/L (the root mean square error is 0.21, and ^R2 is 0.87). Compared with the comparative example, it is shown that the blood glucose concentration prediction method provided in the present application has significant improvement in blood glucose prediction, further reduces the error, and improves the accuracy of the prediction results, which is better than the prediction effect of directly using the regression model without classification in the prior art.

As shown in FIG6 , accordingly, the embodiment of the present application further provides a blood glucose concentration prediction system based on machine learning, comprising:

The data acquisition unit 100 is used to acquire the odor characteristic data set of the subject to be tested and the corresponding measured original blood glucose concentration data, wherein the odor characteristic data set is acquired by collecting data through a gas sensor array;

A preliminary processing unit 110 is used to perform wavelet transform denoising and normalization processing on the odor feature data set to obtain a processed odor feature data set;

A classification unit 120, configured to classify the processed odor feature data set by using a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set;

A search unit 130, used to search for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set and the predicted hyperglycemia data set respectively;

The blood glucose concentration prediction unit 140 is used to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set through a pre-built regression model for predicting hypoglycemia concentration, and to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set through a pre-built regression model for predicting hyperglycemia concentration, so as to obtain the predicted result of the blood glucose concentration of the subject to be tested;

The prediction result output unit 150 outputs the blood sugar concentration prediction result.

As shown in FIG. 7 , specifically, the classification unit 120 includes:

The sample data acquisition module 1201 is used to acquire the odor characteristic data set of the experimental sample and the corresponding blood glucose concentration data;

A data preprocessing module 1202 is used to preprocess the data to process missing values and abnormal values in the odor characteristic data set and the blood sugar concentration data;

A division module 1203 is used to divide the odor feature data set of the preprocessed experimental sample into a first training set and a first test set;

Initialization parameter module 1204, used to initialize the parameters of the distributed gradient boosting model (XGboost model), including the type of weak learner, the minimum loss reduction threshold required for leaf node splitting, the learning rate, and the maximum log depth;

The optimization module 1205 is used to optimize the parameters of the XGboost model through the chaotic feedback adaptive whale optimization algorithm (CFAWOA algorithm), output the optimal parameters, assign the optimal parameters to the XGboost model, and obtain the optimized XGboost model;

The classification model training module 1206 is used to train the optimized XGboost model, substitute the first training set into the optimized XGboost model, and train the XGboost model by adding weak learners to update the prediction value of the XGboost model. When the maximum number of iterations of the XGboost model is met or the prediction performance of the XGboost model is no longer improved, the training is stopped and the trained XGboost model is used as the classification model; otherwise, the weak learners are continued to be added to train the XGboost model after the previous round of training;

The classification model evaluation module 1207 is used to evaluate the classification model, calculate the mean square error or root mean square error between the classification result obtained by the trained XGboost model of the first test set and the corresponding blood glucose concentration data, and then evaluate the classification model.

Accordingly, an embodiment of the present application further provides a terminal device, including:

Memory;

Processor; and

Computer programs;

Wherein, the computer program is stored in the memory and is configured to be executed by the processor to implement the blood glucose concentration prediction method as described in any one of the above contents.

Accordingly, an embodiment of the present application further provides a computer-readable storage medium on which a computer program is stored; the computer program is executed by a processor to implement the blood glucose concentration prediction method as described in any of the above contents.

In practical applications, the test subject needs to monitor and obtain accurate blood sugar concentration levels to further manage their health status in real time. The current method of obtaining accurate blood sugar concentration is still in the invasive stage, which brings a bad user experience to the test subject. The present application provides a blood sugar concentration prediction method based on machine learning, which adopts a non-invasive method. It can complete the blood sugar prediction through the odor characteristic data of the test subject and the corresponding measured original blood sugar concentration data. It combines classification and regression, and simplifies the problem by using the method of classification first and then regression, improves the efficiency and generalization ability of blood sugar prediction, improves the interpretability and robustness of the prediction process, can better capture the nonlinear relationship of the odor characteristic data, make the prediction result more accurate, and provide reliable blood sugar prediction data for the test subject.

Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiment, complete software embodiment, or the embodiment in combination with software and hardware. Moreover, the present application can adopt the form of the computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, C language, VHDL language, Verilog language, object-oriented programming language Java and literal scripting language JavaScript, etc.

