A variable-order array type non-invasive blood glucose detection method based on a graph neural network

Abstract

The application discloses a variable-order array type non-invasive blood glucose detection method based on a graph neural network, solves the problem that the data volume increases during detection of an existing non-invasive blood glucose detection method, and reduces the blood glucose detection speed, solves the reconstruction of a low-order array to a high-order array, and realizes higher accuracy with a smaller array, and the method comprises the following steps: collecting low-order impedance data based on an array type bioelectrical impedance front end, performing order reconstruction and feature extraction processing on the low-order impedance data by using an array type bioelectrical impedance blood glucose monitoring model, and outputting a blood glucose detection result; the application proposes to construct an impedance data order reconstruction module of low-order array type impedance data, the array type bioelectrical impedance blood glucose monitoring model uses bioelectrical impedance imaging technology to realize preprocessing processes such as low-order data feature expansion, and further improves the robustness of the array type bioelectrical impedance non-invasive blood glucose detection model, and improves the non-invasive blood glucose detection efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of blood glucose detection technology, specifically relating to a variable-order array-based non-invasive blood glucose detection method based on graph neural networks. Background Technology

[0002] Bioelectrical impedance measurement technology is a non-destructive testing technique that obtains biomedical information about the physiological and pathological states of the human body by examining the electrical properties (impedance, phase, dielectric constant, etc.) of biological tissues and organs and their changes. Its advantages include low cost, non-invasiveness, safety, ease of operation, rich information content, and non-toxicity.

[0003] Body fat percentage and blood glucose are important physiological parameters used to assess a person's health. For disease prevention, diagnosis, and evaluation of treatment effectiveness, the array-type impedance acquisition front-end is a crucial component in blood glucose measurement based on human body impedance technology, directly affecting the accuracy and anti-interference capability of blood glucose detection methods. Increasing the number of measuring electrodes in the detection area can effectively improve the resolution of array-type bioelectrical impedance analysis, thereby improving the performance of non-invasive blood glucose detection. However, this inevitably increases the data volume, reducing the speed of blood glucose detection. To address these issues, we propose a variable-order array-type non-invasive blood glucose detection method based on graph neural networks. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a variable-order array-based non-invasive blood glucose detection method based on graph neural networks, which solves the problem that the increased data volume in existing non-invasive blood glucose detection methods leads to a decrease in blood glucose detection speed.

[0005] This invention is implemented as follows: a variable-order array-based non-invasive blood glucose detection method based on graph neural networks, comprising:

[0006] Low-order impedance data is acquired based on an array-type bioelectrical impedance front-end.

[0007] Low-order impedance data is loaded in real time. Using the low-order impedance data as input, an array-based bioelectrical impedance glucose monitoring model is executed. The array-based bioelectrical impedance glucose monitoring model performs up-order reconstruction and feature extraction processing on the low-order impedance data and outputs the glucose detection results.

[0008] Preferably, the array-based bioelectrical impedance blood glucose monitoring model includes:

[0009] An impedance data upgrade and reconstruction module is used to reconstruct low-order impedance data into high-order data and output high-order array impedance data. The impedance data upgrade and reconstruction module includes a data matrixing module, an EIT conductivity distribution conversion module, an upgrade module, and a reconstruction module. The upgrade module and the reconstruction module each include a U-shaped autoencoder, which is composed of an encoder, a decoder, and a convolutional block.

[0010] The natural feature extraction module is used to supplement the signal features of low-order impedance data to compensate for the influence of ignored blood glucose fluctuation information on the classification results.

[0011] A multi-scale constraint module is used to add constraint requirements for the blood glucose classification task in the low-order impedance data reconstruction stage and reduce the introduction of new noise in the process of upgrading the low-order impedance data.

[0012] A topology module, which is used to convert high-order array impedance data into graph topology;

[0013] The Diff-PNA classification model is used to obtain graph topology and, based on the Diff-PNA algorithm, splices natural features to achieve blood glucose concentration classification.

[0014] Preferably, the method for reconstructing high-order data from low-order impedance data specifically includes:

[0015] Based on electrical impedance imaging technology, low-order impedance data are preprocessed to reconstruct a two-dimensional image matrix of conductivity distribution.

[0016] The reconstructed two-dimensional image matrix is ​​obtained, and the two-dimensional image matrix is ​​reconstructed by an up-order U-shaped autoencoder to output high-order array impedance data.

[0017] Preferably, the method for preprocessing low-order impedance data based on electrical impedance imaging technology specifically includes:

[0018] The electrode array applies an excitation current to the tissue through the excitation electrode pair and measures the voltage through the acquisition electrode pair. The array impedance data is acquired by combining the excitation electrode pair and the acquisition electrode pair.

[0019] The finite element method is used to solve the forward problem, that is, based on the physical model and boundary conditions, to accurately calculate the distribution of current inside the object and the resulting potential distribution.

[0020] The Gauss-Newton method is used to solve the inverse EIT problem. By approximating the nonlinear relationship of the objective function, the optimal solution, i.e. the true value of the conductivity distribution, is found iteratively.

[0021] Preferably, the method for order-up reconstruction of a two-dimensional image matrix based on a U-shaped autoencoder specifically includes:

[0022] Using the finite element method, each region Ω is modeled as a triangle, and the Gauss-Newton method is used for electrical impedance imaging to construct a high-dimensional conductivity distribution map from the 2×2 electrode array data, dividing the imaging space Ω into a standard 100×100 matrix grid Ω. ψ Each element ψ in the grid is a region in its space, i.e., ψ∈Ω ψ The grid is then overlapped with the conductivity distribution image grid generated by EIT, yielding the intersection of the regions in the matrix grid and the EIT grid regions. Since the conductivity value is the same in each EIT grid, the conductivity value of the matrix grid region is the weighted average of its intersection with the EIT grid regions. Therefore, the conductivity value σ in the rectangular grid is... ψ for:

[0023]

[0024] Where O={e|e∈Ω e ,ψ∈Ω ψ ,e∪ψ≠φ} denotes the triangular EIT mesh Ω that intersects with the region ψ. e The set of regions e, S e∪ψ S represents the area of ​​the intersection of the EIT grid region and the rectangular grid region. ψ σ represents the area of ​​the rectangular grid region. e This represents the conductivity of a region within the EIT grid, thus a standard conductivity distribution image matrix Ψ = (σ ψij ) 100×100 Construction complete;

[0025] The upsampling module constructs an autoencoder according to a U-shaped network structure. The autoencoder is used to convert the low-order conductivity distribution image to the high-order conductivity distribution image. The autoencoder consists of a downsampling layer and an upsampling layer. The downsampling layer extracts features by applying max pooling layers and convolutional blocks defined above in two successive stages. The upsampling layer gradually restores the spatial size of the feature map by two transposed convolutional layers, expanding it by a factor of 2 in length and width each time. At the same time, skip connections are used to concatenate the corresponding feature map of the encoder stage with the feature map of the current decoding stage.

[0026] The reconstruction module uses the same U-shaped network structure as the upgrade module to perform upgrade reconstruction of the two-dimensional image matrix.

[0027] Preferably, the Diff-PNA classification model uses a graph neural network method to extract features and classify the constructed graph topology;

[0028] The Diff-PNA classification model uses a 5-layer multilayer perceptron to achieve the final classification result X. out The output is represented as follows:

[0029]

[0030] in, X represents the characteristics of node j output by the 5th layer PNA network. p This is an embedded feature of the entire image.

[0031] Preferably, the natural feature extraction module is used to extract low-order array data and to convert between low-order array data and high-order array data;

[0032] When the low-order array data is a 2×2 low-order array, the natural feature extraction module extracts the most original low-order 2×2 electrode array data features and splices them with the PNA features of the 4th layer.

