CN110575181A - Near-infrared spectroscopy non-invasive blood glucose detection network model training method - Google Patents

Abstract

The invention relates to a near-infrared spectrum non-invasive blood sugar detection network model training method, which belongs to the technical fields of physiological signal collection technology and digital signal analysis. Firstly, the BP artificial neural network is obtained by training the data of ambient temperature, ambient humidity, systolic blood pressure, diastolic blood pressure, pulse rate, body temperature, single-wavelength near-infrared absorbance and corresponding invasive blood sugar concentration, and then conducts sensitivity analysis on the basis of which to screen out Four variables including systolic blood pressure, pulse rate, body temperature and single-wavelength near-infrared absorbance are used as input variables of the NARX model, and the final NARX detection model is obtained through training; By collecting the above four variables and two invasive blood sugar concentration values, the subsequent near-infrared light non-invasive blood sugar concentration detection results can be obtained. Using this detection network model for near-infrared light non-invasive blood sugar detection has high detection accuracy and can be very good. meet clinical needs.

Description

Near-infrared spectroscopy non-invasive blood glucose detection network model training method

technical field

The invention belongs to the technical fields of physiological signal collection technology and digital signal analysis, and relates to a near-infrared spectrum non-invasive blood sugar detection network model training method.

Background technique

At present, the existing blood glucose detection methods are mainly divided into three categories: invasive, minimally invasive and non-invasive. Among them, invasive and minimally invasive detection methods include automatic biochemical instrument measurement method and rapid blood glucose meter measurement method, etc., which have high detection accuracy, but frequent detection will bring unnecessary trouble and mental stress to patients, and long-term trauma will cause The patient is left with pain and even a psychological shadow. If it is not handled properly, it will also leave the possibility of infection. Therefore, a non-invasive blood glucose concentration monitoring method has very important practical significance for diabetic patients, which can reduce the suffering of patients, and the drugs related to blood glucose dependence can be controlled more effectively and improve the quality of life of patients.

The main non-invasive blood glucose concentration monitoring methods are mainly divided into two categories: liquid collection methods and optical methods. The former includes transdermal collection of interstitial fluid and iontophoresis, while the latter includes mid-infrared spectroscopy, near-infrared spectroscopy, photoacoustic spectroscopy, Raman spectroscopy, optical rotation, and light scattering coefficient methods. Among the many non-invasive blood glucose measurement methods, near-infrared spectroscopy has become the most widely used method for human blood glucose concentration due to its deep penetration into the skin, relatively low equipment cost, relatively small light energy, and no harm to the human body. One of the most promising approaches in the field of noninvasive measurement research. This technology has been used for non-invasive detection of blood glucose concentration in fingers, lower lip, forearm, earlobe, tongue, cheek and other parts.

The theoretical basis for the non-invasive detection of blood glucose concentration by near-infrared spectroscopy is the Beer-Lambert law: Where A _λ is the absorbance of blood sugar at a specific wavelength λ, I ₀ (λ) is the transmitted light intensity, I(λ) is the transmitted light intensity after passing through human tissue, ε(λ) is the absorption coefficient, l is the optical path length, c is the concentration of blood sugar. Beer-Lambert's law applies to uniform non-scattering systems, requiring no interaction between light-absorbing particles.

A stable and reliable quantitative model is a key technology of near-infrared non-invasive blood glucose detection technology. Linear models such as Multiple Linear Regression (MLR), Principal Components Regression (PCR) and Partial Least Squares Regression (PLS) have been used in many studies. However, due to the complexity of blood components and the interaction between components, the linear method based on the Lambert-Beer law may not be able to fit the relationship between near-infrared absorbance and human blood glucose concentration well. On the other hand, the optical signal about blood glucose detected in vitro is very weak and is extremely susceptible to changes in human physiological parameters and in vitro environment. Such as temperature, including the body temperature of the human body during the measurement process and the influence of parameters such as ambient temperature, blood pressure, and ambient humidity. At the same time, near-infrared spectroscopy is also affected by blood volume pulsations throughout the cardiac cycle. However, most existing studies focus on the relationship between near-infrared absorption and blood glucose concentration, but do not consider their fluctuations with respect to time, as well as the influence of environmental factors and human physiological state on human blood glucose concentration.

