CN103310113B - A kind ofly to be separated and the general blood glucose prediction method of data modeling based on frequency band - Google Patents

Abstract

The invention discloses and to be a kind ofly separated and the general blood glucose prediction method of data modeling based on frequency band, the method is analyzed for the subcutaneous blood glucose measurement signal of human body, extract its potential time-series dynamics characteristic, and subcutaneous blood glucose signal is divided into height two frequency ranges by the threshold value defining frequency band separation; Analyze its sequential autocorrelation for low frequency blood glucose signal, set up autoregression blood glucose prediction model; General blood glucose prediction method proposed by the invention for new object without the need to etc. re-start modeling after abundant blood glucose measurement signal to be obtained, but the forecast model that directly can call other individuality carries out real time blood sugar prediction, enormously simplify modeling work amount and complexity, greatly will reduce modeling cost; And universal model adopts the method based on frequency band separation and latent variable modeling, improves precision of prediction.The present invention is easy to implement, and the research for blood glucose prediction modeling method specifies new direction.

Description

Universal blood glucose prediction method based on frequency band separation and data modeling

Technical Field

The invention belongs to the field of blood sugar data analysis and prediction research, and particularly relates to a general blood sugar prediction method based on frequency band separation and data modeling.

Background

One significant characteristic of human blood glucose levels is temporal variability, which is particularly reflected in the existence of significant autocorrelation relationships in time series signal measurements. The time-sequence correlation characteristics of the blood sugar signals can be extracted by analyzing and modeling the blood sugar signals, and the future blood sugar change condition can be obtained according to the dynamics of the historical blood sugar. Bremer and Gough, the foreign scholars in 1999, first suggested that blood glucose time series data have a potentially relevant structure, which can be described by a simple linear dynamic model. At present, a data-driven method is mostly adopted for establishing a blood glucose prediction model. Existing prediction models can be classified into linear (typified by a least-squares-based autoregressive model) and nonlinear (typified by a neural network method). Linear models have gained wide application due to their simple model structure and algorithms. More sophisticated modeling methods include autoregression, stimulus-response (IR), and the like. The autoregressive method is a relatively mature data modeling method, and obtains a future blood glucose predicted value through linear combination of historical blood glucose data by only utilizing the information of blood glucose. However, the conventional autoregressive model has two main defects: (a) the blood sugar prediction relation is directly fitted by using the least square which is the most basic identification method aiming at the measured data, so that the defects of the method cannot be avoided, and satisfactory prediction precision cannot be obtained; (b) different prediction models (namely individualized models) are directly established for online prediction without pre-analysis aiming at the blood sugar dynamics of different individuals, so that a large amount of manpower and material resources are consumed.

Disclosure of Invention

The invention aims to provide a general blood glucose prediction method based on frequency band separation and data modeling, aiming at the defects of the existing blood glucose prediction method.

The purpose of the invention is realized by the following technical scheme: a method for universal blood glucose prediction based on band separation and data modeling, the method comprising the steps of:

step 1: modeling blood sugar signal preprocessing: combining individual subcutaneous blood glucose signals obtained with a certain sampling period delta t into one-dimensional time sequence data x^T(1 × Z), wherein x is the measured value of the blood sugar signal, Z is the number of samples, and the spike noise is removed.

Step 2: separation of blood glucose signal frequency bands: and analyzing the change rule of the dynamics of the blood sugar in different high and low frequency bands, distinguishing a key frequency band from a secondary frequency band, and determining the optimal threshold for dividing the frequency bands. Performing frequency band separation on the blood glucose signal by using a Butterworth low-pass filter according to a defined separation threshold value.

And step 3: acquiring a predictive variable matrix and a response matrix: sliding a one-dimensional sliding window with the length of K sampling points over x^LT(1 × Z), moving a sampling point each time, moving for Z-K +1 times, taking the data in the sliding window each time as a new row vector, and combining the new row vector into a two-dimensional data matrix X^L(N × K), where N is Z-K +1, K is PL + H (PH/5), PL represents the length of the prediction variable, H represents the prediction step size, PH represents the prediction interval, and both H and PH have different measurement units, but both represent how many steps of blood glucose are predicted.

