Noninvasive blood sugar concentration detection method based on near infrared spectrum detection

Abstract

The invention discloses a noninvasive blood glucose concentration detection method based on near infrared spectrum detection, and relates to the technical field of blood glucose concentration detection.The noninvasive blood glucose concentration detection method comprises the following steps that target near infrared spectrum data are obtained, and noise calibration is conducted; dividing the spectral domain frequency band of the calibrated spectrum into a short wave region sensitive to melanin and a long wave region insensitive to melanin; estimating a first gain of each region based on partition gain white balance; performing constraint optimization on the first gain in combination with a priori function of melanin content-spectral shape change fitted by an offline Monte Carlo library, limiting the gain to a priori low-dimensional projection online to obtain a second gain combined with a preset cross-region smoothing function, outputting a spectrum after white balance correction as a first spectrum, and outputting a second spectrum after white balance correction as a second spectrum; the method is applied to blood glucose concentration detection, and the problems of spectrum baseline deviation and blood glucose sensitive wave band peak shape distortion caused by individual melanin difference are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood glucose concentration detection, and more particularly to a non-invasive blood glucose concentration detection method based on near-infrared spectroscopy. BACKGROUND

[0002] The non-invasive blood glucose concentration detection method based on near-infrared spectroscopy obtains a near-infrared spectroscopy signal by irradiating the skin, and establishes a mapping relationship by using a statistical or machine learning model of the spectroscopy feature and the blood glucose concentration to realize the prediction of the blood glucose concentration. Such a method is widely used in the prior art and can avoid the discomfort and infection risk caused by traditional invasive blood sampling, and has the potential of non-invasive, rapid and continuous monitoring. The conventional process includes spectroscopy acquisition, preprocessing, feature extraction, modeling and blood glucose prediction, which is the core step of the existing non-invasive blood glucose detection system.

[0003] However, the prior art has significant deficiencies. Due to the difference in skin color of individuals, the difference in melanin content causes the difference in absorption of near-infrared light in the skin, which causes the baseline shift of the spectrum, especially in the short-wave region sensitive to melanin. This baseline shift interferes with the blood glucose sensitive feature signal and directly affects the accuracy of the blood glucose concentration prediction. In addition, the difference in response of different light sources and detectors and the change in environmental light conditions may also cause inconsistency in the spectroscopy signal, making it difficult for traditional spectroscopy processing methods to ensure the comparability and stability across individuals.

[0004] In other fields of optical imaging and signal processing, the white balance technique is commonly used to compensate for the overall brightness and color deviation in the image or spectroscopy signal, and by adjusting the gain or reference, the signal remains consistent under different acquisition conditions. The white balance method can quickly correct the overall spectral intensity and reduce the signal shift caused by changes in external conditions, and is a mature technology in existing image processing and spectroscopy correction.

[0005] However, directly applying the white balance method to non-invasive blood glucose detection based on near-infrared spectroscopy has limitations. Traditional white balance usually processes the overall gain of the full spectrum, lacks a partition correction mechanism for individual melanin differences, and is prone to cause local spectral peak distortion, especially in the blood glucose sensitive band, which may affect the feature stability. In addition, the white balance freely adjusts the gain, which cannot constrain the physically feasible range of spectral morphology, making it difficult to ensure the comparability of the spectrum between individuals with different skin colors and the detection accuracy. Therefore, it is necessary to combine skin color information and optical priors to perform customized gain constraint and partition smoothing optimization to realize reliable blood glucose concentration detection.

[0006] To solve the above problems, the present application provides a solution. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, embodiments of the present application provide a non-invasive blood glucose concentration detection method based on near-infrared spectrum detection, which solves the problems of spectral baseline offset and blood glucose sensitive band peak shape distortion caused by individual melanin difference by combining white balance based on partition gain with offline Monte Carlo prior function low-dimensional projection and cross-zone smoothing optimization.

[0008] To achieve the above object, the present application provides the following technical solutions: The non-invasive blood glucose concentration detection method based on near-infrared spectrum detection comprises the following steps: acquiring target near-infrared spectrum data and performing noise calibration; dividing the spectral domain frequency band of the calibrated spectrum into a short-wave zone sensitive to melanin and a long-wave zone insensitive to melanin; estimating a first gain of each zone based on partition gain white balance; combining a prior function of melanin content-spectral shape change fitted by an offline Monte Carlo library to constrain and optimize the first gain, and limiting the gain to a low-dimensional projection of the prior function online to obtain a second gain; combining the second gain with a preset cross-zone smoothing function, outputting the white balance corrected spectrum as a first spectrum, and applying it to blood glucose concentration detection.

[0009] In a preferred embodiment, the step of acquiring target near-infrared spectrum data and performing noise calibration specifically comprises: irradiating the surface of the measured skin with a near-infrared light source, receiving the reflection through a detector, amplifying and analog-to-digital converting the spectrum signal through an analog front-end circuit to obtain digitized spectrum data, and performing noise calibration on the digitized spectrum data, wherein the noise calibration includes dark current correction, detector noise suppression, and system response normalization processing, thereby obtaining the target near-infrared spectrum data after noise calibration.