Contents of the invention

In view of this, the purpose of the present invention is to provide a near-infrared spectrum non-invasive blood glucose detection network model training method.

To achieve the above object, the present invention provides the following technical solutions:

A near-infrared spectrum non-invasive blood glucose detection network model training method, the method includes the following steps:

1) For multiple designated individual subjects, the near-infrared absorbance of human blood glucose, invasive blood glucose concentration, systolic blood pressure, diastolic blood pressure and pulse rate are detected for each individual subject at different detection periods, as well as the ambient temperature and ambient humidity collection; the above 7 input parameters and invasive blood glucose concentration values detected by each individual subject every day are used as the sample data time series of the corresponding individual subject on the day, so as to obtain the sample data groups of multiple individual subjects; some of the individual subjects The sample data of the individual objects is used as the training sample data group, the sample data of a part of the individual objects is used as the verification sample data group, and the sample data of the remaining individual objects is used as the test sample data group;

2) Input the 7 input parameters and the invasive blood glucose concentration value in the training sample data group as input variables and target variables respectively into the BP artificial neural network, train the artificial neural network, and then obtain an artificial neural network with definite parameters and structure. network;

3) Sensitivity analysis is performed on the basis of the trained BP network; all samples are used for sensitivity analysis, and unimportant variables are excluded according to the analysis results until no variables can be excluded, and the remaining variables are the blood glucose Important variables with significant concentration effects;

4) Create a new sample data set from the filtered variables, wherein the sample data of a part of the individual subjects is used as the training sample data set, and the sample data of the remaining individual subjects is used as the test sample data set; the filtered variables and invasive blood glucose concentration values The input variables and target variables are respectively input into the NARX network, and the network is trained to obtain a detection model for near-infrared non-invasive blood glucose concentration detection.

Optionally, the step 2) is specifically:

21) Determine that the network structure includes three layers, an output layer, an output layer, and a hidden layer; the number of input neurons is 7, corresponding to the number of output variables, and the number of output neurons is 1, corresponding to the predicted value of blood sugar concentration; The number of neurons in the hidden layer is determined to be 15 according to the Kolmogorov formula;

22) Use the Levenberg-Marquardt algorithm to train the network; randomly select 60% of all samples according to the blood glucose concentration reference value as a training set, and train the network during the training process; 20% are used as a verification set to measure the network performance. Generalization ability and terminate the network training process; the remaining 20% is used as a test set to evaluate the prediction performance of the trained model.

Optionally, the step 4) is specifically:

41) Considering the clinical application scenario, for each user, the number of fingertip blood collection should be reduced as much as possible under the premise of ensuring better prediction accuracy, so the delay order d=2 is determined;

42) Also determine the number of neurons in the hidden layer according to the Kolmogorov formula;

43) Use the Levenberg-Marquardt algorithm to train the network; during the NARX network training process, the network adopts an open-loop mode, which provides a standard reference value for the past output value when predicting the current output, which is conducive to improving the accuracy of the model; training After that, the network switches to closed-loop mode for prediction in real problems.

4. The method for training a near-infrared spectrum non-invasive blood sugar detection network model according to claim 1, wherein the infrared spectrum is near-infrared light with a single wavelength of 1550nm.

Based on the near infrared spectrum non-invasive blood sugar detection method of described training method, this method comprises the following steps:

A) The obtained NARX detection model is used for near-infrared non-invasive blood glucose detection;

B) collecting the input parameters after screening the individual object to be tested, and obtaining the input sample data of the individual object to be tested;

C) Input the input sample data into the NARX detection model obtained by the above method to obtain the predicted value of the blood glucose concentration of the individual subject to be tested as the near-infrared non-invasive blood glucose detection result of the individual subject to be tested.

The beneficial effects of the present invention are:

1. Realize the application of near-infrared spectroscopy with few wavelengths, effectively extract blood sugar concentration, and avoid the use of near-infrared spectrometers. Diabetic patients or their family members do not need very complicated professional knowledge during use to use the detection system, which is conducive to the promotion of non-invasive blood glucose detectors at home.