And 4, step 4: blood glucose prediction modeling based on band separation: the blood glucose prediction modeling is performed by a latent variable based method.

And 5: and (4) predicting the future blood sugar value of any individual according to the blood sugar prediction model established in the step (4). The method is completed by the following steps:

(5.1) in online prediction, any new data of an individual is collectedLast (subscript new stands for new sample, J ═ PL), pairPerforming the band separation process as described in step 2 to obtain

(5.2) calling a general low-frequency model for online prediction:

(5.2.1) calling a general low-frequency model built based on least square to perform online prediction: for each new test dataThe predicted value after PH sampling points is generated by the following steps

${\hat{y}}_{new}^L=x_{new}^{LT}{\mathrm{\θ}}^L;$

Wherein, theta^LThe regression coefficient vector previously found from the training data.

(5.2.2) calling a universal low-frequency model built based on latent variables to perform online prediction:

for each new test dataThe predicted value after PH sampling points are obtained through calculation according to the following steps

${\hat{y}}_{new}^L=x_{new}^{LT}{\mathrm{\σ}}^L;$

Wherein σ^LIs a regression coefficient vector calculated by a partial least squares-typical correlation analysis method.

(5.3) measuring the prediction accuracy: the obtained quality prediction result is compared with the actual measurement value. After obtaining a series of new prediction signals, the blood glucose prediction accuracy for a new sample can be calculated according to the RMSE (root mean square error) performance indicator defined in the following equation:

$RMSE=\sqrt{\frac{1}{N}\underset{i\∈N}{\Σ}{\left(y\right(i)-{\hat{y}}^{*}(i\left)\right)}^2};$

wherein y (i) represents the original blood glucose measurement,representing the predicted blood sugar value obtained by different models and using the first same individual model or different individual modelsOr the full-frequency predicted value obtained by adopting the second same individual modelN is the number of samples.

Compared with the prior art, the invention has the beneficial effects that: the general blood sugar prediction method provided by the invention has the advantages that the new object does not need to be modeled after a sufficient blood sugar measurement signal is obtained, but can directly call prediction models of other individuals to predict the blood sugar in real time, so that the modeling workload and complexity are greatly simplified, and the modeling cost is greatly reduced; and the general model adopts a method based on frequency band separation and latent variable modeling, so that the prediction precision is improved. The method is easy to implement, and a new direction is indicated for the research of the blood sugar prediction modeling method.

Drawings

FIG. 1 is a flow chart of a general blood glucose prediction model modeling according to the present invention;

FIG. 2 is a graph comparing a full-band subcutaneous glucose signal provided by a CGM (continuous glucose monitor) with a band-separated high and low band glucose signal;

FIG. 3 is a graph of the effect of PL on the glycemic prediction performance of the generic low frequency model (GL) (mean RMSE on the ordinate);

FIG. 4 is a graph of the effect of PL on the glycemic performance of the generic Low frequency model (GL) (MAD values in RMSE on the ordinate);

FIG. 5 is a graph of the effect of pH on glycemic prediction performance of the generic low frequency model (GL) (mean RMSE on ordinate);

FIG. 6 is a graph of the effect of pH on glycemic prediction performance of the generic low frequency model (GL) (with the MAD value in RMSE on the ordinate);

FIG. 7 is a graph comparing a blood glucose measurement curve and a blood glucose prediction curve of a generic low frequency (GL) model (blood glucose signals for which the analysis object is a first person of a first group).

Fig. 8 is a graph comparing a blood glucose measurement curve and a blood glucose prediction curve of a standard independent individual (SD) model (blood glucose signals for which the analysis object is a first person of a first group).

FIG. 9 is a graph comparing a blood glucose measurement curve and a blood glucose prediction curve of a generic low frequency (GL) model (analyzed for blood glucose signals of a second group of first persons).

FIG. 10 is a graph comparing a blood glucose measurement curve and a blood glucose prediction curve for a standard independent individual (SD) model (subject of analysis is the blood glucose signal of the second group of first persons).