[0010] In a preferred embodiment, the step of dividing the spectral domain frequency band of the calibrated spectrum into a short-wave zone sensitive to melanin and a long-wave zone insensitive to melanin specifically comprises: determining the short-wave zone sensitive to melanin and the long-wave zone insensitive to melanin in the spectrum according to a preset band division criterion; segmenting the calibrated spectrum data according to the boundary conditions of the short-wave zone and the long-wave zone to obtain short-wave zone spectrum and long-wave zone spectrum; and introducing a transition band at the boundary and performing smoothing processing on the data in the transition band.

[0011] In a preferred embodiment, the first gain of each region based on the partition gain white balance estimation is specifically: the first gain is the gain value after correction by the cross-region smoothing constraint on the basis of the normalized gain, respectively corresponding to the short-wave region and the long-wave region, used for white balance correction of the spectrum, specifically: the average spectral intensity of the short-wave region spectrum and the long-wave region spectrum is calculated respectively, used for representing the spectral response characteristics of each region; the corresponding initial white balance gain value is obtained by analyzing and calculating the difference between the spectral response characteristics and the preset reference benchmark; the reference benchmark is a standard spectrum obtained in advance; the initial white balance gain value of each region is normalized; the first gain of each region is obtained by introducing the cross-region smoothing constraint based on the continuity principle of the tissue spectrum in the wavelength dimension on the basis of the normalized gain.

[0012] In a preferred embodiment, the first gain is constrained and optimized by combining the prior function of the melanin content-spectral shape change of the offline Monte Carlo library fitting, and the gain is limited to the low-dimensional projection of the prior function online to obtain the second gain, specifically: the offline Monte Carlo photon transport library is used to simulate the optical response of the tissue for different melanin contents, and the prior function of the melanin content-spectral shape change is fitted; the first gain is projected to the low-dimensional basis function space of the prior function as a low-dimensional projection; the low-dimensional projection constrains the online gain from deviating from the physically feasible spectral shape change range; a regularization term is introduced to limit the spectral gain of the glucose sensitive region; the gain after optimization of the low-dimensional projection and the regularization term is defined as the second gain.

[0013] In a preferred embodiment, the second gain combines a preset cross-region smoothing function, and outputs the white balance corrected spectrum as the first spectrum, specifically: the cross-region smoothing function is preset according to the short-wave region and long-wave region segmentation boundary of the calibrated spectrum and the continuity principle of the tissue spectrum in the wavelength dimension; the second gain is combined with the cross-region smoothing function to obtain a smoothed gain curve; the smoothed gain curve is applied to the corresponding wavelength points of the calibrated spectrum data to obtain the white balance corrected spectrum, i.e. the first spectrum.

[0014] In a preferred embodiment, the application is applied to blood glucose concentration detection, specifically: the first spectrum is subjected to conventional spectral pretreatment; the pretreated spectral data is subjected to feature dimension reduction or screening; a blood glucose concentration prediction model is established based on the extracted spectral features, used for mapping the spectral features to the blood glucose concentration prediction value; and the predicted blood glucose concentration is output.

[0015] The non-invasive blood glucose concentration detection method based on near-infrared spectrum detection has the following technical effects and advantages: 1.The present application solves the problem of baseline shift of near-infrared spectrum caused by the difference in melanin content of different skin color individuals by using a white balance method based on partition gain to estimate and correct the spectrum in the short wave region and the long wave region respectively. By calculating the average spectral intensity of each region first, then comparing it with the reference spectrum to get the initial gain, and then through normalization and cross-region smoothing constraint processing to generate the first gain, the partition correction of white balance is realized. This step can effectively reduce the spectral baseline shift caused by the difference in melanin, so that the corrected spectrum of individuals with different skin colors has comparability, and ensures the stability of the blood glucose characteristic signal in the short wave region, thereby providing consistent and reliable spectral input for subsequent blood glucose concentration detection. Through this method, the accuracy and applicability of non-invasive blood glucose detection in multi-skin color population can be significantly improved, while maintaining the overall continuity of the spectrum, avoiding the baseline drift problem caused by traditional methods under the difference in skin color.

[0016] 2.The present application optimizes the white balance correction effect, combines the melanin content-spectrum shape change prior function fitted by the off-line Monte Carlo photon transport library, and constrains the first gain in low dimension projection, and introduces a regularization term for the gain in the glucose sensitive region to obtain the second gain. At the same time, the second gain is combined with the preset cross-region smoothing function to act on the calibration spectrum, and a continuous and smooth white balance corrected spectrum is output. This step not only retains the correction effect of white balance on skin color difference, but also avoids the spectral peak shape distortion caused by the free adjustment of the first gain, and ensures the stability of the blood glucose sensitive band. Through this gain constraint optimization and cross-region smoothing processing, the present application effectively solves the local spectral mutation problem that may occur in the actual application of traditional white balance, so that the corrected spectrum not only meets the physical feasibility, but also enhances the accuracy and reliability of blood glucose concentration detection, and is suitable for non-invasive blood glucose detection scenes with multi-skin color and multi-tissue optical properties. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The present application is based on the process diagram of non-invasive blood glucose concentration detection method based on near-infrared spectrum detection. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] Embodiment 1, Figure 1 The present application is based on the process diagram of non-invasive blood glucose concentration detection method based on near-infrared spectrum detection. S1, obtain the target near-infrared spectrum data and perform noise calibration.

