CN103310113A - Universal blood glucose prediction method based on frequency band separation and data modeling - Google Patents

Abstract

The invention discloses a universal blood glucose prediction method based on frequency band separation and data modeling, which comprises the steps that a human body subcutaneous blood glucose measurement signal is analyzed; the latent timing sequence dynamic characteristic of the human body subcutaneous blood glucose measurement signal is extracted; a frequency band separation threshold is defined; the subcutaneous blood glucose measurement signal is divided into a high frequency band and a low frequency band; timing sequence autocorrelation of a low frequency blood glucose signal is analyzed; and an autoegression blood glucose prediction model is established. According to the universal blood glucose prediction method, re-modeling for a new object is not required after the sufficient blood glucose measurement signals are acquired, and real-time blood glucose prediction can be performed by directly calling a prediction model of other individual, so that the modeling working capacity and the complexity are simplified greatly; the modeling cost can be lowered greatly; the prediction precision is improved due to the fact that a universal model adopts a method based on frequency band separation and latent variable modeling; and the universal blood glucose prediction method is easy to implement, and indicates a new direction for research of a blood glucose prediction modeling method.

Description

A kind of general blood sugar Forecasting Methodology based on frequency band separation and data modeling

Technical field

The invention belongs to blood glucose level data analysis and prediction research field, particularly relate to a kind of general blood sugar Forecasting Methodology based on frequency band separation and data modeling.

Background technology

A distinguishing feature of blood sugar for human body level is exactly time variation, is embodied in the clock signal measured value and has significant auto-correlation relation.For the blood sugar signal analyze, modeling, can extract its sequential correlation properties, obtain following change of blood sugar situation according to historical blood sugar dynamic.Foreign scholar Bremer in 1999 and Gough propose first the blood sugar time series data and have a kind of potential dependency structure, can be described by a kind of simple linear dynamic model.At present, the method that based on data drives is adopted in the foundation of blood sugar forecast model more.Already present forecast model can be divided into linearity (take based on the autoregressive model of least square as its Typical Representative) and non-linear (take neural net method as its Typical Representative) two classes.Linear model is applied widely owing to its simple model structure and algorithm.The modeling method of comparative maturity comprises autoregression (autoregressive, AR), exciter response (impulse-response, IR) etc.Autoregression (AR) method is the Data Modeling Method of comparative maturity, and it only utilizes the information of blood sugar itself, and the linear combination by historical glucose data obtains following blood sugar predicted value.Yet traditional AR model major defect has 2 points: (a) directly utilize least square the most basic this discrimination method match blood sugar projected relationship for measurement data, the defective that can't avoid the method itself to have can not obtain satisfied precision of prediction; (b) do not analyze in advance for the blood sugar dynamic of Different Individual, directly set up different forecast model (being the individuation model) and be used for on-line prediction, this will expend a large amount of manpower and materials.

Summary of the invention

The object of the invention is to the deficiency for existing blood sugar Forecasting Methodology, a kind of general blood sugar Forecasting Methodology based on frequency band separation and data modeling is provided.

The objective of the invention is to be achieved through the following technical solutions: a kind of general blood sugar Forecasting Methodology based on frequency band separation and data modeling, the method may further comprise the steps:

Step 1: modeling blood sugar Signal Pretreatment: will become one dimension time series data x with the individual subcutaneous blood sugar signal combination that certain sampling period Δ t obtains ^T(1 * Z), wherein, x is the measured value of blood sugar signal, and Z is number of samples, removes spike noise wherein.The sequential correlativity and the dynamic-change information that have comprised the blood sugar signal in this one dimension time series data.

Step 2: the blood sugar signal band separates: analyze the Changing Pattern of dynamic in high and low different frequency range of blood sugar, distinguish critical frequencies section and less important frequency band, determine the optimal threshold that frequency band is divided.Separation threshold value according to definition adopts Butterworth LPF that the blood sugar signal is carried out the frequency band separation.