Detailed Description

As shown in FIG. 1, the present invention relates to a general blood glucose prediction method based on frequency band separation and data modeling, which comprises the following steps:

step 1: modeling blood glucose signal preprocessing

For an individual subcutaneous blood glucose signal obtained with a certain sampling period Δ t (here Δ t ═ 5min), it can be combined into one-dimensional time series data x^Tin this example, we have blood glucose time series signals from two groups of subjects in common, the 1 st group includes 12 persons, the 2 nd group includes 14 persons, and the two groups include 26 persons, the blood glucose time series signal of each subject includes data of two or three days.

Step 2: blood glucose signal band separation

The step mainly analyzes the change rule of the dynamics of the blood sugar in different high and low frequency bands, distinguishes a key frequency band from a secondary frequency band and determines the optimal threshold value for dividing the frequency bands. Performing frequency band separation on the blood glucose signal by using a Butterworth low-pass filter according to a defined separation threshold value. The main purpose is to remove the noise influence of high frequency band and keep the effective blood sugar change information of low frequency band by frequency band separation.

(2.1) selecting a first order low-pass butterworth filter of the form:

$\tilde{x}\left(k\right)={\mathrm{\β}}_{1}x\left(k\right)+{\mathrm{\β}}_{2}x(k-1)-\mathrm{\α}\tilde{x}\left(k\right);---\left(1\right)$

the threshold period is set to P (min). Wherein x is a measure of the blood glucose signal,is the filtered value of, alpha, beta₁and beta₂Is the filter parameter and k is the sample point. The filtered output is therefore a linear combination of the historical filtered values, the historical and current measured values. According to the actual knowledge of blood sugar dynamics, the value of P is generally 40-80 min.

(2.2) for the data x collected in step 1^T(1 × Z) filtering with a first order Butterworth low pass filter to isolate low frequency blood glucose data x^LT(1 × Z) and high frequency blood glucose data x^HT(1 × Z.) the high frequency blood glucose data is considered noise and can be removed by:

x^LT(1×Z)＝x^T(1×Z)-x^HT(1×Z)；(2)

where superscript L, H represents low and high frequency data, respectively.

And step 3: acquiring a predictive variable matrix and a response matrix:

sliding a one-dimensional sliding window with the length of K sampling points over x^LT(1 × Z), moving one sampling point each time, moving for Z-K +1 times, taking the data in the sliding window each time as a new row vector, and combining the data into a twoDimension data matrix X^L(N × K), where N ═ Z-K +1, K ═ PL + H (PH/5), PL (PredictorLength) represents the length of the prediction variable, H represents the prediction step size, PH (PredictionHorizon) represents the prediction interval, and H and PH have different measurement units, but both represent how many steps (typically 5min per step) of blood glucose are predicted, and since the prediction step size is 5 minutes per step, PH ═ 5 × H.

And 4, step 4: blood glucose prediction modeling based on band separation:

the prediction model is established by two methods, one is the existing least square-based method, and the other is a latent variable-based modeling method. The method comprises the following specific steps:

(4.1) selecting a prediction variable and an output variable: selecting according to the value of PLX ^L(NxK) preceding PL columns as the prediction matrix X^L(NxJ), J ═ PL; then X^Lcolumn (NxK) PL + PH as output variable y^L(N×1)。

(4.2.) if the blood glucose prediction modeling is performed by a least square based method, a low frequency autoregressive prediction model is established by a latent variable method from the low frequency blood glucose signal:

$y^L=X^L{\mathrm{\θ}}^L+f={\hat{y}}^L+f;---\left(3\right)$

wherein,to obtain low frequency blood glucose prediction values，θ^L(J × 1) is a regression coefficient, and f (N × 1) is a model error.

