Noninvasive blood glucose detection system and equipment based on photoelectric signal and impedance spectrum signal

Abstract

The invention discloses a non-invasive blood glucose detection system and equipment based on photoelectric signals and impedance spectrum signals, and the system comprises an acquisition module which is used for acquiring the photoelectric signals and the impedance spectrum signals of the finger end part of a to-be-detected object; the feature extraction module is used for respectively extracting the features of the photoelectric signals and the features of the impedance spectrum signals; and the feature fusion module is used for carrying out feature fusion on the photoelectric signal features and the impedance spectrum signal features and then predicting the fused features to obtain a blood glucose prediction result. The invention innovatively provides noninvasive blood glucose detection equipment and method based on combination of red visible light, green visible light, blue visible light, near-infrared light, an impedance spectroscopy method and a deep learning technology. In the system, two different modes of PPG detection and impedance spectroscopy are ingeniously integrated, physical information obtained by the PPG detection and impedance spectroscopy is fully fused, the problem of uncertainty of data in a single detection mode is effectively solved, and then the accuracy and reliability of a detection result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of blood glucose monitoring, and particularly to a non-invasive blood glucose detection system and device based on optoelectronic signals and impedance spectrum signals. Background Art

[0002] Diabetes, as a long-term chronic metabolic disease, currently has no complete cure. This means that patients need to continuously monitor and manage their conditions, mainly by regularly checking blood glucose concentrations, so as to adjust control strategies accordingly and maintain blood glucose at a stable level. Therefore, finding an efficient and convenient blood glucose detection method is crucial for diabetes management.

[0003] Traditional blood glucose meters detect the glucose concentration in blood based on enzymatic reactions and electrochemical reactions. This requires patients to perform multiple finger pricks every day, which not only brings physical discomfort and psychological burden to patients, but also may increase the risk of infection due to frequent skin punctures. Moreover, it cannot comprehensively reflect the changing trend of blood glucose levels over time.

[0004] In the field of blood glucose monitoring, the limitations of traditional methods have given rise to a research hotspot in non-invasive blood glucose detection technology. In recent years, technologies such as optical methods, metabolic heat methods, and bioelectrical impedance spectroscopy have continued to progress, strongly promoting the development of non-invasive blood glucose detection devices.

[0005] Among them, optical technology shows unique potential with its non-invasive nature and high accuracy. Blood glucose detection technologies based on spectral analysis, such as near-infrared, mid-infrared, and Raman spectroscopy, etc., have been widely studied and applied. The bioelectrical impedance spectroscopy method can effectively correlate physiological and pathological information by monitoring the human impedance spectrum, and has obvious advantages in terms of real-time performance, power consumption, and multi-parameter measurement. Therefore, in view of the requirements for non-invasiveness, portability, and high precision in blood glucose detection, organically integrating optical technology and bioelectrical impedance spectroscopy method, and performing non-invasive blood glucose detection based on the multi-modal of optical signals and bioelectrical impedance spectrum signals becomes an optimal solution. This integration is not a simple addition, but through synergistic effects, giving full play to the advantages of the two technologies and overcoming their respective limitations, which is expected to achieve significant breakthroughs in aspects such as the accuracy and convenience of blood glucose detection, provide more effective support for clinical applications and patient health management, and promote the blood glucose monitoring technology to a new stage of development.

[0006] Prior to this, there have been a small number of patents and literatures using optical signals or bioimpedance signals to monitor blood glucose, but they all use a single optical signal or bioimpedance signal. For example, the publication number is CN117838114A, and the patent name is: A non-invasive blood glucose detector for human body and detection method. An LED light source in the range of 878 - 1554nm with 4 bands is used to emit near-infrared light. The optimal detection pressure is achieved by adjusting the fixture unit to collect the photoplethysmogram signal. The data processing unit uses a multi-model fusion algorithm with feature-optimized weighted Euclidean distance to predict the blood glucose value. For example, the publication number is CN117617960A, and the patent name is: A circuit structure and method for wearable non-invasive blood glucose detection, which is based on bioimpedance spectroscopy signals to detect blood glucose values. These are all blood glucose detection devices based on a single optical signal or impedance spectroscopy signal. However, the anti-interference ability of a single optical signal or bioimpedance signal is weak and is easily affected by other components in the skin (such as moisture, fat, hemoglobin, etc.) and the temperature and humidity of the test environment, and there are large individual differences, thus affecting the measurement accuracy of blood glucose. Summary of the Invention

[0007] To solve the deficiencies of the prior art, the present invention provides a non-invasive blood glucose detection system and device based on optoelectronic signals and impedance spectroscopy signals; the present invention innovatively proposes a non-invasive blood glucose detection device and method based on the combination of three visible lights of red, green, and blue, near-infrared light, impedance spectroscopy method, and deep learning technology. In this system, by skillfully integrating two different modalities of PPG detection and impedance spectroscopy method, fully fusing the physical information obtained by them, effectively making up for the uncertainty problem of data in a single detection mode, and thus significantly improving the accuracy and reliability of the detection results.

[0008] On the one hand, a non-invasive blood glucose detection system based on optoelectronic signals and impedance spectroscopy signals is provided, including:

[0009] An acquisition module, which is configured to: acquire the optoelectronic signal and impedance spectroscopy signal at the finger tip of the object to be detected;

[0010] A feature extraction module, which is configured to: extract the features of the optoelectronic signal and the features of the impedance spectroscopy signal respectively;

[0011] A feature fusion module, which is configured to: perform feature fusion on the features of the optoelectronic signal and the features of the impedance spectroscopy signal, and then predict the fused features to obtain a blood glucose prediction result.

[0012] On the other hand, a non-invasive blood glucose detection device based on optoelectronic signals and impedance spectroscopy signals is provided, including:

[0013] An optoelectronic signal acquisition subsystem and an impedance spectroscopy signal acquisition subsystem;

[0014] The optoelectronic signal acquisition subsystem includes: a visible light LED and a near-infrared LED. The visible light LED and the near-infrared LED emit light to a lens, the lens transmits the light to the first end of an emission optical fiber, the emission optical fiber transmits the light to the second end of the emission optical fiber, and the second end of the emission optical fiber divides the light into multiple beams and emits them to the skin at the finger tip of the object to be detected. The first end of a receiving optical fiber receives each beam of reflected light feedback from the skin at the finger tip, the second end of the receiving optical fiber is connected to the first end of a photodiode, the second end of the photodiode is connected to an analog-to-digital conversion unit through a preamplification circuit, and the analog-to-digital conversion unit is connected to a microcontroller;

[0015] The impedance spectrum signal acquisition subsystem includes: a DDS signal generator. The DDS signal generator transmits the emitted voltage signal to a first electrode plate and a second electrode plate, the first electrode plate and the second electrode plate transmit the voltage signal to the finger tip of the object to be detected, and through a third electrode plate and a fourth electrode plate, the voltage signal feedback from the finger tip of the object to be detected is collected. The feedback voltage signal is amplified by a differential amplification circuit, and the amplified voltage signal is converted into a DC voltage signal through a synchronous demodulation circuit. Finally, the DC voltage signal is converted from an analog signal to a digital signal through an analog-to-digital conversion circuit, and the digital signal is transmitted to the microcontroller;

[0016] The microcontroller transmits the acquired optoelectronic signals and impedance spectrum signals to a host computer, and the host computer processes the two signals to obtain a blood glucose prediction value.

[0017] The above technical solution has the following advantages or beneficial effects:

[0018] (1) The present invention combines near-infrared spectroscopy and impedance spectrum technology to obtain more comprehensive physiological information. Through comprehensive analysis of two different signal sources, it can calculate and predict blood glucose levels more accurately. Near-infrared spectroscopy is mainly used to obtain the optical properties in blood and extract relevant blood glucose physiological parameters, while impedance spectrum technology enhances the perception ability of blood glucose fluctuations by measuring the impedance of human tissues. The two work together to achieve non-invasive blood glucose detection with higher accuracy. Compared with traditional blood glucose measurement technologies such as finger pricking or venous blood sampling, the present invention can not only significantly reduce the physical pain and psychological pressure of patients, but also simplify the operation process, reduce the infection risk and detection cost, and has a broader clinical application prospect.

[0019] (2) In the present invention, multiple visible light and near-infrared LEDs with different wavelengths are used as light sources, and multiple PDs capable of responding to different wavelengths are used to collect the reflected light from the fingertip. Compared with the existing non-invasive blood glucose measurement techniques using a single-wavelength near-infrared light, the present invention has successfully overcome the problem of the narrow detectable range of the blood glucose absorption spectrum in the past, greatly improving the accuracy of blood glucose value prediction, making the blood glucose detection results more reliable and accurate, providing a more effective solution for the field of non-invasive blood glucose detection, and is expected to better serve the medical detection needs in practical applications and promote the further development of blood glucose detection technology.

