Non-invasive blood glucose detection method and device based on ultra-wide near-infrared spectrum

Abstract

The application provides a non-invasive blood glucose detection method based on ultra-wide near-infrared spectroscopy, comprising: collecting spectral data of each finger of a test object in an ultra-wide near-infrared wave band range at different time periods, and synchronously collecting real blood glucose concentration data to construct a data set; constructing a physical model constrained neural network for extracting blood glucose concentration from near-infrared spectroscopy; training the neural network, introducing a physical constraint loss function based on the Kubelka-Munk theory, the Beer-Lambert law and a scattering model in the training process to obtain a trained neural network; and applying the trained neural network to predict blood glucose concentration by taking near-infrared spectroscopy data of any finger of a test object as input. The application solves the problem of insufficient blood glucose prediction accuracy and stability caused by the complex optical properties of human tissues and the interference of non-blood glucose components on spectral signals, and realizes high-precision non-invasive blood glucose detection.

Description

Technical Field

[0001] This invention belongs to the field of non-invasive blood glucose detection, specifically relating to a non-invasive blood glucose detection method and device based on ultrawide near-infrared spectroscopy. Background Technology

[0002] Diabetes mellitus is a common chronic metabolic disease characterized by persistently elevated blood glucose levels. Patients need to monitor their blood glucose levels long-term to guide their diet, exercise, and the use of insulin or other medications. Long-term control of blood glucose levels is crucial for preventing various acute and chronic complications of diabetes. Common diabetic complications include retinopathy, nephropathy, and nerve damage, and large fluctuations in blood glucose levels increase the risk of these complications. Therefore, diabetic patients need to regularly monitor their blood glucose levels to adjust their treatment plans and maintain blood glucose within a reasonable range.

[0003] Currently, diabetic patients often rely on invasive blood glucose monitoring methods, such as finger-prick or venous blood sampling. While these methods typically provide accurate blood glucose values, the procedures are cumbersome and can be quite painful; long-term use may lead to infection or tissue damage. Furthermore, traditional blood glucose monitoring methods cannot provide continuous monitoring, making it difficult for patients to track blood glucose changes in real time and inconvenient for daily use, especially during exercise or after meals, where patients may need to take multiple measurements, increasing adherence issues. Therefore, developing a non-invasive, convenient, and long-term continuous blood glucose monitoring technology has become an important research direction in the field of diabetes management.

[0004] With the development of technology, non-invasive blood glucose monitoring technology has gradually become a potential alternative to overcome the limitations of traditional methods. Spectroscopic analysis, as an important non-invasive method, has been widely used in blood glucose monitoring in recent years. By analyzing the reflection and absorption characteristics of light in the skin, blood, and other tissues, spectral analysis can indirectly extract blood glucose-related information, thereby enabling blood glucose prediction. Compared with traditional invasive methods, spectral analysis has advantages such as being non-invasive, real-time, and convenient, which can greatly improve patient compliance and reduce discomfort and risks associated with blood collection.

[0005] For example, the invention patent with publication number CN121331482A discloses a non-invasive dynamic detection method for blood glucose concentration using near-infrared spectroscopy, which includes: collecting multiple training samples and current near-infrared spectral data of blood at the current blood glucose concentration to be measured; the training samples are historical near-infrared spectral data of blood at different blood glucose concentrations; determining the difference in the peak value and corresponding wavelength of the water molecule absorption peak in the historical near-infrared spectral data and the current near-infrared spectral data of blood, as well as the difference in the width of the water molecule absorption peak, to obtain the usability of each training sample; and obtaining the mathematical regression model weight of each training sample based on the usability of each training sample and the collection time of each training sample.

[0006] However, despite the potential of spectral analysis in non-invasive blood glucose monitoring, several challenges remain in practical applications. The optical properties of different parts of the human body are complex and varied; tissues such as skin, blood vessels, fat, and red blood cells have different absorption, scattering, and reflection characteristics of light. These factors make spectral signals susceptible to interference from non-glucose components.