[0020] In the embodiment, the target near-infrared spectrum data is acquired and noise calibration is performed, specifically as follows: The skin surface to be measured is irradiated by a near-infrared light source, and a reflected or transmitted spectrum signal is received by a detector, the spectrum signal is amplified and analog-to-digital converted by an analog front-end circuit to obtain digitized spectrum data; The digitized spectrum data is subjected to noise calibration, the noise calibration including dark current correction, detector background noise suppression and system response normalization processing, thereby obtaining target near-infrared spectrum data subjected to noise calibration.

[0021] It should be noted that the near-infrared light source can be a wide-spectrum near-infrared light source or a narrow-band near-infrared laser light source for irradiating the skin surface to be measured. The wide-spectrum light source has the advantage of wide coverage, and multiple waveband information can be acquired at one time; the narrow-band light source has the characteristics of strong monochromaticity and high signal-to-noise ratio, facilitating detection of specific sensitive wavebands.

[0022] It should be noted that the analog front-end circuit can use a transimpedance amplifier or a low-noise amplifier to convert and amplify the photoelectric current to a voltage signal, and a low-pass filter is used to suppress high-frequency noise to improve the signal-to-noise ratio. These circuit modules are common electronic signal conditioning methods.

[0023] It should be noted that the noise calibration further includes detector background noise suppression. The background noise can be reduced by repeated sampling and averaging, that is, using the statistical principle that "random noise tends to cancel out after multiple superimpositions", thereby improving the quality of the effective signal.

[0024] It should be noted that the noise calibration further includes system response normalization. Specifically, a reference standard (such as a white plate or a standard reflector) is used to collect a reference spectrum, and the reference spectrum is used to normalize the target spectrum to eliminate the effects of light source spectral non-uniformity and detector response curve differences.

[0025] S2, the spectral domain band of the calibrated spectrum is divided into a short wave region sensitive to melanin and a long wave region not sensitive to melanin.

[0026] In the embodiment, the spectral domain band of the calibrated spectrum is divided into a short wave region sensitive to melanin and a long wave region not sensitive to melanin, specifically as follows: According to a preset waveband division criterion, a short wave region sensitive to melanin and a long wave region not sensitive to melanin in the spectrum are determined; According to the boundary conditions of the short wave region and the long wave region, the calibrated spectrum data is segmented to obtain a short wave region spectrum and a long wave region spectrum; The boundary also includes introducing a transition band, and the data in the transition band is subjected to smoothing processing.

[0027] It should be noted that the "preset wavelength band division criterion" can be obtained in various ways, for example, by referring to an existing tissue spectrum database or by collecting and statistically summarizing spectrum data of different skin color populations through pre-experiments.

[0028] It should be noted that the boundary conditions of the short-wave region and the long-wave region are not fixed and can be adjusted according to different application scenarios. For example, in a medical detection scenario, a wavelength band with significant melanin absorption can be selected as the short-wave region; and in a non-medical imaging or detection scenario, a wavelength band with a high signal-to-noise ratio can be selected as a preferred division basis.

[0029] It should be noted that the division of the spectrum data is realized by software, that is, after data acquisition is completed, the spectrum data is cut according to the wavelength index through a programming algorithm; or the division of the spectrum data is realized by hardware, for example, different narrow-band optical filters are configured at the spectrum acquisition end, and the spectrum is divided into a short-wave region and a long-wave region during the acquisition.

[0030] It should be noted that the introduction of the transition band is mainly used to reduce the influence of the boundary effect on subsequent white balance calculation. The boundary effect can cause discontinuity or discontinuity when the short-wave region and the long-wave region are spliced, and the continuity and stability of the spectrum transition can be ensured by smoothing the data in the transition band (for example, using a linear weighting function or a Gaussian weighting function).

[0031] It should be noted that the smoothing process is not limited to a single algorithm, and different methods such as mean filtering, wavelet denoising, and low-rank approximation can be used to adapt to different noise environments and constraints on computing resources.

[0032] In this embodiment, the calibrated spectrum is divided into a short-wave region sensitive to melanin and a long-wave region not sensitive to melanin, which means that the part of the spectrum affected by the strong absorption of melanin and the relatively stable part can be distinguished. Due to the difference in skin color between individuals, the content of melanin is significantly different, which will cause the baseline of the spectrum to shift in the short-wave region, thereby seriously interfering with the accurate extraction of blood glucose concentration. By dividing the spectrum, on the one hand, a compensation mechanism can be established in the short-wave region to weaken the influence of the difference in melanin absorption, and on the other hand, the long-wave region can be used as a reference region to provide a stable reference for the calculation of the white balance gain. This step not only makes the subsequent white balance correction more targeted, but also significantly improves the consistency of spectrum processing between individuals with different skin colors, thereby improving the reliability and universality of the blood glucose concentration detection result.