2. Preliminary introduction of the physiological parameters of the environment and the subject’s body, taking into account the influence of these factors on the near-infrared absorbance during the detection process, will help reduce the influence of individual differences and environmental factors on the prediction results, and improve the accuracy of the prediction model accuracy and robustness.

3. The SA analysis method is used to analyze the initially introduced variables, and several key variables that have a greater impact on the model output results are screened out. On the one hand, it is beneficial to reduce the complexity of the model and improve the calculation speed; on the other hand, it eliminates redundant Residual variables can effectively prevent overfitting and improve the prediction accuracy and robustness of the model.

4. The fluctuation of human blood sugar concentration in a day has certain regularity. The NARX model not only has the simulation function of time series, but also can better describe the nonlinear relationship. Based on the results of input variable screening, the establishment of the NARX model can effectively use the time series fluctuation of blood glucose concentration and improve the prediction accuracy.

Other advantages, objects and features of the present invention will be set forth in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the investigation and research below, or can be obtained from It is taught in the practice of the present invention. The objects and other advantages of the invention may be realized and attained by the following specification.

Description of drawings

In order to make the purpose of the present invention, technical solutions and advantages clearer, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

Fig. 1 is a flowchart of the present invention;

Fig. 2 is a typical BP artificial neural network model topological structure diagram;

Figure 3 is a typical NARX model topology diagram;

Fig. 4 is the NARX detection network model of different individual objects in the embodiment on different days to carry out near-infrared spectrum non-invasive blood sugar detection results and deviation;

Fig. 5 is a Clark error grid diagram of the present invention using the NARX detection network model for near-infrared spectrum non-invasive blood glucose detection.

Detailed ways

Embodiments of the present invention are described below through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation modes, and various modifications or changes can be made to the details in this specification based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments are only schematically illustrating the basic concept of the present invention, and the following embodiments and the features in the embodiments can be combined with each other in the case of no conflict.

Wherein, the accompanying drawings are for illustrative purposes only, and represent only schematic diagrams, rather than physical drawings, and should not be construed as limiting the present invention; in order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, Enlargement or reduction does not represent the size of the actual product; for those skilled in the art, it is understandable that certain known structures and their descriptions in the drawings may be omitted.

In the drawings of the embodiments of the present invention, the same or similar symbols correspond to the same or similar components; , "front", "rear" and other indicated orientations or positional relationships are based on the orientations or positional relationships shown in the drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred devices or elements must It has a specific orientation, is constructed and operated in a specific orientation, so the terms describing the positional relationship in the drawings are for illustrative purposes only, and should not be construed as limiting the present invention. For those of ordinary skill in the art, the understanding of the specific meaning of the above terms.

In this embodiment, a plurality of volunteers are used as individual objects for training the detection network model. In this embodiment, a plurality of volunteers are used as individual objects for the detection network model training. The card is transmitted to the computer for superposition and average filtering processing to obtain the near-infrared spectrum of 1550nm single-wavelength near-infrared light.

During the specific collection, the sampling frequency of 1550nm near-infrared spectrum is 200Hz, continuous sampling is 20 seconds, and the near-infrared spectrum data in the middle 10 seconds are superimposed and averaged as the near-infrared spectrum data for the final human blood sugar detection.

Before data collection, volunteers washed their hands with water and soap at 20-30°C to keep their hands clean. Then wipe the measurement site with rubbing alcohol and wait for it to dry. In order to avoid the influence of finger structure differences, the spectral detection part is fixed as the left index finger. Each volunteer is asked to sit comfortably in a chair and place the left index finger into the test chamber of the optical subsystem. A fingertip is required to maintain slight contact with the collimator to avoid deformation. During the sampling process, the measuring position and the measuring pressure are kept as constant as possible. After the 1550nm near-infrared light signal is conditioned and amplified, it is transmitted to the computer through the data acquisition card for superposition and average filtering processing, and the near-infrared spectrum of the 1550nm single-wavelength near-infrared light is obtained. The sampling frequency is 200 Hz, the sampling is continuous for 20 seconds, and the near-infrared spectrum data in the middle 10 seconds are superimposed and averaged to obtain the near-infrared absorbance value at 1550 nm. The above near-infrared light signal acquisition operation was repeated three times, and the average value was used as the final experimental data. Then measure the corresponding blood glucose concentration reference value through the One Touch Ultra Easy blood glucose meter. Ambient temperature and humidity are obtained by digital thermometer and hygrometer. The systolic blood pressure, diastolic blood pressure and pulse rate of the volunteers were measured with an Omron electronic sphygmomanometer. The body temperature of the volunteers was measured by an infrared electronic thermometer. The measurement of the above environmental and human physiological parameters was repeated three times, and the average value was recorded as the final result.