The present application is described with reference to the flowchart and/or block diagram of the method, device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and/or box in the flowchart and/or block diagram, and the combination of the process and/or box in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one process or multiple processes in the flowchart and/or one box or multiple boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims (10)

1. A method for predicting blood glucose concentration based on machine learning, characterized in that it comprises the following steps: Acquire an odor characteristic data set of the subject to be tested and the corresponding measured original blood glucose concentration data, wherein the odor characteristic data set is acquired by collecting data through a gas sensor array; The odor feature data set is subjected to wavelet transform denoising and normalization processing to obtain a processed odor feature data set; The processed odor feature data set is classified by a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set; Searching for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia dataset and the predicted hyperglycemia dataset respectively; The measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set are regressed through a pre-constructed regression model for predicting hypoglycemia concentration, and the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set are regressed through a pre-constructed regression model for predicting hyperglycemia concentration, to obtain the predicted result of the blood glucose concentration of the subject, and the predicted result of the blood glucose concentration is output. 2. The method for predicting blood glucose concentration based on machine learning according to claim 1, wherein the method for constructing the classification model comprises the following steps: Obtain the odor feature data set and corresponding blood glucose concentration data of the experimental samples; Preprocess the data to handle missing values and outliers in the odor feature dataset and blood glucose concentration data; Dividing the odor feature data set of the preprocessed experimental samples into a first training set and a first test set; Initialize the parameters of the XGboost model, including the type of weak learner, the minimum loss reduction threshold required for leaf node splitting, the learning rate, and the maximum log depth; The parameters of the XGboost model are optimized through the CFAWOA algorithm, the optimal parameters are output, and the optimal parameters are assigned to the XGboost model to obtain the optimized XGboost model; Train the optimized XGboost model, substitute the first training set into the optimized XGboost model, and train the XGboost model by adding weak learners to update the prediction value of the XGboost model. When the maximum number of iterations of the XGboost model is met or the prediction performance of the XGboost model is no longer improved, stop training and use the trained XGboost model as the classification model; otherwise, continue to add weak learners to train the XGboost model after the previous round of training; Evaluate the classification model, calculate the mean square error or root mean square error of the classification result obtained by the trained XGboost model for the first test set and the corresponding blood glucose concentration data, and then evaluate the classification model. 3. The method for predicting blood glucose concentration based on machine learning according to claim 2, characterized in that the method for constructing the regression model for predicting low blood glucose concentration and the regression model for predicting high blood glucose concentration both comprise the following steps: The odor feature dataset of the preprocessed experimental samples is divided into a low-glucose training set and a high-glucose training set according to the blood glucose concentration; the low-glucose training set consists of S experimental samples: Dlow = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _S , y _S ), where g _s ∈ R ^m , m is the number of odor features in each experimental sample, y _s is the blood glucose concentration of the s-th low-glucose experimental sample, s = 1, 2, K, S, and S is the number of experimental samples in the low-glucose training set; the high-glucose training set consists of V samples, Dhigh = (g ₁ , y ₁ ), (g ₂ , y ₂ ), K, (g _V , y _V ), where g _v ∈ R ^m , m is the number of odor features in each experimental sample, y _v is the actual blood glucose value of the v-th high-glucose experimental sample, v = 1, 2, K, V, and V is the number of experimental samples in the high-glucose training set; A regression model for predicting low blood sugar concentration is constructed using the low blood sugar training set, and a regression model for predicting high blood sugar concentration is constructed using the high blood sugar training set. 4. The method for predicting blood glucose concentration based on machine learning according to claim 2, characterized in that the optimization of the parameters of the XGboost model by the CFAWOA algorithm, outputting the optimal parameters, assigning the optimal parameters to the XGboost model, and obtaining the optimized XGboost model comprises the following steps: S10, initialize the number of agents and the maximum number of iterations of the WOA algorithm, and randomly initialize the position values of all agents in each dimension within the value range; S20, calculating the decision variable length of the WOA algorithm; S30, setting parameters of the WOA algorithm, randomly generating initial whale individual position parameters, the parameters of the WOA algorithm include: population size N, dimension of search space, whale speed, and adaptive weight; S40, taking the mean square error as the fitness value, calculating the fitness f(x) of each individual whale, and recording the current optimal individual and position; S50, the sin chaos self-mapping model is applied to initialize the population and improve the population distribution; the sin chaos self-mapping model is: n＝0,1,2,...,N-1≤x _n ≤1,x _n ≠0; S60, dynamically adjust the adaptive