If the blood sugar prediction modeling is carried out by a method based on latent variables, an autoregressive prediction model is established by adopting the latent variable method according to the low-frequency blood sugar signal:

the autoregressive low-frequency prediction model based on the latent variables is modeled by a feature extraction method of partial least square-typical correlation analysis. The preliminary latent variable group is extracted by partial least squares, and because the latent variable extracted by the partial least squares method cannot ensure the close correlation between the latent variable and the response variable, the latent variable is post-processed by typical correlation analysis, so that the part of the latent variable which is closely related to the corresponding variable is extracted for regression modeling and prediction. The method comprises the following steps:

(4.2.1) data preprocessing

Combining the predictive variable matrix and the response matrix of all individuals together, and aiming at the variable x of any point in the combined predictive variable matrix and response matrix_i,jAnd performing global normalization processing of subtracting the mean value of the variable and dividing the mean value by the standard deviation, wherein the subscript i represents the batch, j represents the variable, and the calculation formula of the normalization processing is as follows:

$x_{i,j}=\frac{x_{i,j}-{\stackrel{\‾}{x}}_j}{s_j};---\left(4\right)$

wherein:is the mean, s, of any column of the combined matrix_jIs the standard deviation of the corresponding column. The calculation formula is as follows:

$\begin{array}{c}{\stackrel{\‾}{x}}_j=\frac{1}{M}\underset{i=1}{\overset{M}{\Σ}}{x}_{i,j}\\ s_j=\sqrt{\underset{i=1}{\overset{M}{\Σ}}{(x_{i,j}-{\stackrel{\‾}{x}}_j)}^2/(M-1)};\end{array}---\left(5\right)$

wherein: m is the total number of samples in the combined matrix.

(4.2.2) extracting the latent variable group T by using partial least squares:

T＝X^LR

；(6)

R＝W(P^TW)^-1

where T is a latent variable matrix composed of a plurality of partial least squares latent variables, R is a coefficient matrix of the partial least squares method, W is a weight matrix found by partial least squares, and P is a load matrix corresponding to T.

(4.2.3) post-processing the partial least square latent variable by using typical correlation analysis to obtain a final latent variable u:

u＝Tv；(7)

where v is the weight vector corresponding to T. Because y is a univariate output response, only one latent variable needs to be extracted finally due to the characteristics of the type correlation analysis method.

It should be noted that the weight matrix of the partial least squares and correlation analysis method can be obtained by solving the eigenvector of the specific matrix, and is an effective statistical analysis method for analyzing the correlation between data variables.

(4.2.4) establishing a low-frequency autoregressive model:

and (3) solving a regression coefficient q between the latent variable and the response variable by using a least square method:

q＝(u^Tu)^-1u^Ty^L；(8)

therefore, the autoregressive model finally established based on the latent variable method is as follows:

$\begin{array}{c}{\hat{y}}^L=X^L{\mathrm{\σ}}^L\\ {\mathrm{\σ}}^L=Rvq\end{array};---\left(9\right)$

wherein σ^LIs a prediction model regression coefficient obtained by combining two methods of partial least squares and typical correlation analysis,for obtaining low-frequency blood sugar pre-treatmentAnd (6) measuring.

After the prediction models established for different objects are obtained, the universality of the prediction models needs to be verified. Verifying the universality of the model, namely verifying whether the potential blood sugar time sequence dynamics of different individuals have similar autocorrelation; whether the model established for any individual can be applied to other individuals for real-time blood glucose prediction. Therefore, it is necessary to compare the prediction accuracy of two models for any one object, one model created using the same individual information (the same individual model), and the other model created using the other individual information (different individual models). Here, we consider two identical individual models, the first one being an individual prediction model built based on low-frequency blood glucose signals obtained by frequency band separation, and the second one being an individual prediction model built based on raw or full-frequency blood glucose measurement signals. Comparison of these two identical individual models will show whether the removal of the high frequency signal affects the prediction accuracy.

The prediction modeling based on the original or full-frequency blood glucose signal comprises the following specific steps:

for the original blood glucose signal, steps 3 and 4 are performed directly without performing the band separation of step 2. Thus, the modeling herein utilizes the full range of blood glucose signals.

Therefore, the autoregressive model finally established by the least square method based on the original or full-frequency blood glucose signal is as follows:

$\hat{y}=X\mathrm{\θ};---\left(10\right)$

where θ is a regression coefficient of the prediction model obtained by combining two methods of partial least squares and canonical correlation analysis,obtaining the predicted value of blood sugar.