[0020] (3) The human impedance measurement device of the present invention can use an operational amplifier to amplify the detected current, which can amplify the effective signal and reduce the influence of noise. The operational amplifier has a high input impedance and a low output impedance, which can overcome the influence caused by the contact impedance between the human body and the electrode sheet, making the measurement result more accurate.

[0021] (4) The present invention innovatively introduces artificial intelligence deep learning technology to perform real-time processing operations on the collected optoelectronic signals and impedance spectrum signals, skillfully integrating the current cutting-edge deep learning technology with traditional blood glucose detection technology, and thus achieving the characteristics of real-time, intelligent and simple operation of blood glucose monitoring work, opening up a new development direction for the field of blood glucose monitoring, greatly improving the monitoring efficiency and effect, and is expected to be widely applied to various medical and health scenarios, providing users with a more high-quality and convenient blood glucose monitoring service experience to meet the high requirements and new expectations of modern society for health management.

[0022] (5) The non-invasive blood glucose monitoring system constructed by the present invention has gone through a rigorous experimental process of multiple cycles, strongly confirming that the system can achieve the function of synchronous blood glucose detection, and the obtained detection results show a high level of accuracy. Its blood glucose detection accuracy fully complies with the standard specifications formulated by the International Organization for Standardization (ISO), and has high reliability and practicality, laying a solid foundation for the clinical application and popularization of non-invasive blood glucose detection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0024] Figure 1 It is a module diagram of a non-invasive blood glucose monitoring system for optoelectronic signals, impedance spectrum signals and deep learning according to the present invention;

[0025] Figure 2 Flow chart of a non-invasive blood glucose monitoring system based on optoelectronic signals, impedance spectroscopy and deep learning according to an embodiment of the present invention;

[0026] FIG. 3(a) is a schematic internal structure diagram of a deep learning model for PPG signal feature extraction according to an embodiment of the present invention.

[0027] FIG. 3(b) is a schematic internal structure diagram of a CNN block according to an embodiment of the present invention.

[0028] FIG. 3(c) is a schematic internal structure diagram of a Dense Block in a deep learning model for PPG signal feature extraction according to an embodiment of the present invention.

[0029] FIG. 3(d) is a schematic internal structure diagram of a Transition Layer in a deep learning model for PPG signal feature extraction according to an embodiment of the present invention.

[0030] Figure 4 Schematic internal structure diagram of a multi-feature fusion blood glucose prediction deep learning model according to an embodiment of the present invention. Detailed implementation manners

[0031] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which the present invention belongs.

[0032] Embodiment 1

[0033] This embodiment provides a non-invasive blood glucose detection system based on optoelectronic signals and impedance spectrum signals;

[0034] A non-invasive blood glucose detection system based on optoelectronic signals and impedance spectrum signals includes:

[0035] An acquisition module, which is configured to: acquire optoelectronic signals and impedance spectrum signals at the finger tip of an object to be detected;

[0036] A feature extraction module, which is configured to: respectively extract features of the optoelectronic signals and features of the impedance spectrum signals;

[0037] A feature fusion module, which is configured to: perform feature fusion on the optoelectronic signal features and the impedance spectrum signal features, and then predict the fused features to obtain a blood glucose prediction result.

[0038] Furthermore, the acquisition of the optoelectronic signals and impedance spectrum signals at the finger tip of the object to be detected means:

[0039] Visible light and infrared light are emitted to the skin at the fingertip of the object to be detected, and the optoelectronic signals feedback from the skin at the fingertip are collected by a photodiode.

[0040] A voltage signal is emitted to the skin at the fingertip of the object to be detected by a DDS signal generator, and the impedance spectrum signal feedback from the skin at the fingertip is collected by an electrode plate.

[0041] Further, the features of the optoelectronic signal and the impedance spectrum signal are extracted. Among them, the extraction process of the features of the optoelectronic signal includes:

[0042] (1-1): The optoelectronic signal is decomposed into five layers by using a wavelet basis function to obtain wavelet coefficients with information in different frequency ranges. By analyzing the amplitude change of the high-frequency coefficients at a set scale, the first feature of the optoelectronic signal is constructed.

[0043] (1-2): The optoelectronic signal is input into a feature learning model to obtain the second feature of the optoelectronic signal.

[0044] Further, the step (1-1): The optoelectronic signal is decomposed into five layers by using a wavelet basis function to obtain wavelet coefficients with information in different frequency ranges. By analyzing the amplitude change of the high-frequency coefficients at a set scale, the first feature of the optoelectronic signal is constructed, which specifically includes:

[0045] (1-1-1) Wavelet transform formula:

[0046]

[0047] Among them, X(n) is the input optoelectronic signal (discrete sequence, n = 0, 1,..., N-1, N is the signal length), is the wavelet basis function, j represents the scale parameter, k represents the translation parameter, and W f(j,k) is the wavelet coefficient. The selection of the wavelet basis function : Daubechies wavelet. Scale j: Here, five-layer decomposition is performed, j = 1, 2, 3, 4, 5. The value range of k is determined according to the signal length and scale. In discrete wavelet transform, k is used to translate the wavelet basis function on the time axis to match the signal.

[0048] (1-1-2) Assume that the input optoelectronic signal is X(n). The following is the specific five-layer decomposition process:

[0049] First - layer decomposition: Decompose the signal X(n) using the selected wavelet basis function to obtain the low - frequency coefficient a1(n) and the high - frequency coefficient d1(n). The low - frequency coefficient a1(n) contains the low - frequency part information of the signal, and the high - frequency coefficient d1(n) contains the high - frequency part information of the signal. This process can be achieved through a filter bank, that is, the low - frequency coefficient is obtained through a low - pass filter, and the high - frequency coefficient is obtained through a high - pass filter.

[0050]

[0051] Among them, h0(m - 2n) is the low - pass filter coefficient, and h1(m - 2n) is the high - pass filter coefficient, which are related to the selected wavelet basis function.

[0052] Second - layer decomposition: Decompose the low - frequency coefficient a1(n) obtained from the first - layer decomposition again to obtain the next - layer low - frequency coefficient a2(n) and high - frequency coefficient d2(n).

[0053]

[0054]

[0055] Third - layer decomposition: Decompose the low - frequency coefficient a2(n) obtained from the second - layer decomposition to obtain a3(n) and d3(n).

[0056]

[0057] Fourth - layer decomposition: Decompose the low - frequency coefficient a3(n) obtained from the third - layer decomposition to obtain a4(n) and d4(n).

[0058]

[0059] Fifth - layer decomposition: Decompose the low - frequency coefficient a4(n) obtained from the fourth - layer decomposition to obtain a5(n) and d5(n).

[0060]

[0061] After five - layer decomposition, the low - frequency coefficients a1(n), a2(n), a3(n), a4(n), a5(n) and high - frequency coefficients d1(n), d2(n), d3(n), d4(n), d5(n) at different scales are obtained, which respectively correspond to the information in different frequency ranges.

[0062] (1 - 1 - 3) Steps for analyzing the amplitude change of the high - frequency coefficient at the set scale:

[0063] Select the set scale: Assume that the high - frequency coefficient d s (n) at the s - th layer (s ∈ {1, 2, 3, 4, 5}) is selected for analysis.

[0064] Calculate the amplitude: Calculate the amplitude of the high-frequency coefficient d s (n), where the absolute value is used to represent the amplitude, that is

[0065] |d s (n)| = abs(d s (n))

[0066] Analyze the amplitude change, including: calculating statistical features, analyzing the change trend, and analyzing extreme points;

[0067] Calculation of statistical features: Calculate the statistical features of the amplitude sequence |d s (n)|, such as the mean μ s 、variance σ 2 etc. L is the length of d s (n).

[0068]

[0069] Analysis of the change trend: Observe the change trend of the amplitude through a sliding window or polynomial fitting.

[0070] Analysis of extreme points: Find the maximum and minimum points in the amplitude sequence |d s (n)| and analyze their distribution and changes.

[0071] Furthermore, the step (1-2): Input the optoelectronic signal into the feature learning model to obtain the second feature of the optoelectronic signal. Among them, as shown in Fig. 3(a), the feature learning model includes:

[0072] The CNN module, the first deep learning module, the second deep learning module, the global average pooling layer, the first fully connected layer, and the output layer connected in sequence;

[0073] As shown in Fig. 3(b), the CNN module includes: the first convolutional layer, the first batch normalization layer, the first activation function layer, the first max pooling layer, the second convolutional layer, the second batch normalization layer, the second activation function layer, the second max pooling layer, the third convolutional layer, the third batch normalization layer, and the third activation function layer connected in sequence; Among them, the input end of the first convolutional layer is also connected to the input end of the second convolutional layer; the input end of the second convolutional layer is also connected to the input end of the third convolutional layer.