[0007] Furthermore, many existing technologies rely on light of a single or limited number of wavelengths, or analyze only the characteristic peaks of glucose or water molecules at specific sites. However, these methods often fail to effectively eliminate interference from the skin layer and other blood components (such as red blood cells and plasma). Therefore, the noise and errors in the spectral signal are often significant, making it difficult to guarantee the accuracy and stability of blood glucose prediction models. Especially when the measured subject experiences large fluctuations in blood glucose levels under different physiological states, during exercise, or after meals, these interfering factors in the spectral signal can significantly affect the accurate measurement of blood glucose concentration. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a non-invasive blood glucose detection method and device based on ultrawide near-infrared spectroscopy. By acquiring ultrawide multi-wavelength channel near-infrared spectral data and real blood glucose data of all fingers of the test subject, a blood glucose prediction model is constructed to achieve non-invasive blood glucose detection. This solves the problem of insufficient accuracy and stability of blood glucose prediction in existing non-invasive blood glucose detection technologies due to the complex optical properties of human tissues and the susceptibility of spectral signals to interference from non-blood glucose components.

[0009] To achieve the above-mentioned objectives, this invention provides a non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy, comprising the following steps: Step 1: Collect spectral data of each finger of the test subject in the ultra-wide near-infrared band at different time periods, and simultaneously collect the real blood glucose concentration data of the test subject to construct a dataset and perform preprocessing; Step 2: Based on the preprocessed dataset, construct a physical model-constrained neural network to extract blood glucose concentration from the near-infrared spectrum; train the neural network and introduce a physical constraint loss function based on Kubelka-Munk theory, Beer-Lambert law and scattering model during the training process, so that the output of blood glucose concentration by the neural network conforms to both spectral data and physical laws, and obtain the trained neural network. Step 3: When applying the method, the near-infrared spectral data of any subject's finger is used as input to predict blood glucose concentration through a trained neural network.

[0010] To address the shortcomings of existing technologies that rely on single or a few wavelengths of light, or analyze only the characteristic peaks of glucose or water molecules in specific areas, thus failing to effectively eliminate the influence of other components in the skin and blood, this invention collects spectral data across an ultra-wide near-infrared band across the fingers as a dataset for model construction. This allows for comprehensive analysis of the absorption and scattering characteristics of light signals by various tissue structures such as skin, blood vessels, and blood, thereby improving the accuracy of blood glucose prediction. Furthermore, a physical model-constrained neural network is constructed, enabling the neural network to understand the differences in reflectance spectra caused by the different physical structures of different fingers (such as skin thickness), thereby reducing the errors caused by factors such as finger thickness in the final blood glucose prediction.

[0011] Preferably, in step 1, the ultrawide near-infrared band range covers 200-1800 nm and includes more than 2000 wavelength channels.

[0012] This invention acquires near-infrared spectral information over an ultra-wide spectral range, enabling the model to comprehensively analyze the absorption and scattering characteristics of light signals by various tissue structures such as skin, blood vessels, and blood, thereby improving the accuracy of blood glucose prediction.

[0013] More preferably, in step 1, the ultrawide near-infrared band ranges from 206.87 to 1130.92 nm, containing 2068 wavelength channels.

[0014] Optionally, in step 1, the different time periods include before lunch, after lunch, two hours after lunch, and after eating dessert.

[0015] Preferably, in step 1, the preprocessing includes: normalizing the amplitude of the spectral data in the dataset, and stitching together the normalized spectral data of different fingers of the same test subject.

[0016] Preferably, in step 2, the construction of the neural network includes: Multi-scale convolutional neural networks are used to extract multi-scale spectral features of each finger from preprocessed spectral data and then feature concatenation is performed. Using each finger as a node, a fully connected graph is constructed. The spectral features of multiple fingers are then processed by a graph convolutional neural network to perform message passing and feature fusion, resulting in the fused finger node features. A corrected spectral decoder is introduced to estimate the skin tissue reflectance spectrum of each finger from the fused finger node features, and subtract it from the preprocessed spectral data to obtain the corrected blood reflectance spectrum; A concentration decoder is introduced to regress the concentrations of multiple components in the blood from the corrected blood reflectance spectrum and the fused finger node features, and output one of them as the blood glucose concentration; A learnable absorbance parameter matrix is ​​introduced to model the absorbance distribution of various components in blood at different wavelengths, which is then used for the calculation of the physical constraint loss function.

[0017] Preferably, before extracting the multi-scale spectral features of each finger from the preprocessed spectral data, average pooling is used to perform dimensionality reduction and noise reduction on the preprocessed spectral data. The dimensionality-reduced and noise-reduced spectral data is then input into a multi-scale convolutional neural network to extract the multi-scale spectral features of each finger.