In a similar way, the autoregressive model finally established by adopting a latent variable method based on the original or full-frequency blood glucose signal is as follows:

$\begin{array}{c}\hat{y}=X\mathrm{\σ}\\ \mathrm{\σ}=Rvq\end{array};---\left(11\right)$

wherein σ is a regression coefficient of the prediction model obtained by combining two methods of partial least squares and typical correlation analysis,obtaining the predicted value of blood sugar.

Studies on individual models show that at PL <7, the prediction accuracy increases with increasing PL; and with the increase of PH, the prediction accuracy of the model shows a descending trend, and the future blood sugar prediction is more meaningful when PH is selected for 30-60 min generally, so that certain prediction accuracy and reliability can be ensured. We next verified the generality of the model for PL 7 and PH 30 min. The method comprises the following specific steps:

(I) blood glucose prediction using the same individual model

(a) And (3) for each object, after the blood glucose signal is processed by adopting the frequency band separation method in the step (2), calling the first same individual model to predict the blood glucose.

Firstly, calling a low-frequency autoregressive model established based on least square to predict blood sugar to obtain a predicted valueWherein, theta^LThe regression coefficient vector previously found from the training data.

Then, calling a low-frequency autoregressive model established based on latent variables to predict blood sugar to obtain a predicted valueWherein σ^LIs a regression coefficient vector calculated by a partial least squares-typical correlation analysis method.

(b) For each subject, a second identical individual model is invoked for blood glucose prediction against the raw blood glucose measurement signal.

Firstly, calling an autoregressive model established based on least square to predict blood sugar to obtain a predicted valueWhere θ is the regression coefficient vector previously found from the training data.

Then, calling an autoregressive model established based on latent variables to predict blood sugar to obtain a predicted valueWhere σ is a regression coefficient vector obtained by a partial least squares-typical correlation analysis method.

(II) blood glucose prediction using different individual models

And calling different individual models for each object to predict the blood sugar. Here we can use 26 different individual models for each individual to make blood glucose predictions.

Firstly, calling a low-frequency autoregressive model based on least square to predict blood sugar to obtain a predicted value ${\hat{y}}^L=x^{LT}\mathrm{\&theta;}.$

Then, calling a low-frequency autoregressive model based on latent variables to predict blood sugar to obtain a predicted value ${\hat{y}}^L=x^{LT}\mathrm{\&sigma;}.$

(III) comparing the prediction accuracy of the two models

First, after blood glucose prediction was performed for the same individual using the methods described in (6.1) and (6.2), the RMSE (root mean square error) performance index was calculated as follows:

$RMSE=\sqrt{\frac{1}{N}\underset{i\&Element;N}{\&Sigma;}{\left(y\right(i)-{\hat{y}}^{*}(i\left)\right)}^2};---\left(13\right)$

where y (i) represents the original blood glucose measurement,representing the predicted blood glucose results obtained using different models, which may be low-frequency predicted blood glucose values obtained using a first same individual model or different individual modelsOr the full-frequency predicted value obtained by adopting the second same individual modelN is the number of samples. It should be noted that, when blood glucose is predicted by using different individual models, different blood glucose prediction results are averaged, and the average value is used as a prediction result to be compared with a reference amount. Meanwhile, when measuring the prediction accuracy, the original blood sugar measurement value is required to be used as a reference quantity, and the prediction error is the comparison between the blood sugar prediction result and the original blood sugar measurement value.

Then, based on the RMSE indices, the prediction accuracy of the two models (same individual model versus different individual model) was compared using paired t-hypothesis testing.

The results (as shown in table 1) show that the prediction accuracy of the autoregressive model established by the latent variable method is better than that of the least square method. In addition, the blood glucose prediction results obtained based on two same individual models have statistically similar prediction accuracy with blood glucose prediction based on different individual models by using the same modeling method. Therefore, the model established for any individual can be applied to other individuals to predict blood sugar in real time, and the latent variable and the least square method do not influence the universality. And establishing a low-frequency autoregressive prediction model based on frequency band separation, namely the universal low-frequency model. In addition, since the prediction accuracy obtained using two identical individual models is similar, it is shown that removing the high frequency signal at a certain cut-off frequency has no effect on the blood glucose prediction.