[0033] S3, estimating a first gain of each region based on the zoned gain white balance.

[0034] In this embodiment, the estimation of the first gain of each region based on the zoned gain white balance is specifically: The first gain is the gain value modified by the cross-zone smoothing constraint on the basis of the normalized gain, corresponding to the short-wave zone and the long-wave zone respectively, used for white balance correction of the spectrum, specifically: The average spectral intensity of the short-wave zone spectrum and the long-wave zone spectrum is calculated respectively, used to represent the spectral response characteristics of each zone; According to the difference between the spectral response characteristics and the preset reference benchmark, the corresponding initial white balance gain value is calculated and analyzed, and the reference benchmark is a standard spectrum obtained in advance; The initial white balance gain value of each zone is normalized; On the basis of the normalized gain, the cross-zone smoothing constraint is introduced based on the continuity principle of the tissue spectrum in the wavelength dimension, and the first gain of each zone is obtained.

[0035] Further, the average spectral intensity is used to quantify the overall spectral energy level of different wavebands, and the following is a specific calculation formula of the average spectral intensity: ; And ; In the formula, is the short-wave zone spectrum, is the long-wave zone spectrum, and is the number of sampling points, and is the average spectral intensity, short and long are the marks of the long-wave zone and the short-wave zone respectively, and is the wavelength set, is the wavelength.

[0036] Further, the following is a specific calculation formula of the initial white balance gain: , And ; In the formula, and are the average intensities of the reference spectrum in the short-wave zone and the long-wave zone respectively, and the white balance gain ensures that the energy levels of the short-wave zone and the long-wave zone can be corrected to a unified benchmark under different skin color individuals, and are the initial white balance gains of the short-wave zone and the long-wave zone respectively.

[0037] Further, in order to avoid the overall correction imbalance caused by the excessive or insufficient gain of individual zones, the initial gain is normalized, and the following is a specific calculation formula of the initial white balance gain normalization: , In the formula is the initial white balance gain for the k-th region, is the initial white balance gain for the normalized k-th region.

[0038] Further, on the basis of the normalized gain, a cross-region smoothing constraint is introduced to reduce the abrupt change caused by independent estimation of different regions, which is introduced based on the continuity principle of the tissue spectrum. The actual spectrum should remain smooth in the wavelength dimension, so after independent estimation of the gain in different regions, the global mean gain is introduced as a regularization term to constrain the region gain, thereby avoiding the unnatural abrupt change caused by the division of the region. The constraint function can be expressed as: and In the formula, and are the normalized gain of the short-wave region and the long-wave region; is the global mean gain, that is, ; is a smoothing coefficient for controlling the trade-off between cross-region independence and overall continuity, and are the first gain of the short-wave region and the first gain of the long-wave region, respectively.

[0039] It should be noted that in the present embodiment, the average intensity of the short-wave region spectrum and the long-wave region spectrum is obtained by calculating the arithmetic mean of the spectrum values of all sampling wavelengths in the respective region. This step is used to quantify the overall spectral response characteristics of each region, providing basic data for subsequent white balance gain calculation. Through this processing, the spectral energy differences of different skin color individuals in different wavebands can be effectively characterized, providing an operable basis for subsequent gain correction.

[0040] It should be noted that the initial gain is calculated by the ratio of the average spectrum intensity of each region to the average intensity of the preset reference spectrum, wherein the reference spectrum can be measured by a standard light source or obtained from a large number of average human spectra. This step is used to preliminarily correct each region, compensate for the baseline shift of the short-wave region caused by the difference in melanin absorption of different individuals, and lay a foundation for subsequent normalization and cross-region smoothing processing.

[0041] It should be noted that the normalization processing is to divide the initial gain of each region by the mean or a preset normalization coefficient, so that the gain of each region is relatively comparable. This processing can avoid the imbalance of overall spectrum correction caused by too large or too small gain of a certain region, while ensuring the input stability of the cross-region smoothing constraint, thereby providing a reliable basis for generating a continuous and smooth first gain.

[0042] ​It should be noted that the setting of the cross-zone smoothing constraint is based on the continuity principle of the organization spectrum in the wavelength dimension, and the normalized gain is smoothed by using weighted average or regularization function. This step can eliminate the boundary mutation caused by the division of short-wave zone and long-wave zone, ensure the continuous change of the gain of each zone, and thus improve the smoothness of the spectrum after white balance correction and the accuracy of blood glucose concentration detection.