Step 3: obtain predictive variable matrix and response matrix: the one dimension moving window that is K sampled point with a length slips over x ^LT(1 * Z), mobile sampled point is total to mobile Z-K+1 time at every turn.Data in each moving window as a new row vector, then can be combined as a two-dimensional data matrix X ^L(N * K), wherein, N=Z-K+1, K=PL+H (PH/5).PL represents the length of predictive variable, and H represents prediction step, and PH represents forecast interval, and H is different with the PH linear module, but equal blood sugar after to be what go on foot of predicting of expression.Because prediction step is 5 minutes per steps, therefore PH=5 * H is arranged.

Step 4: predict modeling based on the blood sugar that frequency band separates: by carry out blood sugar prediction modeling based on the method for LV.

Step 5: the following blood glucose value of any individual is predicted according to the blood sugar forecast model that step 4 is set up.Finish by following steps:

(5.1) during on-line prediction, collecting any individual new data

(subscript n ew represents new samples, and is J=PL), right afterwards

Carry out obtaining such as the described frequency band separating treatment of step 2

(5.2) call general low frequency model and carry out on-line prediction:

(5.2.1) call the general low frequency model of setting up based on LS and carry out on-line prediction: for each new test data By the predicted value behind PH sampled point of following steps generation

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

Wherein, θ ^LThe regression coefficient vector of obtaining according to training data for the front.

(5.2.2) call the general low frequency model of setting up based on LV and carry out on-line prediction:

For each new test data

By following steps calculate to obtain behind PH the sampled point in advance

Measured value

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

Wherein, σ ^LIt is the regression coefficient vector that the PLS-CCA method is obtained.

(5.3) weigh precision of prediction: the prediction of quality result and the actual measured value that obtain are compared.After obtaining a series of new prediction signal, can be according to the blood sugar precision of prediction of the RMSE that defines in the following formula (root-mean-square error) performance Index Calculation for new samples:

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

Wherein, original blood sugar measured of y (i) expression,

The blood sugar that expression adopts different models to obtain predicts the outcome, the low frequency blood sugar predicted value that adopts the first same individual model or Different Individual model to try to achieve

The full range predicted value that perhaps adopts the second same individual model to try to achieve

N is sample number.

Compared with prior art, the invention has the beneficial effects as follows: general blood sugar Forecasting Methodology proposed by the invention need not to re-start modeling after abundant blood glucose measurement signal is obtained in wait for new object, carry out the real time blood sugar prediction but can directly call other individual forecast model, greatly simplify modeling workload and complexity, will greatly reduce the modeling cost; And universal model adopts the method based on frequency band separation and latent variable modeling, has improved precision of prediction.The present invention is easy to implement, predicts that for blood sugar the research of modeling method has indicated a new direction.

Description of drawings

Fig. 1 is general blood sugar forecast model modeling process flow diagram of the present invention;

Fig. 2 is the comparison diagram of height frequency range blood sugar signal after being separated with frequency band by the subcutaneous blood sugar signal of full range that CGM (continuous blood sugar monitor) provides;

What Fig. 3 was PL on general low frequency model (GL) blood sugar estimated performance affects figure (ordinate is the mean value of RMSE);

What Fig. 4 was PL on general low frequency model (GL) blood sugar estimated performance affects figure (ordinate is the MAD value of RMSE);

What Fig. 5 was PH on general low frequency model (GL) blood sugar estimated performance affects figure (ordinate is the mean value of RMSE);

What Fig. 6 was PH on general low frequency model (GL) blood sugar estimated performance affects figure (ordinate is the MAD value of RMSE);

Fig. 7 is the blood glucose measurement curve of general low frequency (GL) model and the comparison diagram of blood sugar prediction curve (analytic target is first group of the first blood sugar signal).

Fig. 8 is the blood glucose measurement curve of standard independent individual (SD) model and the comparison diagram of blood sugar prediction curve (analytic target is first group of the first blood sugar signal).