In this embodiment, a total of 14 volunteers (9 females and 5 males, aged 22 to 25 years old) were recruited. Before the experiment was collected, the volunteers were informed of all possible risks in the experiment process, and the experimental data were collected after obtaining the consent of the volunteers. The experiment started at 8:30 and data collection was performed every 30 minutes. The end time of data collection for each volunteer varied on an individual basis. In addition, some volunteers participated in two days of data collection, and some volunteers only participated in one day. Table 1 shows the sample size of the experimental data (time series) collected per day for each volunteer. For example, the first volunteer collected a total of 11 samples on the first day. Finally, a total of 205 sample data were collected in this embodiment. Each sample data consists of ambient temperature (X ₁ ), ambient humidity (X ₂ ), systolic blood pressure (X ₃ ), diastolic blood pressure (X ₄ ), pulse rate (X ₅ ), body temperature (X ₆ ) and 1550nm near-infrared absorbance (X ₇ ) consisted of 7 input variables and the reference value of invasive blood glucose concentration (Y). The reference value range of the measured invasive blood glucose concentration was 4.2-10mmol/L, and the standard deviation was 1.35mmol/L.

Table 1 Experimental data information

Then, according to the detection network model training method of the present invention, the detection network model for near-infrared spectrum non-invasive blood sugar detection is trained, and the training process is as shown in Figure 1:

1) Obtain blood glucose sample data sets of the above-mentioned multiple individual subjects.

2) Input the 7 input parameters and the invasive blood glucose concentration value in the training sample data group as input variables and target variables respectively into the BP artificial neural network, train the artificial neural network, and then obtain an artificial neural network with definite parameters and structure. network;

21) The typical topology of BP neural network is shown in Figure 2. Determine that the network structure consists of three layers, the output layer, the output layer and a hidden layer; the number of input neurons is 7, corresponding to the number of output variables, and the number of output neurons is 1, corresponding to the predicted value of blood sugar concentration; implicit The number of layer neurons is determined to be 15 according to the Kolmogorov formula.

22) The network is trained using the Levenberg-Marquardt algorithm. 123 samples were randomly selected from all samples according to the blood glucose concentration reference value as the training set to train the network; 41 samples were used as the verification set to verify the generalization ability of the network and terminate the training process in time; 41 samples were used as the test set. to evaluate the performance of the trained model.

3) Sensitivity analysis is performed on the basis of the trained BP network. All samples are used for sensitivity analysis, and unimportant variables are excluded according to the analysis results until no variables can be excluded, and the remaining variables are important variables that have a significant impact on blood glucose concentration.

31) Firstly, the first round of screening is carried out, and the sensitivity analysis of each variable is carried out on the basis of the trained 7-variable BP network. The results are shown in the first round in Table 2. According to the discriminant criterion of Gap<0.5, in the current 7-variable BP network, the sensitivity of the output to the input variables X1 and X2 is relatively low, which will be excluded in the later model.

32) Then carry out the second round of screening, continue to build the BP network model from the remaining 5 variables, and conduct sensitivity analysis again. The results are shown in the second round in Table 2, and the variable X4 is excluded in the same way.

33) Carry out the third round of screening, continue to build the BP network model from the remaining 4 variables, and conduct sensitivity analysis. The results are shown in the third round in Table 2. Currently, the variables X3, X5, X6 and X7 are composed of In the BP network, the sensitivity of the output to each input variable is relatively high, so the variable screening process ends here, and it can be considered that the four variables (systolic blood pressure, pulse rate, body temperature, and 1550nm absorbance) are all the parameters of the blood glucose concentration prediction model. important variable.