[0043] It should be noted that the first gain of each zone after correction by the cross-zone smoothing constraint is directly used for white balance correction of the corresponding zone spectrum, and the first spectrum is output. This step can effectively reduce the spectral baseline offset caused by individual differences of melanin, provide accurate and comparable spectral input for subsequent blood glucose concentration detection, and improve the reliability and accuracy of the non-invasive blood glucose detection method.

[0044] In the embodiment, the first gain of each zone spectrum is estimated by the white balance method based on the partition gain, and the core purpose is to solve the problem of spectral baseline offset caused by individual skin color difference. Since melanin obviously absorbs near-infrared light in the short-wave zone, the difference in melanin content of different individuals will cause the spectral baseline in the short-wave zone to shift downward, and the long-wave zone is less affected, thereby affecting the consistency of the overall spectrum and the accuracy of blood glucose concentration detection. By first calculating the average spectral intensity of each zone and comparing it with the reference baseline spectrum, the initial gain is obtained, and then the first gain is obtained after normalization and cross-zone smoothing constraint processing. This step can correct the short-wave zone and the long-wave zone respectively, make the corrected spectral baseline consistent, reduce the influence of individual differences, and thus provide stable and comparable spectral data for subsequent blood glucose concentration analysis, improve the reliability and accuracy of the non-invasive detection method.

[0045] S4, combining the prior function of melanin content-spectrum shape change fitted by the offline Monte Carlo library, the first gain is constrained and optimized, and the gain is limited to a low-dimensional projection of the prior function online to obtain the second gain.

[0046] In the embodiment, the first gain is constrained and optimized by combining the prior function of melanin content-spectrum shape change fitted by the offline Monte Carlo library, and the gain is limited to a low-dimensional projection of the prior function online to obtain the second gain, specifically: Using the offline Monte Carlo photon transport library, the optical response of the tissue is simulated for different melanin contents, and the prior function of melanin content-spectrum shape change is fitted; Project the first gain into the low-dimensional basis function space of the prior function as a low-dimensional projection; Through the low-dimensional projection, the online gain is constrained not to deviate from the physically feasible spectrum shape change range; Introducing a regularization term to limit the spectral gain of the glucose sensitive zone; The gain after optimization of the low-dimensional projection and the regularization term is defined as the second gain.

[0047] Further, the off-line Monte Carlo photon transport library is used to simulate the optical response of the tissue with different melanin contents, and a prior function of the melanin content-spectrum shape change is fitted, specifically: For each melanin content , the Monte Carlo photon transport is used to simulate the scattering and absorption of photons in the near-infrared band in the tissue, and the tissue transmission spectrum is obtained ; Each set of simulated spectra is normalized to eliminate the absolute intensity difference The normalized spectrum and the melanin content are functionally fitted to obtain the melanin content-spectrum shape change; The following is an expression of a feasible melanin content-spectrum shape change prior function: , The following is an example of the normalization processing calculation of the simulated spectrum: ; In the formula, is the prior function, is the entire simulated wavelength set; is the fitting coefficient varying with the melanin content, and M is the polynomial order; usually 3-5, which can accurately fit the trend of the spectrum shape change with the melanin content, is the normalized simulated spectrum.

[0048] It should be noted that the optical parameters of the main absorption components of water, fat, and hemoglobin in the tissue are considered in the tissue transmission spectrum simulation process to ensure that the spectrum response approximates the actual human tissue characteristics.

[0049] It should be noted that in the embodiment, the Monte Carlo photon transport is used to simulate the scattering and absorption of photons in the tissue to obtain the transmission spectrum under different melanin content conditions. This step can truly reflect the propagation behavior of light in the skin tissue, provide a physical basis for constructing the melanin content-spectrum shape prior function, and make the online gain constraint have a theoretical support.

[0050] It should be noted that each set of simulated spectra is normalized to eliminate the absolute intensity difference. This step ensures that the spectra with different melanin contents can be compared and fitted in the same dimension, avoids the interference of external factors such as light source power and detector sensitivity on the prior function fitting, and improves the usability and stability of the prior function.

[0051] It should be noted that the normalized spectrum is functionally fitted with the melanin content to obtain a prior function of melanin content-spectrum shape change. The prior function obtained by fitting can describe the trend of the influence of melanin content change on the spectrum shape, provide low-dimensional physical constraints for online gain projection, and prevent spectrum peak shape distortion caused by the first gain adjustment.

[0052] It should be noted that the fitting of the prior function of melanin content-spectrum shape change reflects the change rule of the spectrum shape in the short-wave region and the long-wave region under different melanin levels, and provides a theoretical basis for subsequent gain constraints.

[0053] Further, the first gain is projected to the low-dimensional basis function space of the prior function, which is a low-dimensional projection, specifically: extracting several dominant modes from the melanin content-spectrum shape change prior function obtained offline as low-dimensional basis functions where K is the number of basis functions, k is the basis function label, and is used to describe the main trend of the spectrum shape change with the melanin content; projecting the first gain to the selected low-dimensional basis function space to calculate the low-dimensional coefficient; reconstructing the low-dimensional gain curve according to the projection coefficient, that is, the preliminary form of the second gain; It should be noted that the following is a feasible example of low-dimensional coefficient calculation: , wherein is the kth basis function value, indicating the projection coefficient of the first gain in the basis function direction, is the projection coefficient of the first gain in the basis function direction, is the first gain at the wavelength . The following is a feasible example of the preliminary form of the second gain: , wherein, is the second gain at the wavelength .