[0074] Furthermore, the internal structures of the first deep learning module and the second deep learning module are the same. The first deep learning module includes:

[0075] The first dense block Dense Block and the first transition layer Transition Layer connected in sequence;

[0076] The input end of the first Dense Block is the input end of the first deep learning module, and the output end of the first Transition Layer is the output end of the first deep learning module.

[0077] Furthermore, the second deep learning module includes:

[0078] A second Dense Block and a second Transition Layer connected in sequence;

[0079] The input end of the second Dense Block is the input end of the second deep learning module, and the output end of the second Transition Layer is the output end of the second deep learning module.

[0080] The input end of the CNN module is also connected to the input end of the first Dense Block;

[0081] The output end of the CNN module is also respectively connected to the input end of the first Transition Layer and the input end of the second Dense Block.

[0082] Furthermore, the working process of the first deep learning module includes:

[0083] The first Dense Block is used to extract information features and achieve feature reuse using residual connections to extract features;

[0084] The first Transition Layer is used to reduce the number of feature channels.

[0085] Furthermore, as shown in Fig. 3(c), the first Dense Block includes:

[0086] A fourth convolutional layer, a fourth batch normalization layer, a fourth activation function layer, a first adder, and a fifth convolutional layer connected in sequence; wherein, the input end of the first adder is also connected to the input end of the fourth convolutional layer.

[0087] The fourth convolutional layer is used to perform further convolutional operations on the input features to extract richer local features.

[0088] The fourth batch normalization layer is used to normalize the output of the fourth convolutional layer to accelerate the training process and improve the stability of the model.

[0089] The fourth activation function layer is used to introduce non-linearity to enhance the expressive power of the model.

[0090] The first adder is used to implement the residual connection, adding the input features of the fourth convolutional layer to the features processed by convolution, normalization, and activation function. The residual connection can alleviate the vanishing gradient problem, making the model easier to train, and at the same time can also promote the reuse of features.

[0091] The fifth convolutional layer is used to perform another convolution on the features after the residual connection to further extract and integrate features.

[0092] Further, as shown in Fig. 3(d), the first transition layer, including:

[0093] The fifth batch normalization layer, the sixth convolutional layer, the average pooling layer, and the output layer connected in sequence.

[0094] The fifth batch normalization layer is used to normalize the mean and variance of the feature data to 0 and 1 respectively by the third batch normalization layer, which can make the model training more stable, accelerate the convergence speed, improve the training efficiency, provide a standardized input for the subsequent convolution operation, and improve the training efficiency.

[0095] The sixth convolutional layer is used to reduce the dimension or adjust the number of channels of the features through 2×2 convolution. Reducing the number of feature channels can reduce the computational amount and avoid overfitting.

[0096] The average pooling layer is used to reduce the size of the feature map, reduce the data volume and computational amount, relieve the model burden, and improve the operation speed. At the same time, it enhances the translational invariance of the features and improves the model robustness.

[0097] The output layer is used to output the features processed by the previous layers as the final result of the transition layer to the subsequent network layer. It plays a role in connecting the transition layer and the subsequent network, enabling the features processed by normalization, convolution, and pooling to smoothly enter the next network module for further analysis and processing, and ensuring the smooth progress of the data flow and processing process of the entire model.

[0098] Further, extracting the features of the optoelectronic signal and the features of the impedance spectrum signal, wherein the feature extraction process of the impedance spectrum signal includes:

[0099] Let the impedance value measured at different frequencies ω be Z m (ω);

[0100] According to the Cole-Cole model:

[0101]

[0102] where Z0 represents the low-frequency limit impedance, α represents the polarization degree parameter (0 ≤ α ≤ 1), β represents the dispersion index (0 ≤ β ≤ 1), and τ represents the time constant;

[0103] to minimize the error function

[0104]

[0105] as the goal, iteratively solve for the optimal impedance value Z0 of biological tissue under direct current conditions, the frequency dispersion degree α of biological tissue impedance, and the bending degree β of the impedance spectrum curve.

[0106] Z(ω) represents the theoretical impedance value of biological tissue at angular frequency ω, which is the result calculated by the model based on given parameters and is used for comparative analysis with actual measured values.

[0107] Z m (ω) represents the impedance values measured at different angular frequencies ω, which are the impedance response values of human tissue measured by an actual impedance spectrum signal acquisition device after applying current signals of different frequencies to human tissue.

[0108] Z0 represents the impedance value of biological tissue under direct current (ω = 0, that is, when the frequency approaches zero) conditions. It reflects the resistance characteristics of the tissue at rest and is a basic parameter in the model, which is related to the ionic conduction characteristics inside the tissue, etc.

[0109] α is a dimensionless parameter used to describe the frequency dispersion degree of biological tissue impedance, which reflects the influence degree of different components (including changes in blood glucose concentration) or structures in the tissue on the impedance change with frequency. The value of α ranges between 0 and 1. The closer α is to 0, the less obvious the impedance change with frequency; the closer α is to 1, the more significant the impedance change with frequency.

[0110] β is also a dimensionless parameter used to characterize the bending degree of the impedance spectrum curve in the Cole-Cole model. Its value range is usually between 0 and 1, which reflects the non-ideal capacitive characteristics of biological tissue and the complexity of tissue structure. For different tissues, the value of β is different. For example, there are differences in the β values between healthy tissues and diseased tissues.

[0111] j is the imaginary unit, satisfying j2 = -1, and is used to represent the reactance part of impedance (the part related to capacitance and inductance) in complex number operations.

[0112] τ is the time constant, which is related to the physical processes inside biological tissue and reflects the response speed of the tissue to the applied current. The larger the time constant, the slower the tissue responds to current changes, and the impedance change with frequency will also show different characteristics accordingly.

[0113] Further, the feature fusion module is configured to: perform feature fusion on the optoelectronic signal features and impedance spectrum signal features, and then predict the fused features to obtain a blood glucose prediction result; the feature fusion is implemented using a trained feature fusion model.

[0114] Further, the training process of the trained feature fusion model includes:

[0115] Construct a training set, where the training set is the optoelectronic signal features and impedance spectrum signal features with known blood glucose concentration values;

[0116] Input the training set into the feature fusion model to train the feature fusion model. When the loss function value of the feature fusion model no longer decreases, stop training to obtain the trained feature fusion model.

[0117] Further, as Figure 4 shown, the trained feature fusion model includes:

[0118] A seventh convolutional layer, a sixth batch normalization layer, a first BiLSTM model, a seventh batch normalization layer, a second BiLSTM model, a self-attention mechanism layer, a first multi-layer perceptron MLP, a Dropout layer, a second multi-layer perceptron MLP, a second fully connected layer, and a prediction layer connected in sequence.

[0119] The prediction layer includes: a linear activation function layer for directly outputting continuous values, meeting the requirements of the blood glucose concentration regression task. Regularization: Constraining the model complexity through L2 regularization to prevent overfitting. Loss function: MSE is more sensitive to outliers and is suitable for blood glucose prediction requiring high precision. Optimizer: The Adam optimizer combines an adaptive learning rate to balance the training speed and convergence.

[0120] Further, the working process of the feature fusion model includes:

[0121] The first BiLSTM (Bidirectional Long Short-Term Memory Network) model and the second BiLSTM model: For the input optoelectronic signal feature and impedance spectrum signal feature sequences, the first BiLSTM model and the second BiLSTM model respectively extract key information from the features at different time steps and mine the dynamic change laws of the two signals in the time dimension. Since the change in blood glucose concentration will cause continuous changes in the optoelectronic signal and impedance spectrum signal within a period of time, BiLSTM can effectively learn the correlations between these change trends and provide richer time series information for subsequent feature fusion and prediction.

[0122] Self-attention mechanism layer: The self-attention mechanism calculates the importance of each feature based on the correlation between features, highlights the key features closely associated with blood glucose concentration, and suppresses irrelevant or less contributing features. It can dynamically adjust the weights of these features to improve the accuracy of model prediction.

[0123] Second fully connected layer: Further linearly combines and integrates the features processed by the self-attention mechanism. Fuses features of different dimensions, maps the features to a new feature space, and prepares for the prediction layer. By learning weights, the second fully connected layer can extract more representative comprehensive features, enabling the model to more accurately capture the complex relationship between features and blood glucose values.

[0124] Furthermore, the working process of the feature fusion model includes:

[0125] Assume the optoelectronic signal feature is The impedance spectrum signal feature is The final fusion result is The expression of the adaptive weighting algorithm is:

[0126]

[0127] Among them, the weights w1 and w2 satisfy w1 + w2 = 1, and are dynamically adjusted through the following adaptive strategy:

[0128]

[0129] Among them, r PPG represents the Pearson correlation coefficient between the predicted value of the optoelectronic signal and the actual blood glucose value; reflecting the correlation between the optoelectronic signal and blood glucose; r Z represents the Pearson correlation coefficient between the predicted value of the impedance spectrum signal and the actual blood glucose value; reflecting the correlation between the impedance spectrum signal and blood glucose; is the variance of the predicted value of the optoelectronic signal; is the variance of the predicted value of the impedance spectrum signal.