[0018] Preferably, the modified spectral decoder includes a linear layer and a ReLU activation function, which are used to obtain the skin tissue reflectance spectrum by taking the fused finger node features as input through linear layer operation, and then pass it through the ReLU activation function to introduce nonlinearity and ensure that the output is non-negative; The linear layer parameters adopt a small-amplitude output initialization strategy, which specifically includes: the weights of the linear layer are initialized using a Gaussian distribution with a mean of 0, and the biases are initialized using a Gaussian distribution with a mean of 0.1.

[0019] Preferably, the size of the learnable absorbance parameter matrix is ​​N×M, where N is the number of blood components considered and M is the dimension of the spectral data after dimensionality reduction and noise reduction. In the initialization of the absorbance parameter matrix, the mean of each blood component is uniformly distributed in the range of 200-1800 nm, and the standard deviation is fixed at a higher value of 100 to ensure that the absorbance curve changes smoothly in a wide wavelength range. Specifically, the absorbance curves of blood components with a mean value around 1700 nm are considered as the absorbance of glucose, and their corresponding concentrations are used as the output of blood glucose concentration.

[0020] Preferably, in step 2, the physical constraint loss function includes: blood glucose concentration monitoring loss, Kubelka-Munk constraint loss, JS divergence constraint loss, and non-negativity constraint loss; The blood glucose concentration monitoring loss is measured using mean squared error to determine the difference between the predicted blood glucose concentration and the actual blood glucose concentration. The Kubelka-Munk constraint loss establishes the relationship between reflectance spectrum and absorption / scattering based on Kubelka-Munk theory. It combines Beer-Lambert's law to describe the linear relationship between absorption and substance concentration / absorbance, and a scattering model to describe the power-law variation of the scattering coefficient with wavelength. The forced-corrected reflectance spectrum and the predicted blood glucose concentration satisfy physical laws, expressed as follows: ,in, For Kubelka-Munk constraint loss, The corrected blood reflectance spectrum, For the first The absorbance corresponding to each blood component For the first The blood glucose concentration of various blood components The total number of blood components, and All are set as learnable parameters. For reference wavelength, For actual wavelength, This is the mean square error function, used to measure the difference between the two sides of the equation; The JS divergence constraint loss is used to constrain the distribution of the spectra obtained after subtracting skin reflection from different fingers to be as uniform as possible, in order to eliminate the interference caused by tissue differences between fingers, and is expressed as follows: ,in, For JS divergence constraint loss, The original reflectance spectrum of the j-th finger is... For the first Estimated reflectance spectrum of skin tissue from the finger. For actual wavelength, The total number of fingers, Jensen-Shannon divergence is used to measure the similarity between multiple probability distributions. The nonnegativity constraint loss is used to constrain all elements in the absorbance parameter matrix to be nonnegative, in accordance with the physical meaning of nonnegativity of absorbance in Beer-Lambert's law, and is expressed as follows: ,in, For non-negative constraint loss, No. The absorbance corresponding to each blood component For actual wavelength, The total number of blood components, and These are the lower and upper limits of the wavelength range. This indicates taking the absolute value of the negative absorbance value, used to penalize negative absorbance and ensure non-negative values. It is a function for maximizing the value.

[0021] The present invention also provides a non-invasive blood glucose detection device based on ultrawide near-infrared spectroscopy, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy when the computer program is executed.

[0022] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) A wide-spectrum-range spectrometer-based blood glucose concentration measurement method is proposed. By selecting the near-infrared spectrum that covers a wider range of wavelengths, the influence of other components in the blood on blood glucose prediction can be effectively reduced, thereby improving the accuracy of non-invasive blood glucose concentration prediction.

[0023] 2) By collecting spectral data from multiple fingers, errors caused by single-finger features can be reduced, providing more stable blood glucose prediction results.

[0024] 3) Neural networks are used to perform nonlinear regression modeling on high-dimensional ultrawide near-infrared spectral data. Through convolutional feature extraction and cross-finger feature fusion, the complex mapping relationship between different spectral bands and blood glucose concentration can be more fully explored. At the same time, physical constraint mechanisms are combined to suppress overfitting and improve the stability of neural networks, thereby improving the generalization ability of unknown data and the accuracy of blood glucose prediction. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0026] Figure 1 This is a schematic flowchart of the non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy provided by the present invention.