[0033] The natural feature extraction module is used to extract the features of the original low-order 2×2 electrode array data. When concatenating these features with the PNA features of the fourth layer, the acquisition matrix of the original low-order 2×2 electrode array is first input. The acquisition matrix is ​​described as R = (R ijk ) 4×3×3 The natural feature extraction module uses two deconvolutional layers to expand the acquisition matrix of the 2×2 electrode array, and then uses three convolutional layers and a flattening layer to transform the features into a one-dimensional vector representation, described as follows: This feature is directly derived from the acquisition matrix of the 2×2 electrode impedance array and is directly concatenated with the fourth layer PNA feature of the main classification model. Therefore, the fifth layer PNA of the model is represented as follows:

[0034]

[0035]

[0036]

[0037] Here, O represents the horizontal concatenation of two matrix dimensions and features.

[0038] Preferably, the multi-scale constraint module incorporates constraints for the blood glucose classification task during the impedance data reconstruction stage to compensate for the impact of ignored blood glucose fluctuation information on the classification results. The multi-scale constraint module consists of 5 layers of convolutional layers, flattening layers, and 5 layers of multilayer perceptron. The multi-scale constraint module takes the high-order 4×4 electrode array data results from the impedance data up-order reconstruction module as input and outputs the blood glucose classification results.

[0039] Preferably, when constructing the array-based bioelectrical impedance blood glucose monitoring model, the model is iteratively trained using training and testing sets. Two optimizers are used for the impedance data reconstruction module, natural feature extraction module, multi-scale constraint module, and Diff-PNA classification model, respectively. Both optimizers employ adaptive moment estimation, and the first and second moments that track the gradient are expressed as follows:

[0040]

[0041] m t+1 =β1·m t +(1-β1)·G t ,

[0042] v t+1 =β2·v t +(1-β2)·(G t ) 2 ,

[0043]

[0044] Where η is the learning rate, and G t The loss function at the current parameter θ t gradient at m t and v t These are the first and second moment estimates of the gradient, respectively, where β1 and β2 are decay rate parameters, and ε = 10. -8 It is a constant to prevent division by zero, and ⊙ represents the Hadamard product;

[0045] The parameter set of the array-based bioelectrical impedance blood glucose monitoring model is defined as follows:

[0046] Θ:={Θ rebuild ,Θ na_feature ,Θ PNA ,Θ constraints}

[0047] Where Θ rebuild ,Θ na_feature ,Θ PNA ,Θ constraints Let represent the parameters of the impedance data order reconstruction module, the Diff-PNA classification model, the natural feature extraction module, and the multi-scale constraint network module, respectively. Under the multi-scale loss function, the gradients at different locations of the model are obtained as follows:

[0048]

[0049] Therefore, the two classifiers are represented as follows:

[0050]

[0051] Among them, use This represents the optimizer for the impedance data up-order reconstruction module, where η1 represents its learning rate, f Adma2 η represents the optimizer for other modules, and η2 represents its learning rate.

[0052] Preferably, when iteratively training the array-type bioelectrical impedance glucose monitoring model using training and test sets, the ratio of training set to test set is 4:1. Mean squared error is used to estimate the final classification result and the order reconstruction. Mean squared error, peak signal-to-noise ratio, structural similarity index, and cosine similarity are used to evaluate the impedance data order reconstruction module.

[0053] The mean squared error measures the degree of similarity between the 4×4 electrode reconstruction values ​​of the array-based bioelectrical impedance glucose monitoring model and the data acquired by the real-label 4×4 electrode array, and its formula is expressed as:

[0054]

[0055] Where I(i,j) and K(i,j) are the values ​​of the observation array data at (i,j) respectively, and m and n are the number of rows and columns of the observation array data respectively;

[0056] Peak signal-to-noise ratio (PSNR) evaluates the error between corresponding pixels in two images, i.e., error-sensitive image quality assessment, and is expressed as:

[0057]

[0058] Among them, MAX I It is the maximum possible pixel value in the image data;

[0059] The structural similarity index compares the similarity between two images, including brightness, contrast, and structural information, and is expressed as:

[0060]

[0061] Where, μ I and μ K These represent the average brightness of images I and K, respectively. and Let σ represent the variances of images I and K, respectively. IK Let C1 and C2 represent the covariance of images I and K. C1 and C2 are small constants introduced to avoid the denominator being zero. Let C1 = (k1L). 2 C2 = (k2L) 2 Where L is the dynamic range of pixel values, and k1 = 0.01 and k2 = 0.03 are default constants;

[0062] Cosine similarity compares the similarity of two distributions in a spatial direction, evaluating the directional similarity of conductivity distributions rather than their absolute values, and is expressed as:

[0063]

[0064] Where I·K represents the dot product of two image vectors, and ||I|| and ||K|| represent the magnitudes of I and K, respectively.

[0065] Compared with the prior art, the embodiments of this application have the following main advantages:

[0066] In this embodiment of the invention, a low-order implementation strategy for high-order arrays is proposed, and an impedance data upscaling reconstruction module for low-order array-type impedance data is constructed, thereby realizing a variable array-type bioelectrical impedance blood glucose monitoring model. The array-type bioelectrical impedance blood glucose monitoring model uses electrical impedance tomography (EIT) to perform preprocessing processes such as feature expansion of low-order data, and on this basis, it realizes the resampling process from high-order impedance images to low-order arrays. It also integrates a natural feature extraction module and multi-scale constraint strategies, further improving the robustness of the array-type bioelectrical impedance non-invasive blood glucose detection model, improving the efficiency and accuracy of non-invasive blood glucose detection, and solving the reconstruction from low-order arrays to high-order arrays, enabling higher accuracy with smaller arrays. Attached Figure Description

[0067] Figure 1 This is a schematic diagram illustrating the implementation process of the variable-order array-based non-invasive blood glucose detection method based on graph neural networks provided by the present invention.

[0068] Figure 2 This is a schematic diagram of the structure of the array-type bioelectrical impedance blood glucose monitoring model provided by the present invention.

[0069] Figure 3 A schematic diagram illustrating the implementation process of the method for reconstructing high-order data from low-order impedance data is shown.

[0070] Figure 4 A flowchart of electrical impedance imaging is shown.

[0071] Figure 5 A schematic diagram of the impedance data upgrade reconstruction module is shown.

[0072] Figure 6 A comparative schematic diagram of the conductivity distribution of a triangular EIT grid and a matrix network is shown.

[0073] Figure 7 A schematic diagram of the reconstruction data construction method is shown.

[0074] Figure 8 A schematic diagram showing a comparison of fine-grained imaging of data from different arrays is presented.

[0075] Figure 9 This diagram illustrates the effect of upgraded high-order array impedance data on reconstruction. Detailed Implementation

[0076] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0077] Currently, increasing the number of measuring electrodes in the detection area can effectively improve the resolution of array bioelectrical impedance analysis (BIA) and thus enhance the performance of non-invasive blood glucose detection. However, this inevitably increases the amount of data, reducing the speed of blood glucose detection. To address these issues, we propose a variable-order array-based non-invasive blood glucose detection method based on graph neural networks. This method first acquires low-order impedance data from the array bioelectrical impedance analysis front-end, then executes the array bioelectrical impedance analysis blood glucose monitoring model. The model reconstructs the low-order impedance data by increasing its order and outputs the blood glucose detection results. In this embodiment of the invention, a low-order implementation strategy for high-order arrays is proposed, and an impedance data order-up reconstruction module for low-order array-type impedance data is constructed, thereby realizing a variable array-type bioelectrical impedance blood glucose monitoring model. The array-type bioelectrical impedance blood glucose monitoring model uses electrical impedance tomography (EIT) to perform preprocessing processes such as feature expansion of low-order data, and on this basis, it realizes the resampling process from high-order impedance images to low-order arrays. It also integrates a natural feature extraction module and a multi-scale constraint strategy, further improving the robustness of the array-type bioelectrical impedance non-invasive blood glucose detection model, improving the efficiency and accuracy of non-invasive blood glucose detection, and solving the reconstruction from low-order arrays to high-order arrays, enabling higher accuracy with smaller arrays.