[0130] The beneficial effects of the above technical solutions are: There are limitations in using a single optoelectronic signal or impedance spectrum signal for blood glucose detection. The optoelectronic signal is easily affected by other components in the skin (such as moisture, fat, hemoglobin, etc.) and the temperature and humidity of the test environment; although the impedance spectrum signal can reflect the electrical properties of human tissues, it is also interfered by individual differences (such as differences in tissue conductivity among different people). Therefore, in order to more accurately detect blood glucose, it is necessary to combine the advantages of the two signals.

[0131] The beneficial effects of the above technical solution are as follows: improving detection accuracy: the optoelectronic signal and the impedance spectrum signal reflect human physiological information from different angles. After fusion, they can provide more comprehensive data. The optoelectronic signal can reflect the optical characteristics of blood components, while the impedance spectrum signal can reflect the electrical characteristics of tissues. The combination of the two can more accurately correlate with blood glucose concentration, reduce errors caused by a single signal, and improve the accuracy of detection.

[0132] The beneficial effects of the above technical solution are as follows: enhancing anti-interference ability: the two signals complement each other. When one signal is interfered with, the other signal can provide effective information support. For example, when ambient light interference causes the optoelectronic signal to be inaccurate, the impedance spectrum signal can still provide reliable information, enabling the model to perform blood glucose prediction more stably and enhancing the anti-interference ability of the entire detection system.

[0133] The beneficial effects of the above technical solution are as follows: adapting to individual differences: there are differences in the skin composition and tissue conductivity of different individuals. Single-signal detection is difficult to adapt to these differences. Feature fusion can comprehensively consider various physiological information, better adapt to the characteristics of different individuals, improve the generality and reliability of detection, and make the detection results more applicable to different populations.

[0134] Example Two

[0135] This example provides a non-invasive blood glucose detection device based on optoelectronic signals and impedance spectrum signals;

[0136] As Figure 1 and Figure 2 shown, the non-invasive blood glucose detection device based on optoelectronic signals and impedance spectrum signals includes:

[0137] An optoelectronic signal acquisition subsystem and an impedance spectrum signal acquisition subsystem;

[0138] The optoelectronic signal acquisition subsystem includes: a visible light LED and a near-infrared LED. The visible light LED and the near-infrared LED emit the emitted light to a lens. The lens transmits the light to the first end of an emission optical fiber. The emission optical fiber transmits the light to the second end of the emission optical fiber. The second end of the emission optical fiber divides the light into multiple beams and emits them to the skin at the finger tip of the object to be detected. The first end of a receiving optical fiber receives each beam of reflected light feedback from the finger tip skin. The second end of the receiving optical fiber is connected to the first end of a photodiode. The second end of the photodiode is connected to an analog-to-digital conversion unit through a preamplification circuit. The analog-to-digital conversion unit is connected to a microcontroller;

[0139] The impedance spectroscopy signal acquisition subsystem includes: a DDS signal generator that sends the emitted voltage signal to the first and second electrode pads. The first and second electrode pads transmit the voltage signal to the finger tip of the object to be detected. Through the third and fourth electrode pads, the voltage signal feedback from the finger tip of the object to be detected is collected. The feedback voltage signal is amplified by a differential amplifier circuit, and the amplified voltage signal is converted into a DC voltage signal by a synchronous demodulation circuit. Finally, the DC voltage signal is converted from an analog signal to a digital signal by an analog-to-digital conversion circuit, and the digital signal is transmitted to a microcontroller.

[0140] The microcontroller transmits the collected optoelectronic signal and impedance spectroscopy signal to a host computer, and the host computer processes the two signals to obtain a blood glucose prediction value.

[0141] Optical signal: Light-emitting diodes with a wavelength band ranging from 465 nm to 1600 nm are selected. These diodes emit optical signals, and the optical signals are focused onto one end of a composite quartz fiber bundle by a high-performance optical lens, and then are emitted from the other end of the composite quartz fiber bundle to the finger tip of the person to be monitored for blood glucose. The finger tip reflected light received by the composite quartz fiber is divided into multiple beams, and each beam of reflected light is respectively transmitted to the corresponding photodiode.

[0142] Reasons for selecting the wavelength band from 465 nm to 1600 nm:

[0143] 1. Absorption spectral characteristics of glucose

[0144] Absorption peaks in the near-infrared region: Glucose molecules have multiple characteristic absorption peaks in the near-infrared spectrum (700 - 2500 nm) range. For example:

[0145] 900 - 1100 nm: Related to the vibrations of C-H and O-H bonds in glucose molecules.

[0146] 1200 - 1400 nm: Affected by the combined vibrations of hydroxyl groups (-OH) in the glucose molecular structure.

[0147] Covering the key absorption band: The wavelength range from 465 nm to 1600 nm contains the main absorption region of glucose, especially the near-infrared part, which can effectively detect changes in blood glucose concentration.

[0148] 2. Penetration depth and tissue adaptability of optical signals

[0149] Deep tissue penetration: Near-infrared light (especially wavelengths above 1000 nm) can penetrate the human tissue by several millimeters, reaching the dermis layer rich in blood vessels and reducing the interference of the skin surface layer (such as the stratum corneum and adipose layer) on optical signals.

[0150] Multi - wavelength complementarity:

[0151] 465nm visible light: It may be used to detect the optical properties of the skin surface (such as hemoglobin concentration) and assist in correcting individual differences in near - infrared signals (such as the influence of skin color).

[0152] Long - wavelength near - infrared light (such as 1600nm): It penetrates deeper and reflects the blood glucose information of deep tissues.

[0153] It should be understood that for the impedance spectrum signal: The impedance measurement electrode adopts a flexible micro - nano structure electrode and selects a material with good biocompatibility, such as a silver nanowire / polydimethylsiloxane (PDMS) composite material. It can not only ensure close fitting with the skin, reduce contact impedance, but also adapt to the deformation of the skin, improving the stability and comfort of measurement. The electrode array is designed for multi - band response and can accurately measure the impedance changes of biological tissues in a wide frequency range (1kHz - 1MHz), meeting the acquisition requirements of tissue information at different depths and providing a rich data basis for subsequent blood glucose concentration calculation.

[0154] Signal conditioning circuit: It is mainly responsible for pre - processing operations such as amplifying, filtering, and analog - to - digital conversion of the original signals collected by the sensor. However, for the impedance measurement part, a high - precision constant - current source is required to provide a stable alternating - current excitation current for the electrode to ensure that the current amplitude is constant at different frequencies. Therefore, the voltage signal of a DDS signal generator is used as the input of a high - precision voltage - to - current conversion circuit.

[0155] For the PPG signal, first, the LED driver circuit enables LEDs of different wavelengths; then, the pre - amplifier circuit uses a low - noise, high - gain amplifier (instrumentation amplifier ADA4530) to convert the weak photocurrent signal into a voltage signal and perform preliminary amplification to improve the signal - to - noise ratio of the signal; subsequently, through a Butterworth band - pass filter circuit, ambient light, power - frequency interference, and high - frequency noise and other clutter are filtered out, and only the effective frequency components of the PPG signal (usually 0.5Hz - 5Hz, corresponding to the heart rate range) are retained.

[0156] For the impedance spectrum signal, the voltage signal collected by the measurement voltage electrode is amplified by a differential amplifier circuit to enhance the weak voltage change; then, a synchronous demodulation circuit is used to convert the alternating - current voltage signal into a direct - current signal. Finally, the analog signal is converted into a digital signal by an analog - to - digital converter (ADC) for subsequent processing by the microprocessor.

[0157] Adapted to the above - mentioned optical signal and impedance spectrum signal analysis software architecture, specifically:

[0158] The system software adopts a hierarchical architecture design, including a bottom - layer driver layer, an intermediate data - processing layer, and an upper - layer application layer. Each layer cooperates with each other to ensure the efficient and stable operation of the system.

[0159] The underlying driver layer is mainly responsible for direct interaction with hardware devices, using the microprocessor to control the operation of the entire hardware part, such as LED lighting, ADC acquisition, sampling frequency, and serial communication. For optoelectronic signals, a driver program is written to precisely control the light source by enabling the LED driver chip through PWM output; for impedance measurement electrodes, a constant current source is driven to output an excitation current with a specific frequency sequence, controlling the voltage measurement and data acquisition process; and an analog-to-digital conversion chip is driven to collect the converted optoelectronic signals and impedance spectrum signals. At the same time, it also covers functional modules such as the clock configuration of the microprocessor, interrupt management, storage access drive, and communication interface (such as SPI, UART, etc.) drive, providing a unified hardware operation interface for the upper-layer software, shielding the underlying hardware details, and improving the portability and maintainability of the software.