[0027] Figure 2 A schematic diagram of the structure of a neural network constrained by a physical model provided for an embodiment, and its training process. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and given in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] This invention provides a non-invasive blood glucose detection method and device based on ultrawide near-infrared spectroscopy. By acquiring ultrawide multi-wavelength channel near-infrared spectral data and real blood glucose data of all fingers of the test subject, a blood glucose prediction model is constructed to achieve non-invasive blood glucose detection. This solves the problem of insufficient accuracy and stability of blood glucose prediction in existing non-invasive blood glucose detection technologies due to the complex optical properties of human tissue and the susceptibility of spectral signals to interference from non-blood glucose components.

[0030] like Figure 1 As shown in the embodiment, a non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy is provided, which specifically includes the following steps: S1. At different time periods, collect spectral data of each finger of the test subject in the ultra-wide near-infrared band, and simultaneously collect the actual blood glucose concentration data of the test subject to construct a dataset and perform preprocessing.

[0031] S1.1 Spectral data acquisition.

[0032] A spectrometer was used to measure the spectra of multiple fingers of the test subject, and spectral data of each finger at different wavelengths were obtained.

[0033] In this embodiment, the spectrometer used has a minimum operating wavelength of 206.87 nm and a maximum operating wavelength of 1130.92 nm, containing a total of 2068 different wavelength channels; spectral data of 10 fingers were measured for each subject. By acquiring spectral information over a wide spectral range, the model can comprehensively analyze the absorption and scattering characteristics of light signals by various tissue structures such as skin, blood vessels, and blood, thereby improving the accuracy of blood glucose prediction.

[0034] S1.2 Acquisition of real blood glucose data.

[0035] While completing the spectral data acquisition, blood samples were collected from the test subjects, and their true blood glucose concentration was measured as calibration data for subsequent neural network training and validation.

[0036] S1.3 Multiple Sample Acquisition and Data Preprocessing.

[0037] Repeat steps S1.1 and S1.2 to measure different subjects and the same subject at different time periods, and obtain multiple sets of spectral data and corresponding real blood glucose data.

[0038] In this embodiment, the different time periods include before lunch, after lunch, two hours after lunch, and after eating dessert, to ensure that the sample data can cover different blood glucose concentration ranges.

[0039] The acquired spectral data undergoes preprocessing, specifically including: The amplitude of the spectral data is normalized so that the input spectral data is between 0 and 1, making it easier for the neural network to converge.

[0040] The spectral data of different fingers of the same test subject were stitched together to form the input data. ,in To input the number of fingers, For spectral data dimensions.

[0041] In this embodiment, A value of 10 indicates that spectral data from 10 fingers are considered. The value is 2068, which means that the spectral data contains 2068 different wavelengths. The final dataset contains a total of 47 samples.

[0042] S1.4 Dataset partitioning.

[0043] The dataset is divided into a training set and a test set, with the training set containing 31 samples and the test set containing 16 samples, for neural network training and performance evaluation, respectively.

[0044] S2. Based on the preprocessed dataset, construct a physical model-constrained neural network to extract blood glucose concentration from the near-infrared spectrum; train the neural network and introduce a physical constraint loss function based on Kubelka-Munk theory, Beer-Lambert law and scattering model during the training process, so that the output of blood glucose concentration by the neural network conforms to the spectral data and physical laws, and obtain the trained neural network.

[0045] S2.1 Initialization Section: For the measured reflectance spectrum Reflectance spectra can be established. With incident light intensity absorption spectrum and scattering spectrum Mapping relationship: , in, For mapping functions, The actual wavelength; Since this invention uses near-infrared spectroscopy, the reflection of incident light by skin tissue, etc., can be considered as a linear function of wavelength. Thus, the blood reflectance spectrum is obtained as follows: ; Since the light source intensity is weak, the thickness of the finger can be considered infinite. Therefore, the Kubelka-Munk theory can be used to model the relationship between the reflection (R), absorption (A), and scattering (S) of incident light by blood: ; According to Beer-Lambert's law and the superposition law, we have: , in, Wavelength Under the conditions The absorption coefficients of the various blood components For optical path, The total number of blood components considered for neural networks.

[0046] For absorption coefficient It is related to blood components absorbance and its concentration in the blood The following mapping relationship exists: .

[0047] Furthermore, for blood scattering, a power-law approximation can be used: , in, The scattering coefficient is... For reference wavelength, and It is a constant.