[0078] This invention provides a variable-order array-based non-invasive blood glucose detection method based on graph neural networks. Figure 1A schematic diagram illustrating the implementation process of a variable-order array-based non-invasive blood glucose detection method based on graph neural networks is shown. Specifically, this method includes:

[0079] S10, based on an array-type bioelectrical impedance front-end, acquires low-order impedance data;

[0080] S20 loads low-order impedance data in real time. Using the low-order impedance data as input, it executes an array-based bioelectrical impedance glucose monitoring model. The array-based bioelectrical impedance glucose monitoring model performs up-order reconstruction and feature extraction processing on the low-order impedance data and outputs the glucose detection results.

[0081] In this embodiment of the invention, a low-order implementation strategy for a high-order array is proposed, and an impedance data order-up reconstruction module for low-order array-type impedance data is constructed, thereby realizing a variable array-type bioelectrical impedance blood glucose monitoring model. The array-type bioelectrical impedance blood glucose monitoring model uses electrical impedance tomography (EIT) to perform preprocessing processes such as feature expansion of low-order data, and on this basis, it realizes the process of resampling from high-order impedance images to low-order arrays. It also integrates a natural feature extraction module and a multi-scale constraint strategy, further improving the robustness of the array-type bioelectrical impedance non-invasive blood glucose detection model, as well as improving the efficiency and accuracy of non-invasive blood glucose detection.

[0082] like Figure 2 As shown, the array-based bioelectrical impedance blood glucose monitoring model includes:

[0083] Impedance data order-up and reconstruction module, which is used to reconstruct low-order impedance data into high-order data and output high-order array impedance data. The impedance data order-up and reconstruction module includes a data matrix module, an EIT conductivity distribution conversion module, an order-up module, and a reconstruction module. The order-up module and the reconstruction module each include a U-shaped autoencoder, which is composed of an encoder, a decoder, and a convolutional block.

[0084] The natural feature extraction module is used to supplement the signal features of low-order impedance data to compensate for the influence of ignored blood glucose fluctuation information on the classification results.

[0085] A multi-scale constraint module is used to add constraint requirements for the blood glucose classification task in the low-order impedance data reconstruction stage and reduce the introduction of new noise in the process of upgrading the low-order impedance data.

[0086] A topology module, which is used to convert high-order array impedance data into graph topology;

[0087] The Diff-PNA classification model is used to obtain graph topology and, based on the Diff-PNA algorithm, splices natural features to achieve blood glucose concentration classification.

[0088] It should be noted that the array-based bioelectrical impedance glucose monitoring model first inputs low-order impedance data into an upgraded reconstruction model to enrich the impedance information. Then, the reconstructed impedance data and the method designed in the previous chapter are used to classify blood glucose concentrations. Specifically, such as... Figure 1 As shown, the array-based bioelectrical impedance blood glucose monitoring model is an end-to-end organizational form.

[0089] In the impedance data order reconstruction module, a 2×2 electrode array is used as the low-order electrode array and a 4×4 electrode array is used as the high-order electrode array. The low-order array data is reconstructed into high-order data output in a supervised manner.

[0090] Specifically, the impedance data upgrade and reconstruction module applies current to the surface of a biological organism using low-order array impedance imaging technology and measures the voltage response to detect differences in impedance among different tissues or materials within the organism. This non-invasive imaging provides crucial information about the internal impedance distribution. Subsequently, high-order impedance data is obtained by resampling the impedance distribution map after imaging. Furthermore, to reduce information loss in reconstructing high-resolution impedance images using a 2×2 electrode array, the array-based bioelectrical impedance glucose monitoring model employs a U-shaped autoencoder upgrade module to achieve high-resolution reconstruction of impedance imaging, enriching the information in the impedance image and improving its quality. Simultaneously, the impedance data upgrade and reconstruction module reconstructs the edge potential of the object using high-quality impedance images, i.e., inversely reconstructing the impedance acquisition data, further reducing noise in the data.

[0091] Since the impedance data upscaling and reconstruction module focuses on the characteristics of blood glucose at the impedance imaging level, some non-target features are inevitably introduced during the reconstruction process. These features may be considered noise for blood glucose classification tasks. Therefore, the array-based bioelectrical impedance blood glucose monitoring model employs a multi-scale constraint module to reduce the impact of noise. The multi-scale constraint module constrains the model at three scales: the upscaling result of the low-order conductivity distribution image, the reconstruction result of the high-order array impedance data, and the classification result of the Diff-PNA model. Specifically, the upscaling result of the low-order conductivity distribution image requires that the upscaled conductivity distribution image maintain similarity to the high-order conductivity distribution image; this constraint is achieved using mean squared error loss. For the reconstruction result of the high-order array impedance data, the array-based bioelectrical impedance blood glucose monitoring model uses mean squared error loss to constrain it, making the reconstruction result closer to the actual high-order array acquisition data. Simultaneously, the array-based bioelectrical impedance blood glucose monitoring model constructs a CNN-based classifier, which uses cross-entropy loss to constrain the high-order array impedance data, making the features of the reconstructed result more suitable for blood glucose classification tasks. Finally, cross-entropy loss is used to constrain the final classification result of the Diff-PNA model.

[0092] Finally, the array-based bioelectrical impedance glucose monitoring model transforms impedance data into graph topology data and introduces natural features into the graph topology. The Diff-PNA classification model then processes the graph topology data, achieving an end-to-end process from 2×2 electrode array input to classification output.

[0093] This invention provides a method for reconstructing high-order data from low-order impedance data. Figure 3 The diagram illustrates the implementation flow of a method for reconstructing high-order data from low-order impedance data. Specifically, the method includes:

[0094] S101, based on electrical impedance imaging technology, performs data preprocessing on low-order impedance data to reconstruct a two-dimensional image matrix of conductivity distribution;

[0095] S102: Obtain the reconstructed two-dimensional image matrix, perform up-order reconstruction of the two-dimensional image matrix based on a U-shaped autoencoder, and output high-order array impedance data.

[0096] It should be noted that EIT (Electrical Impedance Imaging) preprocesses low-order impedance data. Specifically, it converts the impedance data into a two-dimensional conductivity distribution matrix to increase the dimensionality of the features and maps them to a matrix space that facilitates neural network computation. The basic EIT process is as follows: Figure 4 The diagram shows a flowchart of electrical impedance imaging. In this embodiment, the method for preprocessing low-order impedance data based on electrical impedance imaging technology specifically includes:

[0097] The electrode array applies an excitation current to the tissue through the excitation electrode pair and measures the voltage through the acquisition electrode pair. The array impedance data is acquired by combining the excitation electrode pair and the acquisition electrode pair.

[0098] The finite element method is used to solve the forward problem, that is, based on the physical model and boundary conditions, to accurately calculate the distribution of current inside the object and the resulting potential distribution.

[0099] When solving the forward problem using the finite element method:

[0100] In EIT (Earth Interchange Ingress), the forward problem is how to predict the boundary potential distribution using the known internal conductivity distribution. Constructing a mathematical model of the conductivity and boundary potential distributions by solving the forward problem is a crucial step in EIT image reconstruction, providing support for the inverse problem of EIT to reconstruct the conductivity distribution image. This section uses the finite element method to solve the forward problem. This method typically divides the computational domain into a finite number of non-overlapping triangular mesh elements.

[0101] Let the solution domain be Ω. For the potential distribution φ and the conductivity distribution σ, they satisfy the Laplace equation:

[0102]

[0103] in, Denotes divergence, To represent the gradient, after discretizing the solution domain Ω into a finite element mesh, each element e in the mesh is represented by a node i, and there are n corner points in total. Within each element e, the electric potential φ e Approximated by a linear combination of nodal potentials:

[0104]

[0105] Where, φ i N represents the nodal potential. i It is a shape function with a value of 1 at node i and 0 at other nodes. It is used to interpolate the potential distribution within the element, transforming the potential φ... e In the expression of the Laplace equation, the system of equations is obtained from the weak form of the partial differential equation:

[0106] Ku = b,

[0107]

[0108] In the system of equations, K is the system stiffness matrix, K ij Let b be an element in K, and b be the source vector determined by the boundary conditions and the external current source. By solving this system of linear equations, the unknown nodal potential value u can be obtained, thus completing the solution of the forward problem of electrical impedance imaging.