weight ω; Where f(x) is the fitness of individual whale x, u is the best fitness value in the whale population in the first iteration calculation, and iter represents the current number of iterations; S70, updating the value of control parameter a, The value of a decreases linearly from 2 to 0; where t is the current iteration number and T _max is the maximum iteration number; Update the current whale individual location X(t); Update the values of A, C, l, and p, A = 2ar ₁ -a, C = 2r ₂ , A is a random number in [-a, a], l is a random number in (-1, 1); where r ₁ and r ₂ are random numbers in (0, 1); Select the worst individual position vector X _worst and the best individual position vector X ^* (t) in the population; S80, dynamically adjusting the b value according to the current search situation, and selecting the optimal constant b to define the shape of the spiral; S90, the whale population selects a behavioral model to update the current position of individual whales; S100, update the optimal whale position X ^* (t) according to the worst whale position X _worst , and calculate the fitness f (x) of the individuals in the group; S110, determining whether the maximum number of iterations has been reached, if so, ending the loop and outputting the optimal parameters; otherwise, returning to step S30; Or determine whether the fitness value of the current optimal solution is the same as the minimum fitness value in the previous iteration. If so, end the loop and output the optimal parameters; otherwise, return to step S30; S120, assigning the optimal parameters to the XGboost model to obtain an optimized XGboost model. 5. The blood glucose concentration prediction method based on machine learning according to claim 4 is characterized in that, in step S90, the whale population selection behavior model is used to update the position of the current whale individual, comprising the following steps: The probability threshold is pre-set as p _i ; If p＜ _pi , then perform the following steps: If |A<1|, the position of the current whale individual is updated using the shrinking and encircling position update model; the shrinking and encircling position update model is: X(t+1)＝ωX ^* (t)-AD, where D＝|CX ^* (t)-X(t)|; If |A≥1|, the search and foraging mode position update model is used to update the current individual whale position, and the current worst individual whale position is updated according to the feedback model; the search mode position update model is: X(t+1)＝X _rand -AD, where D＝|CX _rand -Xt|, X _rand is a randomly selected whale position vector; the feedback model is: X _worstnew ＝X _worst -r·(X _p -X _worst ), where r means that if X _worstnew is better than X _worst , then X _worstnew is accepted; If _p≥pi , the spiral update position model is used to update the current individual whale position; the spiral update position model is: X(t+1) ^＝ ωX ^* (t)+ _Dpeblcos (2πl), where _Dp ＝|X ^* (t)-X(t)|. 6. The method for predicting blood glucose concentration based on machine learning according to claim 3, characterized in that the regression model for predicting low blood glucose concentration is an adaptive boosting regression model: where G(g) is the median of all _{ω'k hk} ( _gs ), k = 1, 2, K, K, K is the number of weak learners, s = 1, 2, K, S, S is the number of experimental samples in the hypoglycemia training set, _ω'k is the weight of the kth weak learner, and _hk ( _gs ) is the prediction result of the kth weak learner for the sth hypoglycemia experimental sample. 7. The method for predicting blood glucose concentration based on machine learning according to claim 6, wherein the regression model for predicting high blood glucose concentration is a gradient boosting regression model: Where g∈R _qj , L(y _v ,c) is the loss function, v=1,2,K,V, V is the number of experimental samples in the hyperglycemia training set, y _v is the actual blood glucose value of the vth hyperglycemia experimental sample, c is the cumulative model obtained in the q-1th round, Q is the number of iterations, c _qj is the best fitting value of the qth regression tree at the jth leaf node, j=1,2,KJ, and J is the number of leaf nodes. 8. A blood glucose concentration prediction system based on machine learning, comprising: A data acquisition unit (100) is used to acquire an odor characteristic data set of a subject to be tested and corresponding original measured blood glucose concentration data, wherein the odor characteristic data set is acquired by collecting data through a gas sensor array; A preliminary processing unit (110) is used to perform wavelet transform denoising and normalization processing on the odor feature data set to obtain a processed odor feature data set; A classification unit (120), used to classify the processed odor feature data set by using a pre-built classification model to obtain a predicted hypoglycemia data set and a predicted hyperglycemia data set; A search unit (130), used to search for the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set and the predicted hyperglycemia data set respectively; A blood glucose concentration prediction unit (140) is used to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hypoglycemia data set through a pre-constructed regression model for predicting hypoglycemia concentration, and to perform regression processing on the measured original blood glucose concentration data corresponding to the predicted hyperglycemia data set through a pre-constructed regression model for predicting hyperglycemia concentration, so as to obtain a predicted result of the blood glucose concentration of the subject to be tested; The prediction result output unit (150) outputs the blood sugar concentration prediction result. 9. A terminal device, comprising: Memory; Processor; and Computer programs; The computer program is stored in the memory and is configured to be executed by the processor to implement the blood glucose concentration prediction method according to any one of claims 1 to 7. 10. A computer-readable storage medium, characterized in that a computer program is stored thereon; the computer program is executed by a processor to implement the blood glucose concentration prediction method according to any one of claims 1 to 7.