[0160] The intermediate data processing layer focuses on the in-depth processing and analysis of the collected raw data. This layer includes core functional modules such as signal preprocessing, feature extraction, model calculation, and dual-signal fusion. The signal preprocessing module inherits the digital signals transmitted from the underlying driver, and uses a baseline correction algorithm to eliminate the baseline drift of the signal, ensuring the stability and accuracy of the signal;

[0161] The feature extraction module extracts feature information from multiple dimensions for the PPG signal, respectively from the time domain (extracting parameters such as peak value, trough value, rise time, pulse width, etc.), frequency domain (calculating spectral features through FFT to obtain heart rate-related frequency components and harmonic energy distribution), and time-frequency domain (analyzing the time-varying features of the signal using wavelet transform). At the same time, for the impedance spectrum signal, based on data such as impedance values and phase angles at different frequencies, combined with the biological tissue equivalent circuit model (Cole-Cole model), key parameters reflecting the changes in tissue electrical properties are calculated by fitting;

[0162] Among them, time-domain feature: Peak: The maximum value of the PPG signal within one cycle. It reflects the maximum filling degree of the arterial blood vessel during heart contraction and is related to the pumping ability of the heart. In actual detection, the peak value is determined by searching for the maximum value point in the PPG signal waveform. Its expression is:

[0163] P peak =max(PPG(t)), t∈T

[0164] Where PPG(t) represents the PPG signal value at time t, and T represents the time range of a complete PPG signal cycle.

[0165] Trough: The minimum value of the PPG signal within one cycle, representing the state of the arterial blood vessel during heart relaxation, reflecting information such as blood vessel elasticity and peripheral resistance. The expression is:

[0166] T trough= min(PPG(t)), t ∈ T

[0167] Rise Time: The time required for the PPG signal to rise from the trough to the peak, which reflects the speed of cardiac contraction and the compliance of blood vessels. The calculation formula is:

[0168] T r i se = t peak - t troug h

[0169] where tpeak is the time point corresponding to the peak and ttrough is the time point corresponding to the trough.

[0170] Pulse Width: In the PPG signal, the time interval between a certain threshold point on the rising edge and the same threshold point on the falling edge, which usually reflects the duration of each cardiac beat and is related to cardiac function and vascular status. Assuming the threshold is th, the pulse width expression is:

[0171]

[0172] and are the start and end times at half of the peak value respectively.

[0173] The model calculation module relies on a pre-trained deep learning model, takes the extracted PPG signal features and impedance spectrum features as inputs, and preliminarily estimates the blood glucose concentration value; the dual-signal fusion module is one of the key innovations of this system software. It uses an adaptive weighted algorithm to dynamically adjust the weights according to the correlation between the two signals and blood glucose concentration under different physiological states and individual differences, fuses the output of the PPG signal model and the output of the impedance spectrum signal model, and obtains the final accurate blood glucose concentration prediction result, effectively making up for the limitations of single-signal detection and improving the overall detection accuracy of the system.

[0174] Specific implementation method:

[0175] 1. Online update:

[0176] Real-time calculate r PPG 、r Z 、 and in the current time window and dynamically adjust the weights. The length of the time window is set to 5 minutes to balance real-time performance and statistical stability.

[0177] 2. Initialization and smoothing:

[0178] The initial weights are (w1 = w2 = 0.5) and will be gradually adjusted after data accumulation. An exponential smoothing factor (α = 0.9) is introduced to avoid drastic fluctuations in the weights.

[0179] The upper application layer is oriented towards users and medical professionals, providing a convenient interaction interface and rich functional applications. The user interface is designed to be simple and intuitive, presenting real-time blood glucose values, blood glucose trend graphs (blood glucose fluctuation curves over the past few hours and days), health reminder information (such as diet advice, exercise reminders), and device connection status through a display screen; users can perform basic operations such as device startup, measurement mode switching, and personal information entry through buttons or touch operations, and can also view historical measurement data reports to understand the long-term change patterns of their own blood glucose. For medical professionals, professional diagnosis interfaces are reserved, and through the supporting computer software or mobile applications, they can deeply analyze the detailed measurement data of patients, combine clinical experience for more accurate disease assessment and treatment plan adjustment, and can remotely monitor the blood glucose status of patients, achieving seamless docking of personalized medical services and greatly improving the efficiency and quality of diabetes management.

[0180] In the key feature extraction stage of data processing, the present invention applies deep learning algorithms to quickly, efficiently, and intelligently process and analyze the rich physiological signals collected, deeply excavating the key features and patterns hidden behind the data, and successfully achieving real-time non-invasive and accurate monitoring of blood glucose concentration.

[0181] In terms of near-infrared light detection, in order to collect the near-infrared spectral information reflected by the fingertip comprehensively and efficiently, the system uses multiple high-performance photoelectric detection elements. Through optimized layout and collaborative working mechanisms, it ensures that more subtle and key spectral features can be accurately captured, providing a rich and accurate data basis for subsequent analysis.

[0182] At the same time, impedance spectroscopy technology is introduced. By applying currents with varying frequencies to human tissues and deeply analyzing the response characteristics of the tissues to these currents, key electrical property data at the cellular level can be accurately obtained, including the capacitance characteristics of cell membranes and the electrical conductivity inside tissues. There is a close and inherent correlation between these electrical parameters and blood glucose concentration, which can provide strong supplementary information for the analysis of near-infrared spectral data. The two complement each other, greatly improving the accuracy and stability of blood glucose concentration measurement.

[0183] In this embodiment, we propose an innovative non-invasive blood glucose detection solution that combines combined near-infrared light, human impedance spectroscopy technology, and deep learning algorithms, aiming to provide a more accurate and convenient solution for blood glucose monitoring.

[0184] Design of PPG Signal Acquisition Circuit: This module uses visible light and near-infrared light (435nm - 1600nm) LEDs as light sources. Matched with it is a highly responsive silicon-based photodiode as the detector, which has extremely low dark current and fast response time, and can accurately capture weak light signal changes. In the driving circuit part, to ensure stable light emission of the light source and controllable light intensity, a constant current driving chip is used, effectively avoiding the problem of unstable light intensity caused by fluctuations in the light source current. By connecting the PWM output pin of the microprocessor to the control end of the constant current driving chip, digital adjustment of the light intensity of the light source can be achieved, meeting the light intensity requirements under different skin tones and different measurement environments. The preamplifier circuit selects a low-noise and high-gain instrumentation amplifier (AD7124), which can convert weak photocurrent signals into voltage signals and amplify them preliminarily. The amplification factor can be flexibly set by external resistors. In this design, according to the requirements for signal amplitude, it is set to 100 times, converting millivolt-level photocurrent signals into volt-level voltage signals, improving the signal-to-noise ratio and laying a foundation for subsequent signal processing. The filtering circuit part uses a second-order Butterworth band-pass filter, and its passband frequency range is designed to be 0.5Hz - 5Hz, accurately corresponding to the human heart rate range (30 - 300 beats per minute). Compared with other types of filters, the Butterworth filter has the advantage of high signal flatness in the passband, which can minimize signal distortion to the greatest extent and ensure the waveform integrity of the PPG signal. Finally, it is converted into a digital signal by a high-precision ADC.

[0185] Design of Impedance Spectroscopy Signal Acquisition Circuit: The impedance measurement electrode uses a flexible micro-nano structured electrode, and a material with good biocompatibility, silver nanowire / polydimethylsiloxane (PDMS) composite material, is selected. It can not only ensure close contact with the skin, reduce contact impedance, but also adapt to skin deformation, improving the stability and comfort of measurement. The electrode array is designed as a 4×4 square array, and the adjacent electrode spacing is 3mm. It can simultaneously collect impedance information at multiple positions, covering a wider skin area and providing a rich data source for subsequent data fusion. Driving circuit part: The excitation source selects a high-precision direct digital frequency synthesis (DDS) chip. By controlling the frequency control word of the DDS chip through the microprocessor, the excitation current can be output according to the preset frequency sequence, meeting the requirements for collecting information at different depths of biological tissues. The measurement circuit uses the four-electrode method. A pair of current electrodes is connected to the output end of the DDS chip to inject an alternating current with a known frequency and amplitude. The current amplitude is constantly 1mA to ensure that the biological tissue is in the linear response region during the measurement process. Another pair of voltage electrodes collects the voltage drop across the tissue through a high-input impedance differential amplifier. Signal conditioning circuit: The voltage signal after differential amplification is converted into a DC signal through a synchronous demodulation circuit. Subsequently, the processed impedance signal is digitized by a 16-bit high-precision ADC.