[0109] The Gauss-Newton method is used to solve the EIT inverse problem. By approximating the nonlinear relationship of the objective function, the optimal solution, i.e. the true value of the conductivity distribution, is found iteratively.

[0110] It should be noted that, Figure 4 In this study, it is assumed that a foreign object is contained in the array space covered by electrodes '0' to 'f'. The electrode array applies an excitation current to the tissue through excitation electrodes. Subsequently, a mathematical model between the conductivity distribution and the boundary voltage measurement is established by solving the forward problem. Finally, the two-dimensional image matrix of the conductivity distribution is reconstructed by solving the inverse problem during the iterative process of fitting the forward problem model.

[0111] In this embodiment, the current excitation adopts an adjacent excitation mode. In this mode, current is injected through adjacent electrode pairs, and the potential is measured on the remaining adjacent electrode pairs. This excitation mode will generate different electric field distributions inside the object.

[0112] When using the Gauss-Newton method to solve the inverse EIT problem:

[0113] In the inverse problem of EIT (Electrical Induction Interference), the objective is to reconstruct the conductivity distribution σ inside an object based on potential measurements at the boundary. The Gauss-Newton method, a widely used iterative optimization algorithm for solving nonlinear least squares problems, is employed in EIT image reconstruction to optimize the imaging results, allowing the reconstructed image to better match the actual observation data and thus obtain a more accurate conductivity distribution image. This method is based on Newton's method, approximating the nonlinear relationship of the objective function by iteratively finding the optimal solution, i.e., the true value of the conductivity distribution. The core idea of ​​this method is to continuously iterate, using the first and second derivatives of the objective function to update the parameters and minimize the objective function.

[0114] According to the Dirichlet boundary conditions, the source vector b represents the excitation voltage on the boundary. Let the measured data be d and the model-predicted data be f(σ). The inverse problem is expressed as minimizing the following objective function:

[0115]

[0116] stf(σ):=u,Ku=b,

[0117] That is, we hope to minimize the error between the predicted data f(σ) and the actual measured data d by adjusting the conductivity distribution parameter σ;

[0118] The Gauss-Newton method solves the above least squares problem by iteratively updating the parameter σ. In each iteration, the Jacobian matrix J of the objective function with respect to σ is first calculated, and then σ is updated using a linear approximation. The linear approximation of the objective function is:

[0119] f(σ)≈f(σ k )+J k (σ-σ k ),

[0120] Where the parameter of the current iteration is σ k J k It is in σ k The Jacobian matrix calculated at point is updated by the following equation:

[0121]

[0122] in, It is the coefficient matrix of the normal equation. It is the gradient of the residual.

[0123] EIT data has high dimensionality and complex nonlinear characteristics, making it difficult to effectively represent its features using conventional methods. Neural networks, as a nonlinear modeling tool, possess the ability to approximate arbitrarily complex functions, thus becoming an ideal choice for processing this type of data. In particular, autoencoders (AEs) are widely used in image reconstruction due to their powerful feature representation capabilities and are a common feature extraction method. U-shaped networks have a structure very similar to autoencoders; the downsampling layer (dimensionality reduction) can be considered as the encoder, and the upsampling layer (dimensionality increase) as the decoder. Furthermore, its unique forward propagation structure allows the network to capture sufficient spatial information, enabling the generation of high-order electrode array data from low-order electrode array data. Therefore, the impedance data upscaling and reconstruction module in this application uses a U-shaped neural network to implement the autoencoder, thereby achieving data representation and processing for array upscaling and reconstruction, such as... Figure 5 The diagram shows a schematic of the impedance data upgrade and reconstruction module.

[0124] This invention also provides a method for order-up reconstruction of a two-dimensional image matrix based on a U-shaped autoencoder. The method specifically includes:

[0125] Using the finite element method, each region Ω is modeled as a triangle. The Gauss-Newton impedance imaging method is then used to construct a high-dimensional conductivity distribution map from the 2×2 electrode array data. The conductivity distribution vector generated by the triangular mesh layout is in a non-standard matrix form. To input the conductivity distribution image into the impedance data upscaling reconstruction module, this embodiment uses linear interpolation to standardize the matrix representation of the conductivity distribution image and enhance its fine-grainedness. A comparison of the two mesh construction methods is shown below. Figure 6 As shown, a comparative schematic diagram of the conductivity distribution of a triangular EIT grid and a matrix network is presented.

[0126] The imaging space Ω is divided into a standard 100×100 matrix grid Ω ψ Each element ψ in the grid is a region in its space, i.e., ψ∈Ω ψ The grid is then overlapped with the conductivity distribution image grid generated by EIT, yielding the intersection of the regions in the matrix grid and the EIT grid regions. Since the conductivity value is the same in each EIT grid, the conductivity value of the matrix grid region is the weighted average of its intersection with the EIT grid regions. Therefore, the conductivity value σ in the rectangular grid is... ψ for:

[0127]

[0128] Where, O={e|e∈Ω e ,ψ∈Ω ψ ,e∪ψ≠φ} denotes the triangular EIT mesh Ω that intersects with the region ψ. e The set of regions e, S e∪ψ S represents the area of ​​the intersection of the EIT grid region and the rectangular grid region. ψ σ represents the area of ​​the rectangular grid region. e This represents the conductivity of a region within the EIT grid, thus a standard conductivity distribution image matrix Ψ = (σ ψij ) 100×100 Construction complete.

[0129] The upsampling module constructs an autoencoder according to a U-shaped network structure. The autoencoder is used to convert the low-order conductivity distribution image to the high-order conductivity distribution image, thereby improving the fineness of the image. The autoencoder consists of a downsampling layer and an upsampling layer.

[0130] The downsampling layer extracts features by applying max pooling layers and the previously defined convolutional blocks in two successive stages. During this process, the max pooling layers halve the length and width of the features, allowing the upsampling module to gradually reduce the spatial dimension of the feature map while preserving feature information, thus capturing more global image information. Finally, two convolutional layers further integrate the features.

[0131] The upsampling layer gradually restores the spatial size of the feature map through two transposed convolutional layers, doubling the length and width dimensions each time. At the same time, skip connections are used to concatenate the corresponding feature map from the encoder stage with the feature map from the current decoding stage. This strategy ensures that during the restoration of image details, the detailed information captured in the encoding stage can be utilized without losing the contextual information generated in the decoding stage, thereby achieving accurate reconstruction of the EIT image.

[0132] The upscaling module takes a conductivity image generated by a 2×2 electrode array as input and outputs a high-precision conductivity image after upscaling. Both have the same shape, W×H×C, where W=100, H=100, and C=1. The core of the model consists of a convolutional block, which includes two convolution operations. Each convolution is followed by batch normalization (BN) and a ReLU activation function to enhance the model's nonlinear expressive power and feature learning efficiency. Based on the relationship between the output image size z of the convolutional layer and the convolutional kernel k, stride s, and padding value p, z′=(z+2×pk) / s+1. When k=3, s=1, and p=1, z′=z, meaning that the size of the input image is not modified in terms of length and width, but the feature dimension is transformed by modifying the number of channels. Its output size is Wl×Hl×n×Cl, where n is a constant, and Wl, Hl, and Cl represent the width, height, and channel dimension of the l-th layer, respectively.

[0133] The reconstruction module uses the same U-shaped network structure as the upgrade module to perform upgrade reconstruction of the two-dimensional image matrix.

[0134] In this embodiment, the upgrade module can be viewed as a pre-constraint in the impedance data upgrade reconstruction module's reconstruction process, constraining the features and results of the intermediate process to achieve the reconstruction of 4×4 higher-order impedance data. The reconstruction module and the upgrade module employ the same U-shaped autoencoder structure. Since the reconstruction module's model output is the same size as the input image (100×100×1), the upgrade module requires two convolutional layers to reduce the data dimensionality to the same level as the 4×4 higher-order impedance data. The impedance data upgrade reconstruction module implements the dimensionality upgrade process of the acquisition array during the reconstruction of the array data. Regarding the output of the reconstruction module...

