Method and system for improving quantitative precision of whole blood hemoglobin NIRS

Through the combination of wavelet packet decomposition, fuzzy contraction and whale optimization algorithms, the problem of background and noise interference in blood near-infrared spectroscopy is solved, the accuracy of hemoglobin quantitative analysis is improved, and a new method of non-invasive rapid detection is realized.

CN118961622APending Publication Date: 2024-11-15HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410916548.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-09
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to synchronously remove background and noise interference in blood near-infrared spectroscopy, resulting in insufficient accuracy of hemoglobin quantitative analysis.

Method used

The wavelet packet decomposition technology is used to perform multi-scale decomposition of blood spectral signals, combined with fuzzy contraction denoising and whale optimization algorithm, and the optimal node combination is selected to extract the characteristic information of hemoglobin.

Benefits of technology

It significantly improves the measurement accuracy of hemoglobin NIRS quantitative analysis, can more effectively mine the spectral characteristic information of hemoglobin, avoiding information loss and noise interference in traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118961622A_ABST
    Figure CN118961622A_ABST
Patent Text Reader

Abstract

The invention provides a method and a system for improving whole blood hemoglobin NIRS quantitative precision. The method comprises the following steps: firstly, performing multi-scale decomposition on blood NIRS by using WPT to obtain wavelet packet node coefficients under different frequency scales; then, a fuzzy contraction technology is adopted to carry out adaptive denoising on wavelet packet node coefficients, and noise interference is effectively suppressed; and finally, WOA is introduced to carry out combinatorial optimization on wavelet packet nodes with different frequencies, and adaptive extraction of the hemoglobin characteristic wave band is realized by taking the prediction performance of a partial least squares (PLS) hemoglobin quantitative model as an optimization target. Experimental results show that compared with a traditional feature extraction method, the method has the advantages that spectral feature information of hemoglobin can be more effectively mined, the measurement precision of hemoglobin NIRS quantitative analysis is remarkably improved, and a new thought is provided for non-invasive rapid detection of hemoglobin.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of non-invasive hemoglobin detection, and in particular to a method and system for improving the NIRS quantitative accuracy of whole blood hemoglobin. Background Art

[0002] Hemoglobin (Hb) is an important component responsible for transporting oxygen in the blood of vertebrates and plays a key role in the occurrence of many diseases, such as cancer, anemia and heart disease. Therefore, Hb plays an important role in the field of biochemistry and biomedical research. In recent decades, the determination of Hb has mainly relied on various biological tests, such as the semi-cyanohemoglobin method and enzyme-linked immunosorbent assay. Although these methods can produce accurate results, the sample preparation process is usually time-consuming, and the blood samples cannot be recycled, resulting in a waste of blood resources. Therefore, it is urgent to develop a rapid and non-destructive method for determining Hb concentration. Near-infrared spectroscopy technology has been widely used in Hb concentration determination due to its advantages of convenience, non-destructiveness and high sensitivity. However, the characteristic information absorption of blood near-infrared spectra is weak, and it is seriously interfered by background and noise, resulting in the accuracy of the effective information quantitative model being difficult to meet the requirements. To this end, the first step in establishing a quantitative model for Hb concentration based on near-infrared blood spectra is to remove interference factors through data preprocessing.

[0003] In recent years, researchers have proposed a variety of methods to process spectral data to remove background and noise interference. Traditional methods of background correction usually use curve fitting techniques, such as least squares polynomial fitting method and polynomial fitting method to estimate the background baseline. However, these methods are easily affected by the degree of fitting and require the selection of appropriate fitting orders. In order to overcome these problems, researchers have proposed some improved methods, such as introducing penalty terms or weight factors to improve the accuracy and efficiency of baseline estimation, such as the penalized least squares method proposed by Whittaker and the adaptive loop weighted penalized least squares method proposed by Zhang. Some other chemometric methods have also been developed to remove system interferences that are not related to the substance to be measured, such as orthogonal signal processing (OSC) proposed by Wold et al. and external parameter orthogonalization (EPO) proposed by Roger, which all remove irrelevant interference information through orthogonal projection. In addition, wavelet transform can also be used to extract spectral feature information.

[0004] Noise is another key factor affecting the performance of spectral multivariate correction models. At present, there are three main strategies for noise suppression: methods based on adaptive filtering, transform domain and machine learning. Adaptive filtering methods, such as normalized least mean square (NLMS) filters and Wiener filters, can adaptively adjust filter parameters according to the statistical characteristics of signals and noise, effectively suppress noise and retain useful signals. Transform domain methods are mainly Fourier transform and wavelet transform. The principle is to convert time domain signals to frequency domain and remove high-frequency noise components by designing appropriate frequency domain filters. Empirical mode decomposition and its improved methods decompose complex signals into the sum of several intrinsic mode functions (IMFs), and achieve signal denoising by reconstructing and filtering high-frequency IMFs. In addition, machine learning methods assume that there are differences between signals and noise in certain statistical characteristics. By mining their differences in different feature spaces and combining the prior knowledge of the signal, denoising rules are adaptively learned from the data, thereby achieving effective noise suppression and signal retention. However, the preprocessing effect of this type of algorithm is heavily dependent on the training set data, and its accuracy and generalization ability need to be further improved. In general, the above methods mainly focus on single background or noise problems, and rarely consider the simultaneous removal of background and noise. However, in actual spectra, background and noise often exist simultaneously and are intertwined. Therefore, it is urgent to develop a method that can simultaneously remove background and noise and accurately extract the characteristics of the substance to be measured, so as to further improve the performance of spectral quantitative analysis. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to improve the NIRS quantitative accuracy of whole blood hemoglobin.