[0186] Optimization of PPG Signal Feature Extraction Algorithm: This patent introduces the wavelet transform algorithm for time-frequency domain feature extraction of PPG signals. Compared with traditional methods, wavelet transform can adaptively select appropriate time-frequency resolution according to the signal frequency components and perform multi-scale decomposition on signals in different frequency bands. During the rising stage of blood glucose concentration, the change in hemorheological properties prompts the adjustment of cardiac pumping pressure, and the high-frequency components of the PPG signal increase, reflecting the acceleration of blood circulation; wavelet transform can focus on high-frequency details and accurately capture this change. Using the db4 wavelet basis function, the PPG signal is decomposed into 5 layers to obtain wavelet coefficients containing information in different frequency ranges. By analyzing the amplitude changes of high-frequency coefficients at a specific scale, a feature vector closely related to blood glucose concentration is constructed.

[0187] Furthermore, combined with the deep learning algorithm, a PPG signal feature learning model based on convolutional neural network (CNN) is constructed. The convolutional layer of CNN can automatically extract local features in the PPG signal, such as the change patterns of the rising and falling edges of the waveform, the subtle differences between different cycles, etc. The pooling layer reduces the data dimension and computational amount, and the fully connected layer integrates global features and outputs a feature representation highly related to blood glucose concentration. A CNN architecture with 3 convolutional layers, 2 pooling layers, and 2 fully connected layers is adopted, and the model is trained with a large number of PPG signal samples of different individuals and different physiological states.

[0188] Selection and Calculation of Impedance Spectrum Feature Parameters: This patent selects the impedance amplitude in the low-frequency band (3 kHz - 5 kHz) and the phase angle in the high-frequency band (800 kHz - 1 MHz) as key feature parameters, and uses the Cole-Cole model to fit and calculate the parameters α and β reflecting tissue electrical properties. In actual measurement, after collecting multi-frequency impedance spectrum data using the four-electrode method, the Cole-Cole model is fitted by the nonlinear least squares method. To adapt to the changes in factors such as different individual skin characteristics and measurement environments, an adaptive adjustment strategy is introduced. The open-circuit voltage, short-circuit current, and other parameters of the contact between the measurement electrode and the skin are monitored in real time to evaluate the contact quality.

[0189] Details of the Adaptive Adjustment Strategy:

[0190] Real-time Parameter Monitoring: Continuously monitor the open-circuit voltage and short-circuit current parameters of the contact between the measurement electrode and the skin, and at the same time record environmental parameters such as the environmental temperature and humidity during measurement and the basic information of the individual (such as age, gender, etc.).

[0191] Establishment of Database: Correlate and store the above monitoring data, the corresponding impedance spectrum measurement results, and the final feature parameter calculation results to construct a database containing measurement data of different individuals and different environmental conditions.

[0192] Data analysis and model construction: Using machine learning or data analysis methods, analyze the data in the database to find the relationships between factors such as skin characteristics, measurement environment, etc. and the calculation results of characteristic parameters as well as measurement accuracy, and construct corresponding prediction models.

[0193] Dynamic adjustment: During a new measurement process, based on the parameters such as open-circuit voltage and short-circuit current monitored in real time and environmental information, estimate the impact of the current conditions on the measurement results through the prediction model. If the prediction result shows that there may be a large error, automatically adjust the parameters in the measurement process, such as the measurement frequency range, measurement time interval, etc., or preprocess the collected data to adapt to different situations.

[0194] Parameter measurement: Measure the open-circuit voltage Voc and short-circuit current Isc of the measurement electrode in contact with the skin in real time, and record the timestamp information during the measurement at the same time.

[0195] Calculate the contact resistance: According to Ohm's law Rcontact = Voc / Isc, calculate the contact resistance between the measurement electrode and the skin.

[0196] Set the threshold: Through a large number of previous experiments, determine a reasonable range of contact resistance thresholds for different types of measurement electrodes and different skin characteristics. A contact resistance between 500Ω - 2000Ω is considered to have good contact quality.

[0197] Evaluation and judgment: Compare the calculated contact resistance with the set threshold range. If the contact resistance is within the threshold range, it is judged that the contact quality is good and the measurement data is reliable; if the contact resistance exceeds the threshold range, it indicates that there may be problems with the contact quality, such as the electrode not being in close contact with the skin, etc. At this time, a warning prompt can be issued, or the measurement data can be marked to consider the influence of the contact quality factor during subsequent analysis and processing.

[0198] When the contact impedance exceeds the normal range of 500Ω - 2000Ω, for example, due to dry skin causing the open-circuit voltage to increase (when the open-circuit voltage increases by 20% - 50% compared to the normal measurement value (1.5V), adjustment is required), automatically adjust the amplitude of the excitation current (when it is detected that the contact impedance exceeds the normal range or the open-circuit voltage abnormally increases, control the digital frequency synthesis (DDS) chip to automatically adjust parameters such as the amplitude of the excitation current and the measurement frequency point according to a preset program), reducing it from the standard 1mA to 0.5mA to ensure the accuracy and reliability of the calculation of characteristic parameters.

[0199] Construction of Multi-Feature Fusion Model: The fusion model adopts the architecture of Bidirectional Long Short-Term Memory Network with Attention Mechanism (BiLSTM-Attention). BiLSTM can simultaneously learn the forward and backward temporal dependencies of the PPG signal feature sequence and the impedance spectrum feature sequence, and fully exploit the dynamic change information of the two signals in the time dimension; the attention mechanism assigns dynamic weights to different features, highlighting the key features closely related to blood glucose concentration. The input layer of the model receives the normalized PPG signal feature vector (including time domain, frequency domain, and time-frequency domain optimized features) and the impedance spectrum feature vector (such as impedance, phase angle, and Cole-Cole model parameters in the selected frequency band) respectively. The hidden layer consists of two layers of BiLSTM units, with 64 neurons in each layer. The attention layer calculates the feature weights based on the output of the hidden layer, and the output layer outputs the final predicted blood glucose concentration value through a fully connected layer.

[0200] To construct an accurate blood glucose concentration prediction model, the construction and preprocessing of the dataset are crucial. First, widely collect physiological data of people of different ages, genders, body mass indices (BMI), and patients with different degrees of diabetes, including simultaneously collected PPG signals, impedance spectrum signals, and corresponding venous blood glucose true values. The data collection scenarios cover various daily states such as fasting, postprandial, before and after exercise, and sleep to ensure the diversity and comprehensiveness of the data. For example, 1000 volunteers were tracked and monitored for one month, and at least 5 groups of data were collected per person per day, accumulating more than 150,000 effective samples.

[0201] Strict preprocessing operations are performed on the collected raw signal data. For PPG signals, an optimized filtering algorithm is used to remove noise and baseline drift, and the signal amplitude is normalized to a specific interval, such as [0,1], through standardization processing to eliminate the influence of individual measurement device differences on signal strength; for impedance spectrum data, outlier removal is performed, and statistical methods are used to identify and remove impedance values that deviate significantly from the normal range due to poor electrode contact, motion artifacts, etc., and at the same time, logarithmic transformation is performed to enhance the stability and analyzability of the data.

[0202] In terms of model selection, a multi-layer perceptron (MLP) is used to construct a basic neural network model to explore the influence of different numbers of layers and nodes on blood glucose prediction; at the same time, long short-term memory networks (LSTM) and their variants, such as bidirectional LSTM (BiLSTM), are introduced to capture the long-term dependencies in the signal sequence considering the temporal characteristics of PPG signals and impedance spectrum signals to cope with the dynamic process of blood glucose concentration changes.

[0203] Taking five-fold cross-validation as an example, the preprocessed dataset is divided into a training set and a test set in an 8:2 ratio, and the training and test processes are repeated five times to comprehensively evaluate the model performance. Evaluation metrics such as mean absolute error (MAE), root mean square error (RMSE), and mean absolute relative error (MARD) are selected.

[0204] Optical signal acquisition component: Several near-infrared light-emitting diodes with specific wavelengths are selected. The light signals emitted by the near-infrared light-emitting diodes are gathered by a high-performance optical lens to one end of a composite quartz fiber bundle, and then shoot from the other end of the composite quartz fiber bundle towards the finger tip of the person to be monitored; the composite quartz fiber receives the reflected light from the finger tip and divides it into multiple beams, and each beam of reflected light is transmitted to the corresponding photodiode.

[0205] Impedance spectrum signal acquisition component: The impedance measurement electrode uses a flexible micro-nano structure electrode, and the material is a silver nanowire / polydimethylsiloxane (PDMS) composite material; the electrode array is designed for multi-band response and can measure the impedance change of biological tissues in the frequency range of 1 kHz - 1 MHz; a high-precision constant current source provides a stable alternating current excitation current for the electrode.

[0206] The electrode array is designed for multi-band response and can measure the impedance change of biological tissues in the frequency range of 1 kHz - 1 MHz.