[0135] It should be noted that, regarding the reconstruction data construction method, this application designs two matrix forms, one of which is a matrix representation R = (r ij ) N×N In this representation, rows represent excitation electrode pairs, and columns represent acquisition electrode pairs. Each row indicates that an excitation current is added to electrode pair i. The process is repeated for all electrode pairs, and impedance data is acquired. The next method is R′=(r ijk ) 4×7×N This array representation constructs an array on each pair of excitation electrodes according to the spatial arrangement of the electrodes. Since the data is not a square array in the spatial array, and the data length of each row alternates between 3 and 4, zero-padding is used to supplement the data to form a standard matrix. See details... Figure 7 , Figure 7 The diagram illustrates the reconstruction data construction method. The matrix array data representation is relatively simple and direct, while the spatial design method is more complex, but it contains spatial feature information and is more in line with the actual array data arrangement.

[0136] It should be noted that the mean square error loss is used to train the impedance data upgrade reconstruction module. For the high-order conductivity distribution image and 4×4 electrode array data generated by the upgrade module, the mean square error loss is calculated separately from the actual measured values. The summation of these values ​​yields the complete loss function expression:

[0137]

[0138] in, and σ ψij These represent the generated conductivity distribution matrix, respectively. The elements in the i-th row and j-th column of the conductivity distribution label matrix Ψ are used. The matrix size is M×M, where M=100. and r ij These represent the reconstructed 4×4 electrode array data. The elements in the i-th row and j-th column of the actual 4×4 electrode array data R are used to construct a matrix of size M×N, where N=16. Figure 7 (b) shows the spatial representation matrix R′, which, after removing the padding values, is rearranged as follows: Figure 7 The matrix shown in (a) represents R for easy computation.

[0139] In this embodiment of the invention, the Diff-PNA classification model uses a graph neural network method to extract features and classify the constructed graph topology. First, the Diff-PNA classification model receives low-order data collected from a 2×2 electrode array and constructs a conductivity distribution image from it using an impedance data order-up reconstruction module, further reconstructing it into high-order boundary impedance data for a 4×4 electrode array. Then, a graph construction strategy is used to map the reconstructed high-order data into graph data in non-Euclidean space, forming a graph topology with N nodes. Finally, the Diff-PNA classification model uses a graph neural network method to extract features and classify the graph constructed in the previous stage.

[0140] The Diff-PNA classification model employs a 5-layer concatenated design, which can be simplified as follows:

[0141]

[0142] Finally, a global average pooling method is used to obtain the embedded features of the entire image, and a 5-layer multilayer perceptron is used to achieve the final classification result X. out Output.

[0143] The Diff-PNA classification model uses a 5-layer multilayer perceptron to achieve the final classification result X. out The output is represented as follows:

[0144]

[0145] in, X represents the characteristics of node j output by the 5th layer PNA network. p This is an embedded feature of the entire image.

[0146] In this embodiment, to supplement blood glucose fluctuation information, a natural feature extraction module is introduced. The natural feature extraction module is used to extract low-order array data and realize the conversion between low-order array data and high-order array data.

[0147] In this embodiment, 2×2 is used as the low-order array and 4×4 is used as the high-order array. In fact, the order can be changed arbitrarily. When the low-order array data is a 2×2 low-order array, the natural feature extraction module extracts the most original low-order 2×2 electrode array data features and splices them with the PNA features of the 4th layer.

[0148] The natural feature extraction module is used to extract the features of the original low-order 2×2 electrode array data. When concatenating these features with the PNA features of the fourth layer, the acquisition matrix of the original low-order 2×2 electrode array is first input. The acquisition matrix is ​​constructed as follows: Figure 7 As shown in (b), the acquisition matrix is ​​described as R = (R ijk )4×3×3 The natural feature extraction module uses two deconvolutional layers to expand the acquisition matrix of the 2×2 electrode array, and then uses three convolutional layers and a flattening layer to transform the features into a one-dimensional vector representation, described as follows: This feature is directly derived from the acquisition matrix of the 2×2 electrode impedance array and is directly concatenated with the fourth layer PNA feature of the main classification model. Therefore, the fifth layer PNA of the model is represented as follows:

[0149]

[0150] Here, O represents the horizontal concatenation of two matrix dimensions and features.

[0151] It should be noted that the array-based bioelectrical impedance blood glucose monitoring model consists of multiple modules, each with a unique function. To ensure that the array-based bioelectrical impedance blood glucose monitoring model can effectively backpropagate and be optimized across multiple modules, this application designs a multi-scale loss function. By introducing a mechanism with different weights, it manages and optimizes the performance of data at various levels, thereby enhancing the model's comprehensive understanding and processing capabilities of the data.

[0152] Specifically, the array-based bioelectrical impedance blood glucose monitoring model is constrained on three scales: the classification effect of the high-order array impedance data reconstruction results, the classification effect of the Diff-PNA network, and the upgrade reconstruction effect of the low-order array data.

[0153] In the reconstruction results of high-order array impedance data, since the impedance data upgrade reconstruction module is not specifically designed for blood glucose classification, its enhanced features may ignore some blood glucose fluctuation information contained in the data itself and introduce some noise. To make the reconstruction results more suitable for blood glucose classification, the multi-scale constraint module incorporates constraints for the blood glucose classification task during the impedance data reconstruction stage. This reduces the impact of new noise introduced during the low-order impedance data upgrade reconstruction process. The multi-scale constraint module consists of 5 layers of convolutional layers, flattening layers, and a 5-layer multilayer perceptron. It takes the high-order 4×4 electrode array data results from the impedance data upgrade reconstruction module as input and outputs the blood glucose classification results. The output of this process is independent of the output of the overall framework model, aiming to constrain the impedance data upgrade reconstruction module and reduce the loss of blood glucose concentration features during reconstruction.

[0154] Therefore, the cross-entropy loss function can be defined as L1, L2, and the mean squared error loss function as L3 for the three scales respectively:

[0155]

[0156] L3: = L ref ,

[0157] Where p = [p0, ..., p] c-1 ] is a probability distribution, where each element p i This represents the probability that a sample belongs to the i-th class. The classification results at two scales, y = [y0, ..., y c-1 ] is the one-hot encoded representation of the sample label, where y represents the sample belonging to class i. i =0, otherwise y i =1, where C is the number of sample labels. L ref To constrain the process of order reconstruction and to balance the contributions of each scale and improve the overall performance of the model, this method assigns a weight α to each loss function. s Where s represents the scale, S is the set of all scales, and the weights are constrained to sum to 1. Therefore, the comprehensive multi-scale loss function L... ms This can be expressed as a weighted sum of all single-scale losses:

[0158]

[0159] Where, α s This not only reflects the relative importance of each scale in the total loss, but can also be used to adjust the learning rate across different scales, ensuring that the model can be effectively trained on all scales. To further improve the model's generalization ability and prediction consistency, a regularization term L is introduced into the multi-scale loss function. reg This is to reduce generalization error and avoid model overfitting. Therefore, the final loss function L final This can be further expanded to:

[0160] L final =L ms +λ1L reg ,

[0161]

[0162] Where λ1 is the penalty term coefficient, and θ i These are the model parameters.