Fig. 9 is the blood glucose measurement curve of general low frequency (GL) model and the comparison diagram of blood sugar prediction curve (analytic target is second group of the first blood sugar signal).

Figure 10 is the blood glucose measurement curve of standard independent individual (SD) model and the comparison diagram of blood sugar prediction curve (analytic target is second group of the first blood sugar signal).

Embodiment

As shown in Figure 1, the present invention is based on the general blood sugar Forecasting Methodology of frequency band separation and data modeling, the method may further comprise the steps:

Step 1: modeling blood sugar Signal Pretreatment

Individual subcutaneous blood sugar signal (Δ t=5min here) for obtaining with certain sampling period Δ t can be combined into it one dimension time series data x ^T(1 * Z), wherein, x is the measured value of blood sugar signal, and Z is number of samples, removes spike noise wherein.The sequential correlativity and the dynamic-change information that have comprised the blood sugar signal in this one dimension time series data.In this example, we have the blood sugar clock signal that comes from two group objects, and the 1st group comprises 12 people, and the 2nd group comprises 14 people, two groups of totally 26 people.The blood sugar clock signal of each object comprises two or three days data.The data of first day are used for model validation, and remainder data is used for the checking of model performance.

Step 2: the blood sugar signal band separates

The Changing Pattern of dynamic in high and low different frequency range of this step Main Analysis blood sugar distinguished critical frequencies section and less important frequency band, determines the optimal threshold that frequency band is divided.Separation threshold value according to definition adopts Butterworth LPF that the blood sugar signal is carried out the frequency band separation.Fundamental purpose is to separate by frequency band, removes the noise effect of high band, keeps effective change of blood sugar information of low-frequency range.

(2.1) choose the following single order low pass Butterworth filter of form:

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

It is P (min) that its threshold values cycle is set.Wherein, x is the measured value of blood sugar signal,

Its filter value, α, β ₁And β ₂Be filtering parameter, k is sampled point.Therefore filtering output is historical filter value, the linear combination of history and current measurement value.According to the understanding of reality to the blood sugar dynamic, the P value is generally 40min～80min.(2.2) the data x to gathering in the step 1 ^T(1 * Z) carries out filtering with the single order Butterworth LPF, isolates low frequency blood glucose level data x ^LT(1 * Z) and high frequency blood glucose level data x ^HT(1 * Z).The high frequency blood glucose level data is considered to noise, can remove, and step is:

x ^LT(1×Z)=x ^T(1×Z)-x ^HT(1×Z)；(2)

Wherein, subscript L, H represent respectively low frequency and high-frequency data.

Step 3: obtain predictive variable matrix and response matrix:

The one dimension moving window that is K sampled point with a length slips over x ^LT(1 * Z), mobile sampled point is total to mobile Z-K+1 time at every turn.Data in each moving window as a new row vector, then can be combined as a two-dimensional data matrix X ^L(N * K), wherein, N=Z-K+1, K=PL+H (PH/5).PL (Predictor Length) represents the length of predictive variable, H represents prediction step, PH (Prediction Horizon) represents forecast interval, and H is different with the PH linear module, but equal blood sugar after to be what go on foot (general per step is 5min) of predicting of expression.Because prediction step among the present invention is 5 minutes per steps, therefore PH=5 * H is arranged.

Step 4: predict modeling based on the blood sugar that frequency band separates:

Forecast model is set up two kinds of methods, and a kind of is existing method based on LS (least square, least square), and another kind is based on the modeling method of LV (latent variable, latent variable).Concrete steps are as follows:

(4.1) choose predictive variable and output variable:, choose according to the value of PL X ^L(N * K) front PL is listed as prediction matrix X ^L(N * J), J=PL; X then ^L(N * K) PL+PH is listed as output variable y ^L(N * 1).

If (4.2.) by carry out blood sugar prediction modeling based on the method for LS, then adopt the LS method to set up low frequency AR forecast model according to low frequency blood sugar signal:

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

Wherein,

Be the low frequency blood sugar predicted value that obtains, θ ^L(J * 1) is regression coefficient, and f (N * 1) is model error.