Table 1 comparison of blood glucose predictions obtained using different least squares/latent variable models for two groups of individuals, group 1(12 subjects) and group 2(14 subjects) (RMSE (mg/dL) (MEAN ± MAD); MEAN is the MEAN absolute deviation.)

In addition, the research on the individual model shows that the selection of P (threshold value of frequency band separation), PL (prediction variable length) and PH (prediction interval) all have influence on the accuracy of the model. The invention combines two groups of individual information to analyze and research the selection of P, PL and PH.

(a) Selecting P: first, fig. 2 shows an effect diagram obtained by band separation, taking P as 60min as an example. As can be seen from the graph, the low-frequency data can reflect the general change trend and the curve is smoother, which also shows that certain noise information of the blood glucose signal can be removed by adopting a frequency band separation method. Second, we analyze the effect of the choice of P on the model accuracy. By r_m,nThe RMSE values representing the prediction of blood glucose in subject m from a generic low frequency model established from the subject n observations are shown. RMSE mean r of object m_mCalculated by averaging the results of the other 25 subjects:these calculations are repeated at different threshold periods P. R of test object m_mThe values and threshold periods may be arranged in a vector because the RMSE values for different objects have different ranges, r_mValue is normalizedAnd is a percentage between 0 and 100% for purposes of the drawing. R of object m transformed_mTo a standard value ofIs defined as:26 individuals of groups 1, 2 served as test subjects for the generic low frequency model. In the range of 20min<P<80min, from the paired t hypothesis test (α ═ 0.05), it can be seen that the accuracy of the generic low frequency model is not statistically superior for a particular value of P.

(b) Selecting PL: for each group of subjects, the effect of the prediction length PL on the prediction accuracy for a certain PH (here exemplified by 60 min) is assessed for all 26 subjects. The general low frequency model for each subject was determined and the average was then used as the test model for each subject for the 60min blood glucose prediction for this set of different subjects. The MEAN (MEAN) and MEAN Absolute Deviation (MAD) of the RMSE indices were averaged across each group of test subjects. RMSE was calculated for the predicted results of the two groups of test subjects at different PLs, the results are shown in fig. 3 and 4, and the prediction accuracy of the model increased with increasing PL until PL > 7.

(c) Selecting PH: similar to the choice of PL, the impact of PH on the accuracy of prediction for a certain PL (here PL ═ 7 is taken as an example) is rated across all 26 subjects. The general low frequency model for each subject is determined, and then the average value of the general low frequency model is used as the test model of each subject and applied to the blood glucose prediction of the different subjects in the group. The mean and median mean deviation of the RMSE index was averaged for each group of test subjects. The RMSE of the predicted results of the two groups of test subjects at different PHs was calculated, and the results are shown in fig. 5 and 6, where the model prediction accuracy decreased with increasing PH. Therefore, future blood sugar prediction is meaningful when PH is selected to be 30-60 min generally, on one hand, enough external input action time can be guaranteed to eliminate future abnormal blood sugar, and on the other hand, certain prediction accuracy and reliability can be guaranteed.

In order to show the prediction accuracy of the general low-frequency model established by the method of the invention more clearly, we compare the prediction accuracy with the prediction value of the SD model and the real measurement value, as shown in FIG. 7, FIG. 8, FIG. 9 and FIG. 10.

And 5: on-line prediction based on a general low-frequency model:

based on the 4 steps, the established general prediction model can be called to predict the future blood sugar value of any individual. The method is completed by the following steps:

(5.1) in online prediction, any new data of an individual is collectedLast (subscript new stands for new sample, J ═ PL), pairPerforming the band separation process as described in step 2 to obtain

(5.2) calling a general low-frequency model for online prediction:

(5.2.1) calling a general low-frequency model built based on least square to perform online prediction:

for each new test dataThe predicted value after PH sampling points is generated by the following steps

${\hat{y}}_{new}^L=x_{new}^{LT}{\mathrm{\&theta;}}^L;---\left(14\right)$

Wherein, theta^LThe regression coefficient vector previously found from the training data.