[0163] In this embodiment of the invention, when constructing the array-based bioelectrical impedance blood glucose monitoring model, the model is iteratively trained using a training set and a test set. Two optimizers are used for the impedance data reconstruction module, the natural feature extraction module, the multi-scale constraint module, and the Diff-PNA classification model, respectively. Both optimizers use adaptive moment estimation, and the first and second moments that track the gradient are expressed as follows:

[0164]

[0165]

[0166] m t+1 =β1·m t +(1-β1)·G t ,

[0167] v t+1 =β2·v t +(1-β2)·(G t ) 2 ,

[0168]

[0169] Where η is the learning rate, and G t The loss function at the current parameter θ t gradient at m t and v t These are the first and second moment estimates of the gradient, respectively, where β1 and β2 are decay rate parameters, and ε = 10. -8 It is a constant to prevent division by zero, and ⊙ represents the Hadamard product;

[0170] The parameter set of the array-based bioelectrical impedance blood glucose monitoring model is defined as follows:

[0171] Θ:={Θ rebuild ,Θ na_feature ,Θ PNA ,Θ constraints}

[0172] Where Θ rebuild ,Θ na_feature ,Θ PNA ,Θ constraints Let represent the parameters of the impedance data order reconstruction module, the Diff-PNA classification model, the natural feature extraction module, and the multi-scale constraint network module, respectively. Under the multi-scale loss function, the gradients at different locations of the model are obtained as follows:

[0173]

[0174] Therefore, the two classifiers are represented as follows:

[0175]

[0176] Among them, f is used Adma1 This represents the optimizer for the impedance data up-order reconstruction module, where η1 represents its learning rate, f Adma2 η represents the optimizer for other modules, and η2 represents its learning rate.

[0177] To verify the performance of the variable-order array-based non-invasive blood glucose detection method based on graph neural networks provided in this invention, this invention uses an impedance simulator to simulate and analyze impedance data, thereby verifying the application effect of the Gauss-Newton method on impedance imaging datasets. Furthermore, an impedance-order reconstruction dataset is constructed based on this, including the following steps:

[0178] A uniform rectangular physical medium of 10×10 cm² was constructed on the simulator. This medium was used to generate the basic dataset for electrical impedance imaging, simulating different electrical impedance conditions encountered in real-world scenarios. By introducing specific foreign objects or anomalous regions into this medium, common challenges in electrical impedance imaging can be simulated.

[0179] Foreign objects with different conductivity characteristics from the background medium are randomly generated in a homogeneous medium, thus producing identifiable signal differences during electrical impedance imaging. These foreign objects are varied in five aspects: shape, location, size, number, and conductivity, generating a series of complex and challenging electrical impedance imaging scenarios.

[0180] Different electrode arrays and excitation methods were used to acquire electrical impedance data. Two types of electrode arrays were used in the simulation. The first was a low-order 2×2 electrode array, employing a combination of adjacent and relative excitation methods. All electrodes were paired to form excitation and acquisition electrode pairs, greatly enriching the acquired data. The second was a high-order 4×4 electrode array. With sufficient data complexity and granularity, the simpler adjacent excitation method was used for ease of data representation. Under the Gauss-Newton method, this method produced a conductivity distribution image with significantly higher clarity compared to the 2×2 electrode array data. Figure 8 , Figure 8 A schematic diagram comparing the fine-grained imaging of different array data is shown, in which... Figure 8 (a) shows impedance imaging under a 2×2 electrode array. Figure 8 (b) shows impedance imaging under a 4×4 electrode array.

[0181] The impedance data of the 2×2 electrode array and the impedance distribution map generated from the data are used as training data, and the impedance data of the 4×4 electrode array and the impedance distribution map generated from the data are used as labels to construct an upgraded reconstruction dataset.

[0182] In this embodiment, when iteratively training the array-type bioelectrical impedance blood glucose monitoring model using training and test sets, the ratio of training set to test set is 4:1. Mean squared error is used to estimate the final classification result and the order reconstruction. Mean squared error, peak signal-to-noise ratio, structural similarity index, and cosine similarity are used to evaluate the impedance data order reconstruction module.

[0183] The mean squared error measures the closeness between the 4×4 electrode reconstruction values ​​of the array-based bioelectrical impedance glucose monitoring model and the data acquired by the real-label 4×4 electrode array, and its formula is expressed as:

[0184]

[0185] Where I(i,j) and K(i,j) are the values ​​of the observation array data at (i,j) respectively, and m and n are the number of rows and columns of the observation array data respectively;

[0186] Peak signal-to-noise ratio (PSNR) evaluates the error between corresponding pixels in two images, i.e., error-sensitive image quality assessment, and is expressed as:

[0187]

[0188] Among them, MAX I It is the maximum possible pixel value in the image data;

[0189] The structural similarity index compares the similarity between two images, including brightness, contrast, and structural information, and is expressed as:

[0190]

[0191] Where, μ I and μ K These represent the average brightness of images I and K, respectively. and Let σ represent the variances of images I and K, respectively. IK Let C1 and C2 represent the covariance of images I and K. C1 and C2 are small constants introduced to avoid the denominator being zero. Let C1 = (k1L). 2 C2 = (k2L) 2 Where L is the dynamic range of pixel values, and k1 = 0.01 and k2 = 0.03 are default constants;

[0192] Cosine similarity compares the similarity of two distributions in a spatial direction, evaluating the directional similarity of conductivity distributions rather than their absolute values, and is expressed as:

[0193]

[0194] Where I·K represents the dot product of two image vectors, and ||I|| and ||K|| represent the magnitudes of I and K, respectively.

[0195] It should be noted that in this method, the mean square error is obtained by directly calculating the difference in the array acquisition data, while the peak signal-to-noise ratio, structural similarity index, and cosine similarity are three indicators that compare the array acquisition data after electrical impedance imaging processing.

[0196] The specific results are shown in Table 1. The first matrix representation method, R, has the simplest structure and the smallest mean squared error, only 0.012. Furthermore, it demonstrates superior similarity in terms of peak signal-to-noise ratio (PSNR), structural similarity index, and cosine similarity (cosine similarity) (35.2, 0.85, 0.93, respectively). Overall, the first matrix construction method exhibits a simpler matrix structure and is easier to learn.

[0197] Table 1 Comparison of the two reconstruction matrix representation methods

[0198] Matrix representation MSE PSNR SSIM CS Matrix representation of matrix R 0.012 35.2 0.85 0.93 Spatial representation matrix R′ 0.018 33.5 0.78 0.91

[0199] In this embodiment, the impedance image is reconstructed using the Gauss-Newton method based on the reconstructed 4×4 array impedance data. Figure 9 This demonstrates the effect of upgraded high-order array impedance data on reconstruction: Figure 9 (a) shows the actual electrical impedance background image; Figure 9 (b) shows the actual 2×2 order array impedance image; Figure 9 (c) shows the actual 4×4 array impedance image; Figure 9 (d) shows the generated 4×4 array impedance image. The 4×4 array impedance imaging position is basically correct, but compared to a true 4×4 array impedance image, the generated image appears blurry, lacks fine granularity, and has indistinct boundaries. However, compared to a true 2×2 array impedance image, the generated image has higher granularity. This indicates that the model has good data reconstruction capabilities and can achieve up-order reconstruction of data.

[0200] On the other hand, based on the implementation of the impedance data upgrade and reconstruction module, this embodiment also verifies the effectiveness of the array-based bioelectrical impedance analysis (BIA) glucose monitoring model. Specifically, the method for constructing the in vitro simulated dataset is first described, under which 1000 data points of 11 glucose concentrations were collected. Then, the role of each module in the model is demonstrated through ablation experiments. Finally, the effectiveness of the proposed model is verified by comparing it with conventional neural network and machine learning methods.

[0201] Among them, the data acquisition results of the in vitro blood glucose simulation system are as follows:

[0202] The in vitro blood glucose simulation system collects in vitro blood glucose simulation data. The data acquisition process is as follows: First, an agar module with hollow tubing is prepared, along with 11 different glucose solutions ranging from 0 g / L to 10 g / L. Then, three motor speed modes are set: stationary, 300 rpm, and 500 rpm. The excitation frequency is set to start at 10 kHz and increased logarithmically by 50 cycles to 150 kHz. Finally, the three configurations are combined, and impedance data are collected using adjacent excitation methods with 2×2 electrode arrays and 4×4 electrode arrays, respectively, under the same configuration.

[0203] This paper proposes a 4×4 electrode array for the acquisition system. To control variables and prevent unknown interference from electrode replacement, the 2×2 electrode array in this experiment consists of four electrodes from the 4×4 array: '0', '3', 'c', and 'f'.

[0204] Specifically, in the method for acquiring adjacent excitations of the 4×4 electrode array, there are 24 electrode pairs to be selected, namely {'01', '04', '12', '25', ..., 'ef'}. The 2×2 electrode array data only uses the {'03', '0c', '3f', 'cf'} electrode pairs under the 4×4 electrode array to complete the data acquisition of the 4×4 electrode array and the 2×2 electrode array.