If by carry out blood sugar prediction modeling based on the method for LV, adopt the LV method to set up the AR forecast model according to low frequency blood sugar signal:

AR low frequency forecast model based on LV adopts the feature extracting method of PLS-CCA (partial least square-canonicalcorrelation analysis, offset minimum binary-canonical correlation analysis) to carry out modeling.Extract preliminary latent variable group with PLS first, because the latent variable LVs that the PLS method is extracted can not guarantee the closely related relation of itself and response variable, with CCA it is carried out aftertreatment again, wherein carry out regression modeling and prediction with the closely-related part of relevant variable thereby extract.Step is as follows:

(4.2.1) data pre-service

Predictive variable matrix and response matrix that all are individual are grouped together, for the variable x of any point in the predictive variable matrix after the combination and the response matrix _{I, j}, to this variable subtract average, divided by the global criteria processing of standard deviation, wherein, subscript i representative batch, j represents variable, the computing formula of standardization is as follows:

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

Wherein:

The average of the arbitrary row of matrix after the combination, s _jIt is the standard deviation of respective column.Its computing formula is as follows:

${\stackrel{\‾}{x}}_j=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{x}_{i,j}$

$s_j=\sqrt{\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{(x_{i,j}-{\stackrel{\‾}{x}}_j)}^2/(M-1)};$ (5)

Wherein: M is the sample total in the matrix after the combination.

(4.2.2) utilize PLS to extract latent variable group T:

T＝X ^LR；

R＝W(P ^TW) ^-1 (6)

Wherein, T is the latent variable matrix that is made of a plurality of PLS latent variable, and R is the matrix of coefficients of PLS method, and W is the weight matrix of being obtained by PLS, and P is the load matrix corresponding to T.

(4.2.3) utilize CCA that the PLS latent variable is carried out aftertreatment, obtain final latent variable u:

u=Tv；(7)

Wherein, v is the weight vectors corresponding to T.Because y is single argument output response, because CCA method itself, therefore finally only need to extract a latent variable.

The weight matrix that it is pointed out that PLS and CCA method can be tried to achieve by the proper vector of asking for particular matrix, is the statistical analysis technique of correlationship between a kind of effective analysis data variable.

(4.2.4) set up low frequency AR model:

Ask for regression coefficient q between latent variable and response variable with least square method:

q=(u ^Tu) ^-1u ^Ty ^L；(8)

Therefore the AR model of finally setting up based on the LV method is:

${\hat{y}}^L=X^L{\mathrm{\σ}}^L$ ；(9)

σ ^L=Rvq

Wherein, σ ^LTo be combined the forecast model regression coefficient that obtains by two kinds of methods of PLS and CCA,

Be the low frequency blood sugar predicted value that obtains.

After obtaining the forecast model of setting up for different objects, need to verify the versatility of forecast model.The checking of model commonality verifies that namely whether the potential blood sugar sequential dynamic of Different Individual has similar autocorrelation; Whether the model of setting up for any individual may be used on carrying out on other individualities the real time blood sugar prediction.Therefore need to contrast the precision of prediction of two kinds of models for any one object, a kind of is the model (same individual model) that utilizes same individual information to set up, and a kind of is the model (Different Individual model) that utilizes other individual information to set up.Here, we consider two kinds of same individual models, and the first is based on the individual forecast model that frequency band separates the low frequency blood sugar signal foundation that obtains, and the second is based on individual forecast model original or that full range blood glucose measurement signal is set up.The contrast of these two kinds of same individual models will illustrate whether the removal of high-frequency signal can the impact prediction precision.

Wherein, the prediction modeling concrete steps based on original or full range blood sugar signal are as follows:

For original blood sugar signal, need not to carry out the frequency band separation of step 2, directly execution in step 3 and 4.Therefore, modeling utilization here is the blood sugar signal of full range.