(5.2.2) calling a universal low-frequency model built based on latent variables to perform online prediction:

for each new test dataThe predicted value after PH sampling points are obtained through calculation according to the following steps

${\hat{y}}_{new}^L=x_{new}^{LT}{\mathrm{\&sigma;}}^L;---\left(15\right)$

Wherein σ^LIs a regression coefficient vector calculated by a partial least squares-typical correlation analysis method.

(5.3) measurement of prediction accuracy

The obtained quality prediction result is compared with the actual measurement value. After a series of new prediction signals are obtained, the blood glucose prediction accuracy for the new samples can be calculated according to the RMSE (root mean square error) performance indicator defined in equation (13).

Claims (3)

1. A method for universal blood glucose prediction based on band separation and data modeling, the method comprising the steps of:

step 1: modeling blood sugar signal preprocessing: combining individual subcutaneous blood glucose signals obtained with a certain sampling period delta t into one-dimensional time sequence data x^T(1 × Z), wherein x is the measured value of the blood sugar signal, Z is the sampling number, and the spike noise is removed;

step 2: separation of blood glucose signal frequency bands: analyzing the change rule of the dynamics of the blood sugar in different high and low frequency bands, distinguishing a key frequency band from a secondary frequency band, and determining the optimal threshold value for dividing the frequency bands; performing frequency band separation on the blood glucose signal by adopting a Butterworth low-pass filter according to a defined separation threshold value;

and step 3: acquiring a predictive variable matrix and a response matrix: sliding a one-dimensional sliding window with the length of K sampling points over x^LT(1 × Z), moving a sampling point each time for Z-K +1 times, taking the data in the sliding window each time as a new row vector, and combining the new row vector into a two-dimensional data matrix X^L(NxK), wherein N is Z-K +1, K is PL + H (PH/5), PL represents the length of a prediction variable, H represents a prediction step length, PH represents a prediction interval, and H and PH have different measurement units and respectively represent the predicted blood sugar after how many steps, and PH is 5 × H because the prediction step length is 5 minutes per step;

and 4, step 4: blood glucose prediction modeling based on band separation: performing blood glucose prediction modeling by a latent variable-based method;

and 5: predicting the future blood sugar value of any individual according to the blood sugar prediction model established in the step 4; the method is completed by the following steps:

(5.1) in online prediction, any new data of an individual is collectedThen, toPerforming the band separation process as described in step 2 to obtainWhere the subscript new represents the new sample, J ═ PL;

(5.2) calling a general low-frequency model for online prediction:

(5.2.1) calling a general low-frequency model built based on least square to perform online prediction: for each new test dataThe predicted value after PH sampling points is generated by the following steps

${\hat{y}}_{new}^L=X_{new}^{LT}{\mathrm{\&theta;}}^L$

Wherein, theta^LThe regression coefficient vector is obtained according to the training data;

(5.2.2) calling a universal low-frequency model built based on latent variables to perform online prediction:

for each new test dataThe predicted value after PH sampling points are obtained through calculation according to the following steps

${\hat{y}}_{new}^L=x_{new}^{LT}{\mathrm{\&sigma;}}^L;$

Wherein σ^LIs a regression coefficient vector calculated by a partial least square-typical correlation analysis method;

(5.3) measuring the prediction accuracy: comparing the obtained quality prediction result with an actual measurement value; after obtaining a new series of prediction signals, the accuracy of the blood glucose prediction for the new sample is calculated according to the RMSE performance index defined in the following equation, where RMSE is the root mean square error:

$RMSE=\sqrt{\frac{1}{N}\underset{i\&Element;N}{\&Sigma;}{\left(y\right(i)-{\hat{y}}^{*}(i\left)\right)}^2};$

wherein y (i) represents the original blood glucose measurement,representing the predicted blood sugar value obtained by different models and using the first same individual model or different individual modelsOr the full-frequency predicted value obtained by adopting the second same individual modelN is the number of samples.