[0207] The silver nanowire / polydimethylsiloxane (PDMS) composite material selected in this patent has high conductivity at both high and low frequencies, meeting the requirements of wide-band signal transmission. At the same time, the microcontroller outputs a frequency within a relatively wide frequency range (1 kHz - 1 MHz) through the control word of the DDS chip to measure the true impedance value at different frequencies.

[0208] In this patent, the electrode array is designed for multi-band response to improve the anti-interference ability and clinical use value. Because during the postprandial blood glucose rising stage, the low-frequency impedance drops significantly (about 15%), and during nocturnal hypoglycemia, the high-frequency phase angle rises significantly (about 8°), and the Cole-Cole parameter α is significantly correlated with diabetic neuropathy (r = 0.68, p < 0.01). Traditional single-band detection (such as 50 kHz) only reflects single tissue characteristics, and multi-band response realizes multi-level information fusion.

[0209] Signal conditioning circuit: It performs preprocessing such as amplifying, filtering, and analog-to-digital conversion on the original signals collected by the sensors. For the impedance measurement part, the voltage signal of the DDS signal generator is used as the input of the high-precision voltage-current conversion circuit. For the PPG signal, different wavelength LEDs are enabled by the LED driving circuit. The preamplifier circuit uses a low-noise and high-gain amplifier to convert the photocurrent signal into a voltage signal and perform preliminary amplification. The Butterworth band-pass filter circuit filters out the clutter and only retains the effective frequency components of 0.5 Hz - 5 Hz. The voltage signal collected by the voltage electrode for measuring the impedance spectrum signal is amplified by the differential amplifier circuit, the synchronous demodulation circuit converts the AC voltage signal into a DC signal, and finally the analog signal is converted into a digital signal by the analog-to-digital converter.

[0210] A non-invasive blood glucose detection device, and its adapted analysis software architecture, includes:

[0211] Underlying driver layer: It directly interacts with the hardware device and uses the microprocessor to control the hardware operation, including LED lighting, ADC acquisition, sampling frequency setting, and serial communication. It writes a driver program to control the LED driver chip to achieve precise control of the light source, drives the constant current source to output an excitation current with a specific frequency sequence, controls the voltage measurement and data acquisition process, and drives the analog-to-digital conversion chip to collect signals. It also includes functional modules such as the clock configuration, interrupt management, memory access drive, and communication interface drive of the microprocessor.

[0212] Middle data processing layer: It includes functional modules such as signal preprocessing, feature extraction, model calculation, and dual-signal fusion. The signal preprocessing module uses the Kalman filter algorithm to remove the residual noise and the baseline correction algorithm to eliminate the signal baseline drift. The feature extraction module extracts feature information from multiple dimensions of time domain, frequency domain, and time-frequency domain for the PPG signal, and fits and calculates the key parameters for the impedance spectrum signal based on data such as impedance values and phase angles at different frequencies in combination with the biological tissue equivalent circuit model. The model calculation module relies on a pre-trained deep learning model and takes the extracted PPG signal features and impedance spectrum features as inputs to initially estimate the blood glucose concentration value. The dual-signal fusion module uses the adaptive weighted algorithm to fuse the outputs of the PPG signal model and the impedance spectrum signal model to obtain the final blood glucose concentration prediction result.

[0213] Upper application layer: It faces users and medical professionals and provides an interaction interface and functional applications. The user interface presents real-time blood glucose values, blood glucose trend charts, health tips, and device connection status, etc. Users can perform operations such as device startup, measurement mode switching, and personal information entry through button or touch operations, and view historical measurement data reports. A professional diagnosis interface is reserved for medical professionals, and they can deeply analyze the patient's measurement data through the supporting software, conduct condition assessment, treatment plan adjustment, and remotely monitor the patient's blood glucose status.

[0214] PPG signal acquisition: An LED with a wavelength range of 435 nm - 1600 nm is used as the light source, and a high-responsivity silicon-based photodiode is used as the detector. A constant-current drive chip ensures stable light emission from the light source and controllable light intensity. The microprocessor controls the constant-current drive chip to adjust the light intensity through the PWM output pin. The preamplifier circuit selects a low-noise, high-gain instrumentation amplifier to convert the photocurrent signal into a voltage signal and amplify it. A second-order Butterworth band-pass filter filters out the clutter and retains the effective frequency components of 0.5 Hz - 5 Hz. Finally, the signal is converted into a digital signal by a high-precision ADC.

[0215] Impedance spectrum signal acquisition: Flexible micro-nano structured electrodes are used, and the electrode array is a 4×4 square array with an adjacent electrode spacing of 3 mm. A high-precision DDS chip is selected as the excitation source, and the microprocessor controls it to output an excitation current with a preset frequency sequence. The four-electrode method is adopted, with a pair of current electrodes injecting a 1 mA alternating current, and a pair of voltage electrodes collecting the voltage drop across both ends of the tissue. The voltage signal after differential amplification is converted into a DC signal by a synchronous demodulation circuit and then digitized by a 16-bit high-precision ADC.

[0216] Feature extraction and fusion steps:

[0217] PPG signal feature extraction: An improved wavelet transform algorithm is introduced to extract time-frequency domain features of the PPG signal. The db4 wavelet basis function is used for 5-layer decomposition to construct a feature vector. Combining with a deep learning algorithm, a PPG signal feature learning model based on a convolutional neural network is constructed.

[0218] Impedance spectrum feature extraction: The impedance amplitude in the low-frequency band of 3 kHz - 5 kHz and the phase angle in the high-frequency band of 800 kHz - 1 MHz are selected as key feature parameters, and the Cole-Cole model is used to fit and calculate the parameters α and β. An adaptive adjustment strategy is introduced to adjust the amplitude of the excitation current and the measurement frequency point according to the contact parameters between the measurement electrode and the skin.

[0219] Multi-feature fusion: A fusion model is constructed using the architecture of a bidirectional long short-term memory network with attention mechanism (BiLSTM-Attention). The input layer of the model receives the normalized PPG signal feature vector and impedance spectrum feature vector. The hidden layer consists of two layers of BiLSTM units. The attention layer calculates the feature weights, and the output layer outputs the final predicted blood glucose concentration value through a fully connected layer.

[0220] Steps for predicting model training and validation: Widely collect physiological data of different individuals in various daily states, including PPG signals, impedance spectrum signals, and true venous blood glucose values; preprocess the original signal data, remove noise, baseline drift and normalize the PPG signals, and remove outliers and perform logarithmic transformation on the impedance spectrum data; use multi-layer perceptrons, long short-term memory networks and their variants to build a prediction model, and evaluate the model performance with five-fold cross-validation. The evaluation metrics include mean absolute error, root mean square error, and mean absolute relative error.

[0221] The present invention focuses on the field of non-invasive blood glucose detection technology, aims to solve the limitations of the existing technology in this field, and proposes a non-invasive blood glucose detection system and device based on the combination of near-infrared light, impedance spectroscopy and deep learning technology.

[0222] At the hardware level of the device, the optical signal acquisition component selects near-infrared light-emitting diodes with specific wavelengths, and transmits the fingertip reflected light to the photodiode through an optical lens and a composite quartz fiber bundle. The impedance spectrum signal acquisition component uses a flexible micro-nano structure electrode to accurately measure the impedance change in the frequency range of 1 kHz - 1 MHz, and the signal conditioning circuit preprocesses the original signal to lay a foundation for subsequent analysis.

[0223] The software architecture adopts a hierarchical design. The underlying driver layer controls the hardware device and provides a unified interface for the upper layer; the middle data processing layer deeply processes the original data, and improves the data accuracy and detection accuracy through signal preprocessing, feature extraction, model calculation and dual-signal fusion; the upper application layer provides a simple interaction interface and professional diagnosis function for users and medical staff respectively.

[0224] In the detection method, first collect PPG signals and impedance spectrum signals respectively, extract targeted features from them, then use a bidirectional long short-term memory network based on the attention mechanism for multi-feature fusion, and train and validate the prediction model by collecting multi-source data.

[0225] The present invention combines near-infrared spectroscopy and impedance spectroscopy technologies, and uses deep learning algorithms to achieve high-precision non-invasive blood glucose detection. Compared with traditional detection technologies, it significantly reduces the pain of patients, reduces the risk of infection, simplifies the operation process, and the detection accuracy meets international standards, showing broad prospects in the fields of clinical application and health management.

[0226] The present invention discloses a non-invasive blood glucose detection system and device based on optical signals and impedance spectrum signals. The system includes, at the device hardware level, an optical signal acquisition circuit and an impedance spectrum signal acquisition circuit. The software architecture adopts a hierarchical design. The underlying driver layer controls the hardware device to acquire the corresponding PPG signal and impedance spectrum signal, and provides a unified interface for the upper layer; the middle data processing layer deeply processes the raw data, and inputs the acquired PPG signal and impedance spectrum signal into the trained blood glucose prediction network. The blood glucose prediction network includes a feature extraction module and a feature fusion and blood glucose prediction module. The upper application layer provides a simple interaction interface and a professional diagnosis function for users and medical staff respectively. The present invention combines near-infrared spectroscopy and impedance spectrum technology, and uses deep learning algorithms to achieve high-precision non-invasive blood glucose detection.