Therefore finally based on AR model original or that full range blood sugar signal adopts the LS method to set up be:

$\hat{y}=\mathrm{X\θ};$ (10)

Wherein, θ is combined the forecast model regression coefficient that obtains by two kinds of methods of PLS and CCA,

Be the blood sugar predicted value that obtains.

In like manner, finally based on AR model original or that full range blood sugar signal adopts the LV method to set up be:

$\hat{y}=\mathrm{X\σ}$ ；(11)

σ=Rvq

Wherein, σ is combined the forecast model regression coefficient that obtains by two kinds of methods of PLS and CCA,

Be the blood sugar predicted value that obtains.

To studies show that of individual body Model, in PL＜7 o'clock, precision of prediction increases with PL; And along with the increase of PH, the precision of prediction of model presents downtrending, when PH generally chooses 30～60min following blood sugar prediction comparatively meaningful, can guarantee certain precision of prediction and reliability.Below we verify the versatility of model for PL=7 and PH=30min.Concrete steps are as follows:

(I) utilize the same individual model to carry out the blood sugar prediction

(a) to each object, after the frequency band separation method in the employing step 2 is processed the blood sugar signal, call the first same individual model and carry out the blood sugar prediction.

At first, call the low frequency AR model of setting up based on LS and carry out the blood sugar prediction, obtain predicted value

Wherein, θ ^LThe regression coefficient vector of obtaining according to training data for the front.

Then, call the low frequency AR model of setting up based on LV and carry out the blood sugar prediction, obtain predicted value

Wherein, σ ^LIt is the regression coefficient vector that the PLS-CCA method is obtained.

(b) to each object, for original blood glucose measurement signal, call the second same individual model and carry out the blood sugar prediction.

At first, call the AR model of setting up based on LS and carry out the blood sugar prediction, obtain predicted value

Wherein, θ is the regression coefficient vector that the front is obtained according to training data.

Then, call the AR model of setting up based on LV and carry out the blood sugar prediction, obtain predicted value

Wherein, σ is the regression coefficient vector that the PLS-CCA method is obtained.

(II) utilize the Different Individual model to carry out the blood sugar prediction

To each object, call the Different Individual model and carry out the blood sugar prediction.Here we can adopt 26 Different Individual models that it is carried out the blood sugar prediction for each individuality.

At first, the low frequency AR model that calls based on LS carries out the blood sugar prediction, obtains predicted value

Then, the low frequency AR model that calls based on LV carries out the blood sugar prediction, obtains predicted value

(III) precision of prediction of two kinds of models of contrast

At first, carry out the blood sugar prediction for same body and function (6.1) and (6.2) described method after, calculate as follows RMSE (root-mean-square error) performance index:

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

Original blood sugar measured of y (i) expression wherein,

The blood sugar that expression adopts different models to obtain predicts the outcome, and can be the low frequency blood sugar predicted value that adopts the first same individual model or Different Individual model to try to achieve here The full range predicted value that perhaps adopts the second same individual model to try to achieve

N is sample number.Here it is pointed out that when adopting the Different Individual model to carry out the blood sugar prediction, the different blood sugar that obtain are predicted the outcome carry out first on average this mean value being compared with reference quantity as predicting the outcome.Simultaneously, when weighing precision of prediction, need original blood sugar measured as the reference amount, predicated error is that blood sugar predicts the outcome and original blood sugar measured contrast.

Then, based on the RMSE index, compare with the precision of prediction of the t hypothesis testing method that matches to two models (same individual model and Different Individual model).

Result's (as shown in table 1) shows, adopts the AR model prediction precision of LV method foundation to be better than the LS method.In addition, adopt identical modeling method, the blood sugar that obtains based on two kinds of same individual models predicts the outcome and carries out the blood sugar prediction based on the Different Individual model and have the similar precision of prediction of statistics.Therefore, any one individual model of setting up all be may be used on carrying out real-time blood sugar prediction on other individualities, and LV and LS method can not affect versatility.Separate the low frequency AR forecast model of setting up based on frequency band and be general low frequency model.In addition, since the precision of prediction that adopts two kinds of same individual models to obtain is similar, illustrate that the high-frequency signal of removing under certain cutoff frequency is on not impact of blood sugar prediction.