[0054] It should be noted that the low-dimensional projection retains the main component of the first gain in the physically feasible spectrum shape, while filtering out noise or deviations that do not conform to the prior. By projecting the first gain to the low-dimensional basis function space of the prior function, the gain change can be constrained online to conform to the spectrum shape of the melanin physical prior, prevent spectrum peak shape distortion caused by free gain adjustment, and ensure that the spectrum after white balance correction has comparability and stability between different skin color individuals.

[0055] The online gain constraint does not deviate from the physically feasible spectral shape variation range through low-dimensional projection; a regularization term is introduced to limit the spectral gain in the glucose-sensitive region; the gain after low-dimensional projection and regularization optimization is defined as the second gain, specifically: The first gain is obtained by projecting to the low-dimensional basis function space of the offline Monte Carlo prior function to obtain the preliminary low-dimensional gain. During online optimization, the reconstructed gain curve is adjusted to fit the first gain as much as possible, while ensuring the physical feasibility of gain variation and avoiding unreasonable spectral peaks or depressions. The optimization method can be realized by using existing least squares fitting or convex optimization technology with constraints.

[0056] During optimization, the spectral band sensitive to blood glucose is set with a separate constraint condition to ensure that the gain variation in this region does not exceed the preset range, and the stability of the blood glucose characteristic spectral peak is ensured. The specific implementation can use a weighted penalty term or threshold clipping method to make the second gain smooth in the sensitive region and avoid affecting the accuracy of blood glucose concentration detection.

[0057] The gain curve after low-dimensional projection optimization and sensitive region constraint processing is the second gain. The second gain can be directly used for spectral white balance correction, so that the corrected spectrum not only corrects the baseline shift caused by individual melanin differences, but also maintains the stability of the blood glucose characteristic signal, providing reliable spectral data for subsequent blood glucose concentration detection.

[0058] It should be noted that in the existing non-invasive blood glucose concentration detection method based on near-infrared spectroscopy, due to the difference in individual skin color, the melanin absorption is different, and the spectral baseline has a significant shift, which directly affects the stability of the blood glucose characteristic signal, and the traditional method is difficult to accurately correct. To alleviate this problem, the white balance technology is introduced, and the white balance is preliminarily corrected by partition gain, but in actual application, the high degree of freedom of the first gain may cause local spectral peak distortion, especially in the blood glucose sensitive band, which may affect the detection accuracy.

[0059] To solve the above problems, this step combines the melanin content-spectral shape variation prior function fitted by the offline Monte Carlo photon transport library to constrain and optimize the first gain, and limits the gain online to be a low-dimensional prior projection to obtain the second gain. In this way, the white balance correction effect on the spectral baseline of different skin color individuals is retained, and the spectral peak distortion caused by the free adjustment of the gain is effectively prevented, the stability and detection accuracy of the spectrum in the blood glucose sensitive band after white balance correction are improved, and reliable non-invasive blood glucose concentration detection in multiple skin color individuals is realized.

[0060] S5, the second gain combines a preset cross-region smoothing function to output the white balance corrected spectrum as the first spectrum, and applies it to blood glucose concentration detection.

[0061] In the embodiment, the second gain is combined with the preset cross-region smoothing function, and a white balance corrected spectrum is output as the first spectrum, specifically: According to the short-wave region and long-wave region division boundary of the calibrated spectrum and the continuity principle of the tissue spectrum in the wavelength dimension, a preset cross-region smoothing function is set; The second gain is combined with the cross-region smoothing function to obtain a smoothed gain curve; The smoothed gain curve is applied to the corresponding wavelength points of the calibrated spectrum data to obtain a white balance corrected spectrum, i.e., the first spectrum.

[0062] It should be noted that the smoothing function can be obtained by existing signal processing technology, for example, using polynomial fitting, moving average or Gaussian convolution method to smooth and fit the spectrum gain of the boundary transition zone. The parameters of the smoothing function can be preset or adjusted according to the experimentally measured different individual skin colors and tissue optical properties, to ensure the smooth and continuous cross-region gain. It should be noted that the gain curve processing makes the short-wave region and long-wave region gain smoothly transition in the boundary region, avoiding local jump or discontinuity causing spectrum distortion.

[0063] It should be noted that the following is a quantized expression of the feasible smoothed gain curve: , In the formula, is the cross-region smoothing function, is the value of the smoothed gain curve at .

[0064] It should be noted that the following is a quantized expression of the feasible first spectrum of each wavelength : , In the formula, is the first spectrum, is the calibrated spectrum data.