[0048] Substituting into the Kubelka-Munk model, we get: , in, The reflectance coefficient is the same as the corrected reflectance spectrum. Proportional.

[0049] The neural network component of the physical model constraints in S2.2: In this embodiment, the target substance and the monitoring signal are blood glucose concentration. The physical model-constrained neural network (…) Figure 2The construction process of ) is as follows: First, considering the similar blood absorption characteristics of adjacent wavelengths in the input high-dimensional spectral data, this invention employs one-dimensional average pooling (denoising) to perform dimensionality reduction and noise reduction on the preprocessed spectral sequence in order to reduce dimensionality and suppress noise, resulting in dimensionality-reduced and noise-reduced spectral data. Then we have: , in, This is the preprocessed spectral sequence. Indicates the pooling kernel size. Indicates the pooling step size. This is an average pooling operation. In this embodiment, k is set to 11 and s is set to 11, meaning that the amplitude corresponding to every 11 adjacent wavelength channels is averaged and downsampled to obtain the processed spectral features.

[0050] Then, a convolutional neural network is used to perform multi-scale convolution on the processed spectrum to obtain the multi-scale spectral features of each finger. : , Then and By concatenating along the feature dimension, a sequence is formed. We collect individual finger features and construct a fully connected graph using fingers as nodes. A three-layer graph convolutional neural network (using skin data from the fingers) is then used to perform cross-finger message passing and feature fusion on these individual finger features, resulting in the fused finger node features. .

[0051] Then, using a modified spectral decoder containing a linear layer and a ReLU activation function, the reflectance spectra produced by the skin tissue of each finger are calculated. and with Subtracting the two values ​​and averaging them along the finger dimension yields the corrected reflectance spectrum. The decoder's linear layer parameters employ a "small-amplitude output" initialization strategy: the weights W of the linear layer are initialized using a Gaussian distribution with a mean of 0, ensuring that the neural network outputs small amplitudes in the early stages of training, thus avoiding excessive interference. Excessive corrections are generated; the bias is initialized uniformly in the positive range of a Gaussian distribution with a mean of 0.1 to ensure that the initial output is a small positive non-zero value, thereby improving the stability of training and facilitating faster convergence.

[0052] Feature extraction from multiple fingers and feature fusion across fingers allow neural networks to understand the differences in reflectance spectra caused by the different physical structures of different fingers (such as skin thickness), thereby reducing the error caused by factors such as finger thickness in the final blood glucose prediction.

[0053] Finally, a concentration decoder containing a linear layer and a ReLU activation function is used to decode the concentrations of N different substances in the blood considered by the neural network. And take a specific concentration as the blood glucose concentration.

[0054] To ensure that the blood glucose concentration predicted by the neural network has a physical meaning, this invention adds a set of learnable parameters to the model. It is used to model the absorbance of different components in blood. Initialize using a Gaussian distribution with N means In wavelength range Uniformly distributed within, with a fixed, relatively high standard deviation. The absorbance of each component is constrained to change gradually over a wide wavelength range, and the initial value is positive.

[0055] In the embodiments, the lower wavelength limit Slightly below the shortest wavelength in the spectrum, take 200 nm, the upper limit of the wavelength range. The absorption peak is slightly larger than that of glucose, taken at 1800 nm, considering a total of N=64 blood components. Among them, since blood glucose has a significant absorption peak around 1700 nm, it was initialized... In this study, the distribution with a mean at 1700 nm was taken as the absorbance curve of blood glucose, and the corresponding values ​​were... It is used as the output of blood glucose concentration.

[0056] S2.3 Training and Evaluation of Neural Networks Constrained by Physical Models To enable neural networks to model the relationship between blood glucose concentration and reflectance spectrum, this invention designs multiple loss functions to optimize neural networks constrained by physical models.

[0057] Blood glucose concentration monitoring loss It is used to directly display the optimized final predicted blood glucose concentration based on the actual blood glucose value: , in, The blood glucose concentration predicted by the model. The true blood glucose concentration is denoted as MSE, which is the mean squared error function used to measure the difference between the predicted and true blood glucose concentrations.