2. The method for universal blood glucose prediction based on band separation and data modeling according to claim 1, wherein the step 2 comprises the sub-steps of:

(2.1) selecting a first order low-pass butterworth filter of the form:

$\tilde{x}\left(k\right)={\mathrm{\&beta;}}_{1}x\left(k\right)+{\mathrm{\&beta;}}_{2}x(k-1)-\mathrm{\&alpha;}\tilde{x}\left(k\right);$

setting the threshold period as P (min); wherein x is a measure of the blood glucose signal,is the filtered value of, alpha, beta₁and beta₂Is a filtering parameter, k is a sampling point; the filtering output is a linear combination of the historical filtering value, the historical and current measurement values; according to the actual knowledge of the dynamics of blood sugar, the value of P is generally 40-80 min; (2.2) for the data x collected in step 1^T(1 × Z) filtering with a first order Butterworth low pass filter to isolate low frequency blood glucose data x^LT(1 × Z) and high frequency blood glucose data x^HT(1 × Z), and removing the high-frequency blood sugar data which is considered as noise, wherein the steps are as follows:

x^LT(1×Z)＝x^T(1×Z)-x^HT(1×Z)；

where superscript L, H represents low and high frequency data, respectively.

3. The universal blood glucose prediction method based on frequency band separation and data modeling as claimed in claim 1, wherein the step 4 employs a latent variable based method for blood glucose prediction modeling; the method comprises the following specific steps:

(4.1) selecting a prediction variable and an output variable: selecting according to the value of PLX ^L(NxK) preceding PL columns as the prediction matrix X^L(NxJ), J ═ PL; then X^Lcolumn (NxK) PL + PH as output variable y^L(N×1)；

(4.2) establishing an autoregressive prediction model by adopting a latent variable method according to the low-frequency blood glucose signal: an autoregressive low-frequency prediction model based on latent variables is modeled by adopting a feature extraction method of partial least square-canonical correlation analysis; firstly, extracting a preliminary latent variable group by using partial least squares, and extracting parts closely related to corresponding variables for regression modeling and prediction by using typical correlation analysis because the latent variables extracted by using the partial least squares method cannot ensure the close correlation between the latent variables and response variables; the method comprises the following steps:

(4.2.1) data preprocessing

Combining the predictive variable matrix and the response matrix of all individuals together, and aiming at the variable x of any point in the combined predictive variable matrix and response matrix_i,jAnd performing global normalization processing of subtracting the mean value of the variable and dividing the mean value by the standard deviation, wherein the subscript i represents the batch, j represents the variable, and the calculation formula of the normalization processing is as follows:

$x_{i,j}=\frac{x_{i,j}-{\stackrel{\&OverBar;}{x}}_j}{s_j};$

wherein:is the mean, s, of any column of the combined matrix_jIs the standard deviation of the corresponding column; the calculation formula is as follows:

${\stackrel{\&OverBar;}{x}}_j=\frac{1}{M}\underset{i=1}{\overset{M}{\&Sigma;}}{x}_{i,j}$

$s_j=\sqrt{\underset{i=1}{\overset{M}{\&Sigma;}}{(x_{i,j}-{\stackrel{\&OverBar;}{x}}_j)}^2/(M-1)};$

wherein: m is the total number of samples in the combined matrix;

(4.2.2) extracting the latent variable group T by using partial least squares:

T＝X^LR

；

R＝W(P^TW)^-1

wherein T is a latent variable matrix composed of a plurality of partial least squares latent variables, R is a coefficient matrix of a partial least squares method, W is a weight matrix found by partial least squares, and P is a load matrix corresponding to T; (4.2.3) post-processing the partial least square latent variable by using typical correlation analysis to obtain a final latent variable u:

u＝Tv；

where v is a weight vector corresponding to T; because y is a univariate output response, only one latent variable needs to be extracted finally due to the characteristics of a typical correlation analysis method;

(4.2.4) establishing a low-frequency autoregressive model:

and (3) solving a regression coefficient q between the latent variable and the response variable by using a least square method:

q＝(u^Tu)^-1u^Ty^L；

therefore, the autoregressive model finally established based on the latent variable method is as follows:

；

σ^L＝Rvq

wherein σ^LIs a prediction model regression coefficient obtained by combining two methods of partial least squares and typical correlation analysis,the low-frequency blood sugar predicted value is obtained.