Table 1 is for the individual blood sugar that adopt the acquisition of Different L S/LV models of group 1 (12 object) and two groups of 2 (14 objects) of group contrast (RMSE (the mg/dL) (MEAN ± MAD) that predicts the outcome.MEAN is mean value, and MAD is mean absolute deviation.)

In addition, to studies show that of individual body Model, choosing the precision of model of P (threshold value that frequency band separates), PL (predictive variable length), PH (forecast interval) is all influential.The present invention combines two groups of individual informations choosing of P, PL, PH is analyzed and researched.

(a) choosing of P: at first, Fig. 2 has shown the design sketch that adopts after frequency band separates as an example of P=60min example.As seen from the figure, low-frequency data can reflect that overall variation tendency and curve are more level and smooth, and this has also embodied and has adopted the frequency band separation method can remove the certain noise information of blood sugar signal.Secondly, we analyze the impact on model accuracy chosen of P.Use r _{M, n}The general low frequency model that expression is set up according to object n observation data is carried out the RMSE value of object m blood sugar prediction.The RMSE average r of object m _mResult's calculating by average other 25 objects gets:

These calculating repeat under different threshold period P.The r of tested object m _mValue and threshold period can be arranged in a vector, because the RMSE value of different objects has different scopes, r _mValue is converted into 0 to 100% percentage in order to draw by standardization.The r that object m is converted _mStandard value

Be defined as: 26 individualities of group 1,2 are as the tested object of general low frequency model.At scope 20min＜P＜80min, by the t test of hypothesis (α=0.05) of pairing as can be known, the accuracy Bu Yin Teding P value of general low frequency model and superior to some extent on adding up.And along with P continues to increase, significantly reducing appears in the accuracy of universal model.

(b) choosing of PL: for every group objects, prediction length PL evaluates by all 26 objects the impact of certain PH (here take 60min as example) accuracy of forecast.Determine first the general low frequency model of each object, again with the test model of its mean value as each object, be applied to the 60min blood sugar prediction that this organizes different objects.The mean value of RMSE index (MEAN) and mean absolute deviation (MAD) on average get by every group of tested object.Calculate two groups of RMSE that tested object predicts the outcome under the different PL, its result as shown in Figure 3 and Figure 4, the precision of prediction of model increases until PL with PL〉7.

(c) choosing of PH: similar with choosing of PL, PH evaluates by all 26 objects for the impact of certain PL (here take PL=7 as example) accuracy of forecast.Determine first the general low frequency model of each object, again with the test model of its mean value as each object, be applied to the blood sugar prediction that this organizes different objects.The mean value of RMSE index and intermediate value mean deviation on average get by every group of tested object.Calculate two groups of RMSE that tested object predicts the outcome under the different PH, its result as shown in Figure 5 and Figure 6, the model prediction precision descends with the increase of PH.Therefore following blood sugar prediction was comparatively meaningful when PH generally chose 30～60min, can guarantee that the enough outside input action time to eliminate following abnormal plasma glucose, can guarantee certain precision of prediction and reliability on the other hand on the one hand.

For the general low frequency model precision of prediction that clearer demonstration the method for the invention is set up, we compare predicted value and the true measurement of itself and SD model, such as Fig. 7, Fig. 8, Fig. 9 and shown in Figure 10.

Step 5: based on the on-line prediction of general low frequency model:

Based on aforementioned 4 steps, we can call the general forecast model of setting up the following blood glucose value of any individual is predicted.Finish by following steps:

(5.1) during on-line prediction, collecting any individual new data

(subscript n ew represents new samples, and is J=PL), right afterwards

Carry out obtaining such as the described frequency band separating treatment of step 2

(5.2) call general low frequency model and carry out on-line prediction:

(5.2.1) call the general low frequency model of setting up based on LS and carry out on-line prediction:

For each new test data

By the predicted value behind PH sampled point of following steps generation

${\hat{y}}_{\mathrm{new}}^L=x_{\mathrm{new}}^{\mathrm{LT}}{\mathrm{\θ}}^L;$ (14)

Wherein, θ ^LThe regression coefficient vector of obtaining according to training data for the front.