[0065] It should be noted that in the embodiment, the preset cross-region smoothing function is to ensure the continuous transition of the short-wave region and long-wave region gain in the boundary region, avoiding gain mutation causing spectrum distortion. The smoothing function can be obtained based on existing signal processing methods, such as polynomial fitting, moving average or Gaussian convolution, to smooth the gain curve in the boundary transition zone. The parameters of the function can be preset or adjusted according to the experimentally measured different individual skin colors and tissue optical properties, to ensure the smooth and operable gain.

[0066] By combining the second gain correction spectrum with the preset cross-region smoothing function, the spectral jump problem caused by the discontinuity of the gain boundary can be effectively solved, the spectral continuity and blood glucose characteristic signal stability are enhanced, the correction effect of the white balance technology on the skin color difference is continued, and therefore the accuracy and reliability of the non-invasive blood glucose concentration detection are improved.

[0067] In the embodiment, by combining the second gain with the preset cross-region smoothing function and acting on the calibrated spectrum to obtain the first spectrum, the discontinuity or jump problem that may occur at the gain boundary of the short-wave region and the long-wave region can be effectively solved. This step not only maintains the correction effect of the white balance correction on the spectral baseline of different skin color individuals, but also ensures the continuity and stability of the blood glucose sensitive band spectrum characteristics, thereby reducing the interference of skin color difference and local gain distortion on the detection accuracy of non-invasive blood glucose concentration, and realizing more reliable blood glucose feature extraction and detection in individuals with different skin colors and different tissue optical properties.

[0068] In the embodiment, the application to blood glucose concentration detection specifically includes: performing conventional spectral preprocessing on the first spectrum; performing feature dimension reduction or screening on the preprocessed spectral data; establishing a blood glucose concentration prediction model based on the extracted spectral features, for mapping the spectral features to a blood glucose concentration prediction value; outputting the predicted blood glucose concentration.

[0069] It should be noted that the preprocessing can include multivariate scatter correction, standard normal transformation, and first or second derivative processing methods, for eliminating spectral noise, baseline drift, and scattering interference, thereby improving the distinguishability of the blood glucose sensitive signal. This step ensures the stability and consistency of the spectral data, facilitating subsequent feature extraction. In the prior art, these spectral preprocessing methods have been widely used in near-infrared analysis and non-invasive blood glucose detection, and are operable and mature.

[0070] It should be noted that this step can use principal component analysis (PCA), partial least squares regression (PLS) load analysis, sparse feature selection, or other dimension reduction algorithms to compress high-dimensional spectral data into low-dimensional features while retaining blood glucose related information. The role of this step is to remove redundant information and noise, and to improve the stability and accuracy of the blood glucose prediction model. In the prior art, PCA and PLS have been widely used for near-infrared blood glucose detection spectral feature extraction, and have a mature theoretical and practical basis.

[0071] It should be noted that regression or machine learning models such as partial least squares regression (PLS), support vector regression (SVR), artificial neural network (NN) can be selected to map the spectral features to the predicted blood glucose concentration. The role of this step is to realize the quantitative relationship between the spectral signal and the blood glucose concentration, and to provide reliable prediction for non-invasive blood glucose monitoring. In the prior art, these models have been proven to be feasible and accurate in spectral analysis and blood glucose prediction.

[0072] It should be noted that the predicted blood glucose concentration obtained by the above modeling can be used for clinical or personal blood glucose monitoring as the final detection result. The role of this step is to convert the spectral signal into a quantifiable blood glucose concentration index, providing a direct reference for non-invasive blood glucose detection. In the existing mature technology, non-invasive blood glucose detection systems usually realize the mapping from spectrum to blood glucose through similar modeling and output steps.

[0073] In this embodiment, in order to solve the problem of baseline shift of near-infrared spectrum caused by melanin content difference of individuals with different skin colors, a white balance method based on partition gain is used to estimate and correct the gain of short-wave and long-wave regions respectively. By first calculating the average spectral intensity of each region, then comparing it with the reference spectrum to obtain the initial gain, and then through normalization and cross-region smoothing constraint processing to generate the first gain, the partition correction of white balance is realized. This step can effectively reduce the spectral baseline shift caused by melanin difference, so that the corrected spectrum of individuals with different skin colors has comparability, and ensures the stability of the blood glucose feature signal in the short-wave region, thereby providing consistent and reliable spectral input for subsequent blood glucose concentration detection. Through this method, the accuracy and applicability of non-invasive blood glucose detection in multi-skin color population can be significantly improved, while maintaining the overall continuity of the spectrum, avoiding the baseline drift problem that may occur in traditional methods under skin color difference.