[0205] To verify the effectiveness of each module in the array-based bioelectrical impedance blood glucose monitoring model proposed in this application, this section uses an ablation experiment to remove a part or function of the model and observe its impact on model performance. Specifically, this section focuses on three key modules: the natural feature extraction module, the multi-scale constraint module, and the impedance data upgrade reconstruction module. The following is a detailed description of the combination of the natural feature extraction module, the multi-scale constraint module, and the impedance data upgrade reconstruction module:

[0206] 1) Both modules include: This setup includes a natural feature extraction module and a multi-scale constraint module. This fully functional model allows for evaluation of the model's optimal performance when all design components are integrated.

[0207] 2) Includes only the natural feature extraction module: In this setting, the model includes only the natural feature extraction module, while the multi-scale constraint module is excluded. This helps to understand the impact of relying solely on natural feature extraction on model performance.

[0208] 3) Include only the multi-scale constraint module: This configuration retains only the multi-scale constraint module while removing the natural feature module. This allows for the evaluation of the contribution of multi-scale processing to model performance in the absence of natural feature support.

[0209] 4) Neither module is included: Completely remove both modules and experiment using only the base model. This setup can serve as a baseline to help you understand the extent of the improvements brought about by adding these modules.

[0210] Based on the four schemes mentioned above, a total of eight ablation comparison experiments were conducted, and the results are shown in Table 2. Analysis of the table shows that the model with the inclusion of the natural feature extraction module, the multi-scale constraint module, and the impedance data upgrade reconstruction module performed best, achieving an accuracy of 91.72%. Further analysis indicates that when the natural feature extraction module and the multi-scale constraint module are not used, the pre-training of the impedance data upgrade reconstruction module has little impact on the model's accuracy and precision. However, when all modules are combined, the weights of the pre-trained impedance data upgrade reconstruction module have a significant optimization effect on the blood glucose classification framework. Without considering the pre-trained model with the impedance data upgrade reconstruction module, the accuracy of the natural feature extraction module and the multi-scale constraint module, when used independently, both showed a significant improvement, increasing by 10.18% and 10.18%, respectively. This demonstrates the role of these three modules in the blood glucose classification framework, effectively improving the model's performance.

[0211] Table 2 Comparison of ablation test results

[0212]

[0213]

[0214] Benchmark Comparison Analysis:

[0215] The array-based bioelectrical impedance analysis (BIA) glucose monitoring model was compared in detail with several conventional models, including LSTM, CNN, GCN, the PNA neural network classifier without upgrade processing (PNA_no_upgrade), and SVM. Specifically, PNA_no_upgrade represents a PNA neural network classifier without an upgrade module. The comparison results are detailed in Table 3.

[0216] Table 3 Comparison of detection results between conventional methods and this method.

[0217] Model Accuracy Precision Recall F1 SVM 83.21 84.42 82.11 83.57 LSTM 85.79 86.27 84.43 85.73 CNN 87.39 88.63 86.96 87.22 GCN 88.40 89.02 87.94 88.23 PNA_no_upgrade 90.49 91.54 89.35 90.65 This model 91.72 92.04 91.89 91.76

[0218] As shown in Table 3, the proposed array-based bioelectrical impedance analysis (BIA) glucose monitoring model outperformed other conventional models across all evaluation metrics, achieving an accuracy of 91.72%, precision of 92.04%, recall of 91.89%, and an F1 score of 91.76%. Compared to traditional machine learning methods (SVM) and other deep learning methods (LSTM, CNN), different types of graph neural networks (GCN, PNA_no_upgrade, and the array-based BIA glucose monitoring model) all achieved better results, validating the powerful ability of graph neural networks in processing array-based data. Furthermore, compared to the PNA neural network classifier PNA_no_upgrade without upgrade reconstruction, the proposed model improved accuracy by 1.23%, precision by 0.50%, recall by 2.54%, and F1 score by 1.11%, indicating that the upgrade module significantly contributes to improving model performance. In summary, the array-based BIA glucose monitoring model can deeply mine complex information within the data, extract effective features from low-level feature information, and possesses strong classification capabilities.

[0219] In summary, this invention provides a variable-order array-based non-invasive blood glucose detection method based on graph neural networks. In this embodiment, a low-order implementation strategy for a high-order array is proposed, and an impedance data order-up reconstruction module for low-order array impedance data is constructed. This realizes a variable-order array-based bioelectrical impedance blood glucose monitoring model. The array-based bioelectrical impedance blood glucose monitoring model uses Electrical Impedance Tomography (EIT) to perform preprocessing processes such as feature expansion of low-order data. Based on this, a resampling process from high-order impedance images to low-order arrays is implemented. Furthermore, a natural feature extraction module and multi-scale constraint strategies are integrated, further improving the robustness of the array-based bioelectrical impedance non-invasive blood glucose detection model, as well as the efficiency and accuracy of non-invasive blood glucose detection.

[0220] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0221] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A variable-order array-based non-invasive blood glucose detection method based on graph neural networks, characterized in that, include: Low-order impedance data is acquired based on an array-type bioelectrical impedance front-end. Low-order impedance data is loaded in real time. Using the low-order impedance data as input, an array-based bioelectrical impedance glucose monitoring model is executed. The array-based bioelectrical impedance glucose monitoring model performs up-order reconstruction and feature extraction processing on the low-order impedance data and outputs the glucose detection results.

2. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 1, characterized in that: The array-based bioelectrical impedance glucose monitoring model includes: Impedance data order-up and reconstruction module, which is used to reconstruct low-order impedance data into high-order data and output high-order array impedance data. The impedance data order-up and reconstruction module includes a data matrix module, an EIT conductivity distribution conversion module, an order-up module, and a reconstruction module. The order-up module and the reconstruction module each include a U-shaped autoencoder, which is composed of an encoder, a decoder, and a convolutional block. The natural feature extraction module is used to supplement the signal features of low-order impedance data to compensate for the influence of ignored blood glucose fluctuation information on the classification results. A multi-scale constraint module is used to add constraint requirements for the blood glucose classification task in the low-order impedance data reconstruction stage and reduce the introduction of new noise in the process of upgrading the low-order impedance data. A topology module, which is used to convert high-order array impedance data into graph topology; The Diff-PNA classification model is used to obtain graph topology and, based on the Diff-PNA algorithm, splices natural features to achieve blood glucose concentration classification.

3. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 2, characterized in that: Methods for reconstructing high-order data from low-order impedance data include: Based on electrical impedance imaging technology, low-order impedance data are preprocessed to reconstruct a two-dimensional image matrix of conductivity distribution. The reconstructed two-dimensional image matrix is ​​obtained, and the two-dimensional image matrix is ​​reconstructed by an up-order U-shaped autoencoder to output high-order array impedance data.

4. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 3, characterized in that: The method for preprocessing low-order impedance data based on electrical impedance imaging technology specifically includes: The electrode array applies an excitation current to the tissue through the excitation electrode pair and measures the voltage through the acquisition electrode pair. The array impedance data is acquired by combining the excitation electrode pair and the acquisition electrode pair. The finite element method is used to solve the forward problem, that is, based on the physical model and boundary conditions, to accurately calculate the distribution of current inside the object and the resulting potential distribution. The Gauss-Newton method is used to solve the inverse EIT problem. By approximating the nonlinear relationship of the objective function, the optimal solution, i.e. the true value of the conductivity distribution, is found iteratively.

5. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 4, characterized in that: The method for order-up reconstruction of a two-dimensional image matrix based on a U-shaped autoencoder specifically includes: Using the finite element method, each region Ω is modeled as a triangle, and the Gauss-Newton method is used for electrical impedance imaging to construct a high-dimensional conductivity distribution map from the 2×2 electrode array data, dividing the imaging space Ω into a standard 100×100 matrix grid Ω. ψ Each element ψ in the grid is a region in its space, i.e., ψ∈Ω ψ The grid is then overlapped with the conductivity distribution image grid generated by EIT, yielding the intersection of the regions in the matrix grid and the EIT grid regions. Since the conductivity value is the same in each EIT grid, the conductivity value of the matrix grid region is the weighted average of its intersection with the EIT grid regions. Therefore, the conductivity value σ in the rectangular grid is... ψ for: Where, O={e|e∈Ω e ,ψ∈Ω ψ ,e∪ψ≠φ} denotes the triangular EIT mesh Ω that intersects with the region ψ. e The set of regions e, S e∪ψ S represents the area of ​​the intersection of the EIT grid region and the rectangular grid region. ψ σ represents the area of ​​the rectangular grid region. e This represents the conductivity of a region within the EIT grid, thus a standard conductivity distribution image matrix Ψ = (σ ψij ) 100×100 Construction complete; The upsampling module constructs an autoencoder according to a U-shaped network structure. The autoencoder is used to convert the low-order conductivity distribution image to the high-order conductivity distribution image. The autoencoder consists of a downsampling layer and an upsampling layer. The downsampling layer extracts features by applying max pooling layers and convolutional blocks defined above in two successive stages. The upsampling layer gradually restores the spatial size of the feature map by two transposed convolutional layers, expanding it by a factor of 2 in length and width each time. At the same time, skip connections are used to concatenate the corresponding feature map of the encoder stage with the feature map of the current decoding stage. The upscaling module takes a conductivity image generated by a 2×2 electrode array as input and outputs a high-precision conductivity image after upscaling. Both have the same shape, W×H×C, where W=100, H=100, and C=1. The core of the model consists of a convolutional block. This module includes two convolution operations, followed by batch normalization and ReLU activation function after each convolution to enhance the nonlinear expressive power and feature learning efficiency of the model. According to the relationship between the output image size z of the convolutional layer and the convolutional kernel k, stride s, and padding value p, z′=(z+2×pk) / s+1. When k=3, s=1, and p=1, z′=z, that is, the size of the input image is not modified in terms of length and width. The feature dimension is transformed by modifying the number of channels. Its output size is Wl×Hl×n×Cl, where n is a constant, and Wl, Hl, and Cl represent the width, height, and channel dimension of the l-th layer, respectively. The reconstruction module uses the same U-shaped network structure as the order-up module to perform order-up reconstruction of the two-dimensional image matrix; The upgrade module is a pre-constraint term in the impedance data upgrade reconstruction module. It constrains the features and results of the intermediate process to achieve the reconstruction of 4×4 high-order impedance data. The reconstruction module and the upgrade module adopt the same U-shaped autoencoder structure. Since the output of the reconstruction module model is the same size as the input image (100×100×1), two convolutional layers are needed in the upgrade module to reduce the dimensionality of the data to the same level as the 4×4 high-order impedance data. The impedance data upgrade reconstruction module realizes the dimensionality upgrade process of the acquisition array during the reconstruction of the array data.

6. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 2, characterized in that: The Diff-PNA classification model uses a graph neural network method to extract features and classify the constructed graph topology; The Diff-PNA classification model uses a 5-layer multilayer perceptron to achieve the final classification result X. out The output is represented as follows: in, X represents the characteristics of node j output by the 5th layer PNA network. p This is an embedded feature of the entire image.

7. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 6, characterized in that: The natural feature extraction module is used to extract low-order array data and to convert between low-order and high-order array data. When the low-order array data is a 2×2 low-order array, the natural feature extraction module extracts the most original low-order 2×2 electrode array data features and splices them with the PNA features of the 4th layer. The natural feature extraction module is used to extract the features of the original low-order 2×2 electrode array data. When concatenating these features with the PNA features of the fourth layer, the acquisition matrix of the original low-order 2×2 electrode array is first input. The acquisition matrix is ​​described as R = (R ijk ) 4×3×3 The natural feature extraction module uses two deconvolutional layers to expand the acquisition matrix of the 2×2 electrode array, and then uses three convolutional layers and a flattening layer to transform the features into a one-dimensional vector representation, described as follows: This feature is directly derived from the acquisition matrix of the 2×2 electrode impedance array and is directly concatenated with the fourth layer PNA feature of the main classification model. Therefore, the fifth layer PNA of the model is represented as follows: Here, O represents the horizontal concatenation of two matrix dimensions and features.

8. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 2, characterized in that: The multi-scale constraint module incorporates constraints for the blood glucose classification task during the impedance data reconstruction stage to compensate for the impact of ignored blood glucose fluctuation information on the classification results. The multi-scale constraint module consists of 5 layers of convolutional layers, flattening layers, and 5 layers of multilayer perceptron. The multi-scale constraint module takes the high-order 4×4 electrode array data results from the impedance data upscaling reconstruction module as input and outputs the blood glucose classification results.

9. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 2, characterized in that: When constructing the array-based bioelectrical impedance analysis (BIA) model, iterative training was conducted using training and testing sets. Two optimizers were used for the impedance data reconstruction module, natural feature extraction module, multi-scale constraint module, and Diff-PNA classification model, respectively. Both optimizers employed adaptive moment estimation, and the first and second moments, which tracked the gradient, were expressed as follows: m t+1 =β1·m t +(1-β1)·G t , v t+1 =β2·v t +(1-β2)·(G t ) 2 , Where η is the learning rate, and G t The loss function at the current parameter θ t gradient at m t and v t These are the first and second moment estimates of the gradient, respectively, where β1 and β2 are decay rate parameters, and ε = 10. -8 It is a constant to prevent division by zero, and ⊙ represents the Hadamard product; The parameter set of the array-based bioelectrical impedance blood glucose monitoring model is defined as follows: Θ:={Θ rebuild ,I na_feature ,I PNA ,I constraints } Where Θ rebuild ,Θ na_feature ,Θ PNA ,Θ constraints Let represent the parameters of the impedance data order reconstruction module, the Diff-PNA classification model, the natural feature extraction module, and the multi-scale constraint network module, respectively. Under the multi-scale loss function, the gradients at different locations of the model are obtained as follows: Therefore, the two classifiers are represented as follows: Among them, use The optimizer represents the impedance data up-order reconstruction module, and η1 represents its learning rate. η represents the optimizer for other modules, and η2 represents its learning rate.

10. The variable-order array-based non-invasive blood glucose detection method based on graph neural networks as described in claim 9, characterized in that: When iteratively training the array-type bioelectrical impedance glucose monitoring model using training and test sets, the ratio of training set to test set is 4:

1. Mean squared error is used to estimate the final classification result and the order reconstruction. Mean squared error, peak signal-to-noise ratio, structural similarity index, and cosine similarity are used to evaluate the impedance data order reconstruction module. The mean squared error measures the degree of similarity between the 4×4 electrode reconstruction values ​​of the array-based bioelectrical impedance glucose monitoring model and the data acquired by the real-label 4×4 electrode array, and its formula is expressed as: Where I(i,j) and K(i,j) are the values ​​of the observation array data at (i,j) respectively, and m and n are the number of rows and columns of the observation array data respectively; Peak signal-to-noise ratio (PSNR) evaluates the error between corresponding pixels in two images, i.e., error-sensitive image quality assessment, and is expressed as: Among them, MAX I It is the maximum possible pixel value in the image data; The structural similarity index compares the similarity between two images, including brightness, contrast, and structural information, and is expressed as: Where, μ I and μ K These represent the average brightness of images I and K, respectively. and Let σ represent the variances of images I and K, respectively. IK Let C1 and C2 represent the covariance of images I and K. C1 and C2 are small constants introduced to avoid the denominator being zero. Let C1 = (k1L). 2 C2 = (k2L) 2 Where L is the dynamic range of pixel values, and k1 = 0.01 and k2 = 0.03 are default constants; Cosine similarity compares the similarity of two distributions in a spatial direction, evaluating the directional similarity of conductivity distributions rather than their absolute values, and is expressed as: Where I·K represents the dot product of two image vectors, and ||I|| and ||K|| represent the magnitudes of I and K, respectively.

Images