(5.2.2) call the general low frequency model of setting up based on LV and carry out on-line prediction:

For each new test data

By the predicted value behind PH sampled point of following steps calculating acquisition

${\hat{y}}_{\mathrm{new}}^L=x_{\mathrm{new}}^{\mathrm{LT}}{\mathrm{\σ}}^L;$ (15)

Wherein, σ ^LIt is the regression coefficient vector that the PLS-CCA method is obtained.

(5.3) weigh precision of prediction

The prediction of quality result and the actual measured value that obtain are compared.After obtaining a series of new prediction signal, can be according to the blood sugar precision of prediction of RMSE (root-mean-square error) performance Index Calculation that defines in the formula (13) for new samples.

Claims (3)

1. the general blood sugar Forecasting Methodology based on frequency band separation and data modeling is characterized in that, the method may further comprise the steps:

Step 1: modeling blood sugar Signal Pretreatment: will become one dimension time series data x with the individual subcutaneous blood sugar signal combination that certain sampling period Δ t obtains ^T(1 * Z), wherein, x is the measured value of blood sugar signal, and Z is number of samples, removes spike noise wherein.The sequential correlativity and the dynamic-change information that have comprised the blood sugar signal in this one dimension time series data.

Step 2: the blood sugar signal band separates: analyze the Changing Pattern of dynamic in high and low different frequency range of blood sugar, distinguish critical frequencies section and less important frequency band, determine the optimal threshold that frequency band is divided.Separation threshold value according to definition adopts Butterworth LPF that the blood sugar signal is carried out the frequency band separation.

Step 3: obtain predictive variable matrix and response matrix: the one dimension moving window that is K sampled point with a length slips over x ^LT(1 * Z), mobile sampled point is total to mobile Z-K+1 time at every turn.Data in each moving window as a new row vector, then can be combined as a two-dimensional data matrix X ^L(N * K), wherein, N=Z-K+1, K=PL+H (PH/5).PL represents the length of predictive variable, and H represents prediction step, and PH represents forecast interval, and H is different with the PH linear module, but equal blood sugar after to be what go on foot of predicting of expression.Because prediction step is 5 minutes per steps, therefore PH=5 * H is arranged.

Step 4: predict modeling based on the blood sugar that frequency band separates: by carry out blood sugar prediction modeling based on the method for LV.Step 5: the following blood glucose value of any individual is predicted according to the blood sugar forecast model that step 4 is set up.Finish by following steps:

(5.1) during on-line prediction, collecting any individual new data

(subscript n ew represents new samples, and is J=PL), right afterwards

Carry out obtaining such as the described frequency band separating treatment of step 2

(5.2) call general low frequency model and carry out on-line prediction:

(5.2.1) call the general low frequency model of setting up based on LS and carry out on-line prediction: for each new test data

By the predicted value behind PH sampled point of following steps generation

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

Wherein, θ ^LThe regression coefficient vector of obtaining according to training data for the front.

(5.2.2) call the general low frequency model of setting up based on LV and carry out on-line prediction:

For each new test data

By the predicted value behind PH sampled point of following steps calculating acquisition

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

Wherein, σ ^LIt is the regression coefficient vector that the PLS-CCA method is obtained.

(5.3) weigh precision of prediction: the prediction of quality result and the actual measured value that obtain are compared.After obtaining a series of new prediction signal, can be according to the blood sugar precision of prediction of the RMSE that defines in the following formula (root-mean-square error) performance Index Calculation for new samples:

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

Wherein, original blood sugar measured of y (i) expression,

The blood sugar that expression adopts different models to obtain predicts the outcome, the low frequency blood sugar predicted value that adopts the first same individual model or Different Individual model to try to achieve

The full range predicted value that perhaps adopts the second same individual model to try to achieve

N is sample number.