[0227] The foregoing are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals, characterized in that: include: An acquisition module is configured to: acquire a photoelectric signal and an impedance spectrum signal from the end of a finger of an object to be detected; A feature extraction module, which is configured to: extract features of the photoelectric signal and features of the impedance spectrum signal respectively; The feature fusion module is configured to: fuse the photoelectric signal features and the impedance spectrum signal features, and then predict the fused features to obtain blood glucose prediction results.

2. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 1, characterized in that: The step of obtaining the photoelectric signal and the impedance spectrum signal of the finger end of the object to be detected refers to: Visible light and infrared light are emitted to the skin at the end of the finger of the object to be detected, and the photoelectric signal fed back by the skin at the end of the finger is collected through the photodiode; The DDS signal generator transmits a voltage signal to the skin at the end of the finger of the object to be detected, and the electrode sheet collects the impedance spectrum signal fed back by the skin at the end of the finger.

3. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 1, characterized in that: The feature of the photoelectric signal and the feature of the impedance spectrum signal are extracted, wherein the process of extracting the feature of the photoelectric signal includes: (1-1): The photoelectric signal is decomposed into five layers using wavelet basis functions to obtain wavelet coefficients of information in different frequency ranges. The first feature of the photoelectric signal is constructed by analyzing the amplitude changes of high-frequency coefficients under the set scale; (1-2): Input the photoelectric signal into the feature learning model to obtain the second feature of the photoelectric signal.

4. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 3, characterized in that: The photoelectric signal is input into a feature learning model to obtain a second feature of the photoelectric signal, wherein the feature learning model includes: A CNN module, a first deep learning module, a second deep learning module, a global average pooling layer, a first fully connected layer, and an output layer connected in sequence; The CNN module comprises: a first convolutional layer, a first batch normalization layer, a first activation function layer, a first maximum pooling layer, a second convolutional layer, a second batch normalization layer, a second activation function layer, a second maximum pooling layer, a third convolutional layer, a third batch normalization layer and a third activation function layer connected in sequence; wherein the input end of the first convolutional layer is also connected to the input end of the second convolutional layer; the input end of the second convolutional layer is also connected to the input end of the third convolutional layer; The internal structure of the first deep learning module is the same as that of the second deep learning module. The first deep learning module includes: a first dense block and a first transition layer connected in sequence; The input end of the first dense block Dense Block is the input end of the first deep learning module, and the output end of the first transition layer Transition Layer is the output end of the first deep learning module; The internal structure of the first deep learning module is the same as that of the second deep learning module. The first deep learning module includes: The first dense block and the first transition layer are connected in sequence; The input end of the first dense block Dense Block is the input end of the first deep learning module, and the output end of the first transition layer Transition Layer is the output end of the first deep learning module.

5. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 4, characterized in that: The working process of the first deep learning module includes: the first dense module Dense Block is used to extract information features and use residual connections to achieve feature reuse and extract features; the first transition layer Transition Layer is used to reduce feature channels; the first dense module Dense Block includes: a fourth convolutional layer, a fourth batch normalization layer, a fourth activation function layer, a first adder and a fifth convolutional layer connected in sequence; wherein the input end of the first adder is also connected to the input end of the fourth convolutional layer; the fourth convolutional layer is used to perform further convolution operations on input features to extract local features; the fourth batch normalization layer is used to normalize the output of the fourth convolutional layer; the fourth activation function layer is used to introduce nonlinear characteristics and enhance the expression ability of the model; the first adder is used to realize residual connections and add the input features of the fourth convolutional layer to the features after convolution, normalization and activation function processing; the fifth convolutional layer is used to re-convolute the features after residual connection to further extract and integrate features; The first transition layer Transition Layer includes: a fifth batch normalization layer, a sixth convolution layer, an average pooling layer and an output layer connected in sequence; the fifth batch normalization layer is used for the third batch normalization layer to normalize the mean and variance of the feature data to 0 and 1 respectively; the sixth convolution layer is used to reduce the dimension of the feature or adjust the number of channels through 2×2 convolution; the average pooling layer is used to reduce the size of the feature map; the output layer is used to output the features processed by the previous layers as the final result of the transition layer to the subsequent network layer.

6. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 1, characterized in that: The feature extraction of the photoelectric signal and the feature extraction of the impedance spectrum signal, wherein the feature extraction process of the impedance spectrum signal includes: Assume that the impedance value measured at different frequencies ω is Z m (ω); According to the Cole-Cole model: Among them, Z0 represents the low-frequency limiting impedance, α represents the polarization degree parameter, β represents the diffusion index, and τ represents the time constant; To minimize the error function As the goal, the optimal impedance value Z0 of biological tissue under DC conditions, the frequency dispersion degree α of biological tissue impedance and the curvature degree β of the impedance spectrum curve are iteratively solved.

7. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 1, characterized in that: The feature fusion module is configured to: perform feature fusion on the photoelectric signal feature and the impedance spectrum signal feature, and then predict the fused feature to obtain a blood glucose prediction result; wherein the feature fusion is achieved by using a trained feature fusion model; The trained feature fusion model includes: The seventh convolutional layer, the sixth batch normalization layer, the first BiLSTM model, the seventh batch normalization layer, the second BiLSTM model, the self-attention mechanism layer, the first multi-layer perceptron MLP, the Dropout layer, the second multi-layer perceptron MLP, the second fully connected layer and the prediction layer are connected in sequence.

8. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 7, characterized in that: The feature fusion model, the working process includes: The first BiLSTM model and the second BiLSTM model: For the input photoelectric signal feature and impedance spectrum signal feature sequences, the first BiLSTM model and the second BiLSTM model extract key information from the features at different time steps, respectively, and explore the dynamic change rules of the two signals in the time dimension; Self-attention mechanism layer: The self-attention mechanism calculates the importance of each feature based on the correlation between features; The second fully connected layer: further linearly combines and integrates the features processed by the self-attention mechanism.

9. The non-invasive blood glucose detection system based on photoelectric signals and impedance spectrum signals as claimed in claim 8, characterized in that: The feature fusion model, the working process includes: Assume that the photoelectric signal characteristics are The impedance spectrum signal characteristics are The final fusion result is The expression of the adaptive weighting algorithm is: Among them, the weights w1 and w2 satisfy w1+w2=1, and are dynamically adjusted through the following adaptive strategy: Among them, r PPG Represents the predicted value of the photoelectric signal Pearson correlation coefficient with actual blood glucose value; reflects the correlation between photoelectric signal and blood glucose; r Z Indicates the predicted value of the impedance spectrum signal Pearson correlation coefficient with actual blood glucose value; reflects the correlation between impedance spectrum signal and blood glucose; is the variance of the predicted value of the photoelectric signal; is the variance of the predicted value of the impedance spectroscopy signal.

10. A non-invasive blood glucose detection device based on photoelectric signals and impedance spectrum signals, characterized in that: include: Photoelectric signal acquisition subsystem and impedance spectrum signal acquisition subsystem; The photoelectric signal acquisition subsystem includes: a visible light LED and a near-infrared LED, the visible light LED and the near-infrared LED transmit the emitted light to a lens, the lens transmits the light to a first end of a transmitting optical fiber, the transmitting optical fiber transmits the light to a second end of the transmitting optical fiber, the second end of the transmitting optical fiber divides the light into multiple beams and then transmits them to the skin at the end of the finger of the object to be detected, the first end of the receiving optical fiber receives each beam of reflected light fed back by the skin at the end of the finger, the second end of the receiving optical fiber is connected to the first end of the photodiode, the second end of the photodiode is connected to the analog-to-digital conversion unit through a preamplifier circuit, and the analog-to-digital conversion unit is connected to a microcontroller; The impedance spectrum signal acquisition subsystem includes: a DDS signal generator, which sends the emitted voltage signal to the first electrode sheet and the second electrode sheet, and the first electrode sheet and the second electrode sheet transmit the voltage signal to the end of the finger of the object to be detected, and collects the voltage signal fed back from the end of the finger of the object to be detected through the third electrode sheet and the fourth electrode sheet, amplifies the fed back voltage signal through the differential amplifier circuit, converts the amplified voltage signal into a DC voltage signal through the synchronous demodulation circuit, and finally converts the DC voltage signal from an analog signal to a digital signal through the analog-to-digital conversion circuit, and transmits the digital signal to the microcontroller; The microcontroller transmits the collected photoelectric signal and impedance spectrum signal to the host computer, which processes the two signals to obtain the predicted blood glucose value.

Images