[0058] Kubelka-Munk Constraint Loss This method is used to establish the relationship between reflectance spectrum and absorption / scattering based on Kubelka-Munk theory. It combines Beer-Lambert's law to describe the linear relationship between absorption and substance concentration / absorbance, and a scattering model to describe the power-law variation of the scattering coefficient with wavelength. The forced correction of the reflectance spectrum and the predicted blood glucose concentration conforms to physical laws, allowing the neural network to make better use of the supervision signal and establish a better mapping relationship between reflectance spectrum and blood glucose concentration. , in, For Kubelka-Munk constraint loss, The corrected blood reflectance spectrum, For the first The absorbance corresponding to each blood component For the first The blood glucose concentration of various blood components The total number of blood components, and All are set as learnable parameters. For reference wavelength, For actual wavelength, This is the mean square error function, used to measure the difference between the two sides of the equation.

[0059] In this embodiment, and The initial value is 1, and the reference wavelength is... Fixed at 800 nm.

[0060] JS divergence constraint loss Reflectance spectrum used for constraint Calculated with different fingers The result after subtraction The information spectra of different fingers are nearly identical in distribution to ensure that the model can understand the differences in reflectance spectra caused by the different physical structures of different fingers (such as skin thickness). ,in, For JS divergence constraint loss, For the first The original reflectance spectrum of the root finger, For the first Estimated reflectance spectrum of skin tissue from the finger. For actual wavelength, The total number of fingers, The Jensen-Shannon divergence measures the similarity between multiple probability distributions.

[0061] Non-negative constraint terms This is used to constrain all elements in the absorbance parameter matrix to be non-negative, in accordance with the physical meaning of non-negativity of absorbance in Beer-Lambert's law, and is expressed as follows: ,in, For non-negative constraint loss, No. The absorbance corresponding to each blood component For actual wavelength, The total number of blood components, and These are the lower and upper limits of the wavelength range. This indicates taking the absolute value of the negative absorbance value, used to penalize negative absorbance and ensure non-negative values. It is a function for maximizing the value.

[0062] Therefore, the constructed physical constraint loss function is expressed as: , in, Weighting for blood glucose concentration monitoring loss. For Kubelka-Munk constrained loss weights, For the JS divergence constraint loss weights, The weights are non-negative constraint loss weights; in this embodiment, Take 0.5, Take 10 -6 , Take 10 -4 , Take 10 -4 .

[0063] The model uses the Adam optimizer with a learning rate of 10. -4 , train 100 times.

[0064] During the evaluation phase, mean squared error and mean relative error were used as performance evaluation metrics. On the test set, the mean squared error of the neural network constrained by the constructed physical model was 0.761, and the mean relative error was 11.696%, indicating that the method proposed in this invention can achieve high-precision prediction of blood glucose concentration under non-invasive conditions and has good generalization performance.

[0065] S3. When applying the technology, the near-infrared spectral data of any subject's finger is used as input, and the blood glucose concentration is predicted through a trained neural network.

[0066] Based on the same inventive concept, the embodiment also provides a non-invasive blood glucose detection device based on ultrawide near-infrared spectroscopy, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy when the computer program is executed.

[0067] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy, characterized in that, Includes the following steps: Step 1: Collect spectral data of each finger of the test subject in the ultra-wide near-infrared band at different time periods, and simultaneously collect the real blood glucose concentration data of the test subject to construct a dataset and perform preprocessing; Step 2: Based on the preprocessed dataset, construct a physical model-constrained neural network to extract blood glucose concentration from the near-infrared spectrum; train the neural network and introduce a physical constraint loss function based on Kubelka-Munk theory, Beer-Lambert law and scattering model during the training process, so that the output of blood glucose concentration by the neural network conforms to both spectral data and physical laws, and obtain the trained neural network. Step 3: When applying the method, the near-infrared spectral data of any subject's finger is used as input to predict blood glucose concentration through a trained neural network.

2. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 1, characterized in that, In step 1, the ultrawide near-infrared band range covers 200-1800 nm and includes more than 2000 wavelength channels.

3. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 1, characterized in that, The preprocessing includes: normalizing the amplitude of the spectral data in the dataset, and stitching together the normalized spectral data of different fingers of the same test subject.

4. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 1, characterized in that, In step 2, the construction of the neural network includes: Multi-scale convolutional neural networks are used to extract multi-scale spectral features of each finger from preprocessed spectral data and then feature concatenation is performed. Using each finger as a node, a fully connected graph is constructed. The spectral features of multiple fingers are then processed by a graph convolutional neural network to perform message passing and feature fusion, resulting in the fused finger node features. A corrected spectral decoder is introduced to estimate the skin tissue reflectance spectrum of each finger from the fused finger node features, and subtract it from the preprocessed spectral data to obtain the corrected blood reflectance spectrum; A concentration decoder is introduced to regress the concentrations of multiple components in the blood from the corrected blood reflectance spectrum and the fused finger node features, and output one of them as the blood glucose concentration; A learnable absorbance parameter matrix is ​​introduced to model the absorbance distribution of various components in blood at different wavelengths, which is then used for the calculation of the physical constraint loss function.

5. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 4, characterized in that, Before extracting the multi-scale spectral features of each finger from the preprocessed spectral data, average pooling is used to perform dimensionality reduction and noise reduction on the preprocessed spectral data. The dimensionality-reduced and noise-reduced spectral data is then input into a multi-scale convolutional neural network to extract the multi-scale spectral features of each finger.

6. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 4, characterized in that, The modified spectral decoder includes a linear layer and a ReLU activation function. It takes the fused finger node features as input, performs linear layer operations to obtain the skin tissue reflectance spectrum, and then passes it through the ReLU activation function to introduce nonlinearity and ensure that the output is non-negative. The linear layer parameters adopt a small-amplitude output initialization strategy, which specifically includes: the weights of the linear layer are initialized using a Gaussian distribution with a mean of 0, and the biases are randomly initialized using a Gaussian distribution with a mean of 0.

1.

7. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 4, characterized in that, The learnable absorbance parameter matrix is ​​N×M in size, where N is the number of blood components considered and M is the dimension of the spectral data after dimensionality reduction and noise reduction. In the initialization of the absorbance parameter matrix, the mean of each blood component is uniformly distributed in the range of 200-1800 nm, and the standard deviation is fixed at a higher value of 100 to ensure that the absorbance curve changes smoothly in a wide wavelength range. Specifically, the absorbance curves of blood components with a mean value around 1700 nm are considered as the absorbance of glucose, and their corresponding concentrations are used as the output of blood glucose concentration.

8. The non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy according to claim 4, characterized in that, In step 2, the physical constraint loss function includes: blood glucose concentration monitoring loss, Kubelka-Munk constraint loss, JS divergence constraint loss, and non-negativity constraint loss; The blood glucose concentration monitoring loss is measured using mean squared error to measure the difference between the predicted blood glucose concentration and the actual blood glucose concentration. The Kubelka-Munk constraint loss establishes the relationship between reflectance spectrum and absorption / scattering based on Kubelka-Munk theory. It combines Beer-Lambert's law to describe the linear relationship between absorption and substance concentration / absorbance, and a scattering model to describe the power-law variation of the scattering coefficient with wavelength. The forced-corrected reflectance spectrum and the predicted blood glucose concentration satisfy physical laws, expressed as follows: ,in, For Kubelka-Munk constraint loss, The corrected blood reflectance spectrum, For the first The absorbance corresponding to each blood component For the first The blood glucose concentration of various blood components The total number of blood components, and All are set as learnable parameters. For reference wavelength, For actual wavelength, This is the mean square error function, used to measure the difference between the two sides of the equation; The JS divergence constraint loss is used to constrain the distribution of the spectra obtained after subtracting skin reflection from different fingers to be as uniform as possible, in order to eliminate the interference caused by tissue differences between fingers, and is expressed as follows: ,in, For JS divergence constraint loss, For the first The original reflectance spectrum of the root finger, For the first Estimated reflectance spectrum of skin tissue from the finger. For actual wavelength, The total number of fingers, Jensen-Shannon divergence is used to measure the similarity between multiple probability distributions. The nonnegativity constraint loss is used to constrain all elements in the absorbance parameter matrix to be nonnegative, in accordance with the physical meaning of nonnegativity of absorbance in Beer-Lambert's law, and is expressed as follows: ,in, For non-negative constraint loss, No. The absorbance corresponding to each blood component For actual wavelength, The total number of blood components, and These are the lower and upper limits of the wavelength range. This represents the absolute value of the negative absorbance value, used to penalize negative absorbance and ensure non-negative absorbance. It is a function for maximizing the value.

9. A non-invasive blood glucose detection device based on ultrawide near-infrared spectroscopy, comprising a memory and a processor, wherein the memory is used to store a computer program, characterized in that, The processor is configured to implement the non-invasive blood glucose detection method based on ultrawide near-infrared spectroscopy as described in any one of claims 1-8 when executing the computer program.

Images