2. describedly according to claim 1 separate and the general blood sugar Forecasting Methodology of data modeling based on frequency band, it is characterized in that, described step 2 comprises following substep:

(2.1) choose the following single order low pass Butterworth filter of form:

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

It is P (min) that its threshold values cycle is set.Wherein, x is the measured value of blood sugar signal,

Its filter value, α, β ₁And β ₂Be filtering parameter, k is sampled point.Therefore filtering output is historical filter value, the linear combination of history and current measurement value.According to the understanding of reality to the blood sugar dynamic, the P value is generally 40min～80min.

(2.2) the data x to gathering in the step 1 ^T(1 * Z) carries out filtering with the single order Butterworth LPF, isolates low frequency blood glucose level data x ^LT(1 * Z) and high frequency blood glucose level data x ^HT(1 * Z).The high frequency blood glucose level data is considered to noise, can remove, and step is:

x ^LT(1×Z)=x ^T(1×Z)-x ^HT(1×Z)；

Wherein, subscript L, H represent respectively low frequency and high-frequency data.

3. describedly according to claim 1 separate and the general blood sugar Forecasting Methodology of data modeling based on frequency band, it is characterized in that, described step 4 adopts carries out blood sugar based on the modeling method of LV and predicts modeling; Concrete steps are as follows:

(4.1) choose predictive variable and output variable:, choose according to the value of PL X ^L(N * K) front PL is listed as prediction matrix X ^L(N * J), J=PL; X then ^L(N * K) PL+PH is listed as output variable y ^L(N * 1).

(4.2) adopt the LV method to set up the AR forecast model according to low frequency blood sugar signal: the AR low frequency forecast model based on LV adopts the feature extracting method of PLS-CCA to carry out modeling.Extract preliminary latent variable group with PLS first, because the latent variable LVs that the PLS method is extracted can not guarantee the closely related relation of itself and response variable, with CCA it is carried out aftertreatment again, wherein carry out regression modeling and prediction with the closely-related part of relevant variable thereby extract.Step is as follows:

(4.2.1) data pre-service

Predictive variable matrix and response matrix that all are individual are grouped together, for the variable x of any point in the predictive variable matrix after the combination and the response matrix _{I, j}, to this variable subtract average, divided by the global criteria processing of standard deviation, wherein, subscript i representative batch, j represents variable, the computing formula of standardization is as follows:

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

Wherein:

The average of the arbitrary row of matrix after the combination, s _jIt is the standard deviation of respective column.Its computing formula is as follows:

${\stackrel{\‾}{x}}_j=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{x}_{i,j}$

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

Wherein: M is the sample total in the matrix after the combination.

(4.2.2) utilize PLS to extract latent variable group T:

T=X ^LR

R=W(P ^TW) ^-1；

Wherein, T is the latent variable matrix that is made of a plurality of PLS latent variable, and R is the matrix of coefficients of PLS method, and W is the weight matrix of being obtained by PLS, and P is the load matrix corresponding to T.

(4.2.3) utilize CCA that the PLS latent variable is carried out aftertreatment, obtain final latent variable u:

u=Tv；

Wherein, v is the weight vectors corresponding to T.Because y is single argument output response, because CCA method itself, therefore finally only need to extract a latent variable.

(4.2.4) set up low frequency AR model:

Ask for regression coefficient q between latent variable and response variable with least square method:

q=(u ^Tu) ^-1u ^Ty ^L；

Therefore the AR model of finally setting up based on the LV method is:

${\hat{y}}^L=X^L{\mathrm{\σ}}^L$ ；

σ ^L=Rvq

Wherein, σ ^LTo be combined the forecast model regression coefficient that obtains by two kinds of methods of PLS and CCA, Be the low frequency blood sugar predicted value that obtains.

Images