Table 2 Sensitivity analysis results

4) Establish a new sample data set by the screened four variables (systolic blood pressure, pulse rate, body temperature, and 1550nm absorbance) and invasive blood glucose concentration reference values, wherein 90% of the sample data of individual subjects are used as training sample data sets, and the remaining The sample data of 10% of the individual subjects is used as the test sample data set. The filtered variables and invasive blood glucose concentration values were input into the NARX network as input variables and target variables, respectively, and the network was trained to obtain a detection model for near-infrared non-invasive blood glucose concentration detection.

41) Figure 3 is a typical topology of a NARX network. Considering the clinical application scenario, for each user, it is necessary to reduce the number of fingertip blood collection as much as possible under the premise of ensuring better prediction accuracy, so the delay order d=2 is determined.

42) Also determine the number of neurons in the hidden layer to be 9 according to the Kolmogorov formula.

43) The network is trained using the Levenberg-Marquardt algorithm. During the training process of the NARX network, the network adopts an open-loop mode, which provides a standard reference value for the past output value when predicting the current output, which is conducive to improving the accuracy of the model. After training, the network switches to closed-loop mode for prediction in practical problems.

44) Since the NARX model requires the data to have strict timing, the original order of the sample data is retained, and the collected sample data is divided into 10 groups of time series as a ten-fold cross-validation data set. Figure 4 shows the prediction results and deviations of NARX in this example using the ten-fold cross-validation method. The label format of the horizontal mark is "Sa-b-c", where a represents the serial number of the individual object to be tested, b represents the number of days of detection (mark of the detection period), and c represents the number of blood glucose sample data groups within the corresponding number of days (within the corresponding detection period). number mark; for example, "S1-1-01" indicates the first blood glucose sample data set detected by the first individual subject within the first day. It can be seen from Figure 3 that most of the predicted values are relatively close to the measured reference values.

Fig. 5 is the Clark error grid analysis diagram of the NARX detection network model of the present invention for the method of non-invasive blood glucose detection by near-infrared spectroscopy. The prediction results are distributed in areas A and B, which meet the clinical requirements. Among them, the proportion of points falling in area A is 90.27%, and the proportion of points falling in area B is 9.73%. It shows that the non-invasive blood glucose detection method based on NARX detection network model of the present invention has high detection accuracy.

In order to further illustrate the performance of the NARX detection network model of the present invention, the root mean square error (RMSE), the correlation coefficient (CORR) and the proportion of Clarke error grid analysis prediction results falling into each region are evaluated respectively. The parameters of the prediction results of each ten-fold cross-validation set are shown in Table 3. The root mean square error of the prediction results of the final overall cross-validation set is 0.72, and the correlation coefficient is 0.85. The above analysis shows that the NARX detection model of the present invention has high detection accuracy and robustness.

Table 3 Prediction results of NARX model ten-fold cross-validation set

To sum up, the present invention is a network model training method for near-infrared spectrum non-invasive blood glucose detection. First of all, according to the near-infrared absorption characteristics of human blood sugar, 1550nm near-infrared has the strongest absorption on human blood sugar, so the 1550nm near-infrared spectrum is determined as an independent variable, avoiding the use of near-infrared spectrometers, making the equipment more portable, easy to operate, and low in price . The structure of human tissue is complex, and the near-infrared absorbance and blood glucose concentration do not strictly conform to Lambert-Beer's law, which may have both linear and nonlinear relationships, and the daily fluctuation of human blood glucose concentration has certain regularity, and the NARX network model can both fit Linear and nonlinear relationships, but also has the simulation function of time series. Considering the influence of environmental factors and human physiological parameters on the near-infrared absorbance, ambient temperature, ambient humidity, systolic blood pressure, diastolic blood pressure, pulse rate and body temperature were added as input variables. In order to overcome information redundancy and model overfitting caused by too many independent variables, a sensitivity analysis method was used to screen the initially introduced seven input variables, and finally screened out systolic blood pressure, pulse rate, body temperature, and 1550nm near-infrared absorbance The four input variables are used as the input variables of NARX to obtain the NARX model. The results show that the NARX model proposed in this paper can meet the accuracy requirements of non-invasive blood glucose detection. On the one hand, it improves the accuracy of non-invasive blood glucose detection, and on the other hand, it improves the robustness and generalization ability of the model.