[0074] In this embodiment, in order to further optimize the white balance correction effect, the first gain is subjected to low-dimensional projection constraint combined with the melanin content-spectrum shape change prior function fitted by the offline Monte Carlo photon transport library, and a regularization term is introduced to the gain of the glucose sensitive region to obtain the second gain. At the same time, the second gain is combined with the preset cross-region smoothing function to act on the calibration spectrum, and a continuous and smooth white balance corrected spectrum is output. This step not only retains the correction effect of white balance on skin color difference, but also avoids the peak shape distortion caused by the free adjustment of the first gain, ensuring the stability of the blood glucose sensitive band. Through this gain constraint optimization and cross-region smoothing processing, the present application effectively solves the local spectral mutation problem that may occur in the actual application of traditional white balance, so that the corrected spectrum not only meets the physical feasibility, but also enhances the accuracy and reliability of blood glucose concentration detection, and is suitable for non-invasive blood glucose detection scenes with multi-skin color and multi-tissue optical properties.

Claims

1. A non-invasive blood glucose concentration detection method based on near-infrared spectroscopy detection, characterized by, The method comprises the following steps: Obtaining target near-infrared spectrum data and performing noise calibration; Dividing the spectral domain band of the calibrated spectrum into a short-wave region sensitive to melanin and a long-wave region insensitive to melanin; Estimating first gains of each region based on zoned gain white balance; Constrained optimization of the first gains in combination with a prior function of melanin content-spectrum shape change fitted by an offline Monte Carlo library, and limiting the gain to a low-dimensional projection of the prior function online to obtain second gains; The second gains combine a preset cross-region smoothing function to output the white balance corrected spectrum as the first spectrum and apply it to blood glucose concentration detection.

2. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 1, characterized by, The method of obtaining target near-infrared spectrum data and performing noise calibration comprises the following steps: Irradiating the surface of the measured skin with a near-infrared light source, and receiving the reflected spectrum signal through a detector, the spectrum signal is amplified and analog-to-digital converted by an analog front-end circuit to obtain digitized spectrum data; Performing noise calibration on the digitized spectrum data, the noise calibration includes dark current correction, detector background noise suppression and system response normalization processing, thereby obtaining the target near-infrared spectrum data after noise calibration.

3. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 2, characterized by, The method of dividing the spectral domain band of the calibrated spectrum into a short-wave region sensitive to melanin and a long-wave region insensitive to melanin comprises the following steps: According to a preset band division criterion, determining the short-wave region sensitive to melanin and the long-wave region insensitive to melanin in the spectrum; According to the boundary conditions of the short-wave region and the long-wave region, segmenting the calibrated spectrum data to obtain short-wave region spectrum and long-wave region spectrum; The boundary also includes introducing a transition band, and the data in the transition band is smoothed.

4. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 3, characterized by, The method of estimating first gains of each region based on zoned gain white balance comprises the following steps: The first gain is a gain value corrected by cross-region smoothing on the basis of normalized gain, corresponding to the short-wave region and the long-wave region respectively, and is used for white balance correction of the spectrum, which comprises the following steps: Calculating the average spectrum intensity of the short-wave region spectrum and the long-wave region spectrum respectively, which is used to represent the spectrum response characteristics of each region; According to the difference between the spectrum response characteristics and a preset reference benchmark, the initial white balance gain value corresponding to the difference is calculated and analyzed, and the reference benchmark is a standard spectrum obtained in advance; Normalizing the initial white balance gain value of each region; On the basis of the normalized gain, introducing a cross-region smoothing constraint based on the continuity principle of tissue spectrum in the wavelength dimension to obtain the first gain of each region.

5. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 4, characterized by, The method of constrained optimization of the first gains in combination with a prior function of melanin content-spectrum shape change fitted by an offline Monte Carlo library, and limiting the gain to a low-dimensional projection of the prior function online to obtain second gains comprises the following steps: Using an offline Monte Carlo photon transport library to simulate the optical response of the tissue for different melanin contents, and fitting a prior function of melanin content-spectrum shape change; Projecting the first gain to the low-dimensional basis function space of the prior function as a low-dimensional projection; Through the low-dimensional projection, the online gain is constrained not to deviate from the physically feasible spectrum shape change range; Introducing a regularization term to limit the spectrum gain of the glucose sensitive region; Defining the gain after optimization of the low-dimensional projection and the regularization term as the second gain.

6. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 5, characterized by, The second gain is combined with a preset cross-region smoothing function to output a white balance corrected spectrum as a first spectrum, specifically: According to the short-wave region and long-wave region segmentation boundary of the calibrated spectrum and the continuity principle of the tissue spectrum in the wavelength dimension, a preset cross-region smoothing function is set; The second gain is combined with the cross-region smoothing function to obtain a smoothed gain curve; The smoothed gain curve is applied to the corresponding wavelength points of the calibrated spectrum data to obtain a white balance corrected spectrum, i.e., a first spectrum.

7. The noninvasive blood glucose concentration detecting method based on near-infrared spectroscopy detection according to claim 6, characterized by, The application is applied to blood glucose concentration detection, specifically: Performing conventional spectrum preprocessing on the first spectrum; Performing feature dimension reduction or screening on the preprocessed spectrum data; Based on the extracted spectrum features, a blood glucose concentration prediction model is established to map the spectrum features to a blood glucose concentration prediction value; Output the predicted blood glucose concentration.

Images