Finally, it is noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be carried out Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should be included in the scope of the claims of the present invention.

Claims (5)

1. the near infrared spectrum noninvasive blood glucose detection network model training method is characterized by comprising the following steps:

1) Aiming at a plurality of specified individual subjects, respectively detecting the near infrared light absorbance, invasive blood glucose concentration, systolic pressure, diastolic pressure and pulse rate of human blood glucose of each individual subject at different detection time periods, and acquiring environmental temperature and environmental humidity; taking the 7 input parameters and the invasive blood glucose concentration value detected by each individual object every day as a sample data time sequence of the corresponding individual object on the same day, thereby obtaining a sample data group of a plurality of individual objects; the sample data of a part of individual objects is used as a training sample data set, the sample data of a part of individual objects is used as a verification sample data set, and the sample data of the rest individual objects is used as a test sample data set;

2) Inputting 7 input parameters and invasive blood sugar concentration values in a training sample data set into a BP artificial neural network as input variables and target variables respectively, training the artificial neural network, and further obtaining an artificial neural network with determined parameters and structure;

3) Carrying out sensitivity analysis on the basis of the trained BP network; all samples are used for sensitivity analysis, unimportant variables are excluded according to analysis results until no variable can be excluded, and the reserved variables are important variables which obviously influence the blood glucose concentration;

4) Establishing a new sample data set by the screened variables, wherein the sample data of a part of individual objects is used as a training sample data set, and the sample data of the rest individual objects is used as a test sample data set; and inputting the screened variable and the invasive blood glucose concentration value into an NARX network as an input variable and a target variable respectively, training the network, and obtaining a detection model for near-infrared noninvasive blood glucose concentration detection.

2. The near infrared spectrum noninvasive blood glucose detection network model training method of claim 1, wherein the step 2) is specifically:

21) Determining that the network structure comprises three layers, namely an output layer, an output layer and a hidden layer; the number of input neurons is 7, the number of output neurons corresponds to the number of output variables, the number of output neurons is 1, and the number corresponds to a blood glucose concentration predicted value; the number of the neurons in the hidden layer is determined to be 15 according to a Kolmogorov formula;

22) Training the network by adopting a Levenberg-Marquardt algorithm; randomly selecting 60% of all samples according to the blood glucose concentration reference value as a training set, and training the network in the training process; 20% is used as a verification set for measuring the generalization capability of the network and terminating the network training process; the remaining 20% was used as a test set to evaluate the prediction performance of the trained model.

3. The near infrared spectrum noninvasive blood glucose detection network model training method of claim 1, wherein the step 4) is specifically:

41) Considering a clinical application scene, for each user, on the premise of ensuring better prediction accuracy, the number of times of blood sampling of a fingertip is reduced as much as possible, so that the delay order d is determined to be 2;

42) determining the number of neurons in an implied layer according to a Kolmogorov formula;

43) Training the network by adopting a Levenberg-Marquardt algorithm; in the NARX network training process, the network adopts an open-loop mode, and the past output value provided by the mode when the current output is predicted is a standard reference value, so that the accuracy of the model is improved; after training is finished, the network is converted into a closed-loop mode for prediction in practical problems.

4. the near infrared spectroscopy noninvasive blood glucose detection network model training method of claim 1, wherein the infrared spectrum is 1550nm single wavelength near infrared light.

5. The near infrared spectrum noninvasive blood glucose detection method based on the training method of any one of claims 1 to 4 is characterized by comprising the following steps:

A) The obtained NARX detection model is used for near-infrared noninvasive blood glucose detection;

B) acquiring input parameters after screening individual objects to be tested to obtain input sample data of the individual objects to be tested;

C) Inputting the input sample data into the NARX detection model obtained by the method to obtain the blood glucose concentration predicted value of the individual object to be detected, and taking the predicted value as the near-infrared noninvasive blood glucose detection result of the individual object to be detected.