[0006] The present invention solves the above technical problems through the following technical means:

[0007] A method for improving the NIRS quantitative accuracy of whole blood hemoglobin, comprising the following steps:

[0008] S1. Decompose the original blood spectrum signal into multiple nodes with different frequency information by wavelet packet decomposition, and classify the nodes. The classification method is as follows:

[0009]

[0010] Among them, N j,n is the coefficient W of node [j,n] j,n The length of j is the current decomposition level, δ j,n is the noise intensity estimate, Mad j,n is the absolute median of node [j,n], when node W j,n The absolute value of j,n When , it is classified as a signal node, otherwise it is identified as a pure noise node and set to zero;

[0011] S2. Node fuzzy shrinkage denoising: Introducing fuzzy membership evaluation and quantizing the wavelet packet node coefficient W j,n With the noise intensity estimate δ j,n The similarity between them; specifically, the shrinkage factor is defined as follows:

[0012] θ(W j,n ,β j,n ,m)=1-ε(W j,n ,β j,n ,m)

[0013]

[0014] Among them, m represents the fuzzy factor, β j,n is the tolerance coefficient, dynamically adjusts the shrinkage factor, so as to better retain the potential signal components in the wavelet packet node coefficients and suppress noise; the node coefficients after fuzzy shrinkage denoising

[0015]

[0016] S3. Based on the denoised blood spectrum model in step S2, the whale optimization algorithm is used to select the optimal node combination, extract effective spectral feature information, and achieve accurate hemoglobin quantitative analysis.

[0017] Furthermore, the specific method of performing wavelet packet decomposition on the original blood spectrum signal into multiple nodes of different frequency information in step S1 is:

[0018] For the original blood spectrum signal f(t)∈L 2 (R), L 2 (R) is expressed as Hilbert space, and the wavelet decomposition coefficients can be obtained as follows:

[0019]

[0020] in, is the mother wavelet; parameter a represents the scale index, τ represents the time shift; the recursive formula of the wavelet packet decomposition and reconstruction algorithm is as follows:

[0021] Wavelet packet decomposition is performed by W l j,n Get W l j+1,2n and W l j+1,2n+1 :

[0022]

[0023] Wavelet packet reconstruction is based on W l j+1,2n and Wl j+1,2n+1 Get W l j,n :

[0024]

[0025] Among them, l, k represent the number of decomposition layers, j, n represent the node numbers, so the original spectrum signal W1 will be decomposed into node signals of different layers and different frequency spans by combining the high-pass filter h and the low-pass filter g; the width of the frequency span is:

[0026]

[0027] where f s is the sampling frequency. When the number of decomposition layers l is preset, the desired frequency span can be obtained.

[0028] Furthermore, the specific process of the whale optimization algorithm selecting the optimal node combination in step S3 is:

[0029] Step 1. Encoding: Represent each candidate solution as a binary vector, X = (x1, x2, ..., x n ), where x i ∈{0,1},i=1,2,…,n. '1' means the node is selected, while '0' means the node is not selected; the vector length n is equal to the total number of nodes;

[0030] Step 2. Fitness function: The cross-validation root mean square error (RMSECV) of the hemoglobin quantitative model is used as the fitness function f(x) = RMSECV, and the goal is to minimize it;

[0031] Step 3. Whale predation behavior simulation: (1) Surrounding the prey: where t is the current iteration, and is the coefficient vector, is the position vector, is the current optimal solution; (2) Swimming towards the prey in a spiral: Where b is a constant and l is a random number; (3) Random search: in is a random solution;

[0032] Step 4. Termination condition: When the maximum number of iterations is reached, the optimal node combination solution is output.

[0033] Furthermore, the original blood spectrum signal in step S1 comes from subjects with similar social and economic backgrounds, and the interval between blood collection and near-infrared spectrum collection does not exceed 24 hours.

[0034] The present invention also provides a system for improving the NIRS quantitative accuracy of whole blood hemoglobin, comprising the following steps:

[0035] The wavelet packet decomposition module is used to decompose the original blood spectrum signal into multiple nodes with different frequency information through wavelet packets, and classify the nodes. The classification method is as follows:

[0036]

[0037] Among them, N j,n is the coefficient W of node [j,n] j,n The length of j is the current decomposition level, δ j,n is the noise intensity estimate, Mad j,n is the absolute median of node [j,n], when node W j,n The absolute value of j,n When , it is classified as a signal node, otherwise it is identified as a pure noise node and set to zero;

[0038] Node fuzzy shrinkage denoising module, introduces fuzzy membership evaluation, quantizes wavelet packet node coefficients W j,n With the noise intensity estimate δ j,n The similarity between them; specifically, the shrinkage factor is defined as follows:

[0039] θ(W j,n ,β j,n ,m)=1-ε(W j,n ,β j,n ,m)

[0040]

[0041] Among them, m represents the fuzzy factor, β j,n is the tolerance coefficient, dynamically adjusts the shrinkage factor, so as to better retain the potential signal components in the wavelet packet node coefficients and suppress noise; the node coefficients after fuzzy shrinkage denoising

[0042]

[0043] The optimal selection module uses the whale optimization algorithm to select the optimal node combination based on the denoised blood spectrum model, extract effective spectral feature information, and achieve accurate hemoglobin quantitative analysis.

[0044] Furthermore, the specific method of performing wavelet packet decomposition on the original blood spectrum signal into multiple nodes of different frequency information in the wavelet packet decomposition module is:

[0045] For the original blood spectrum signal f(t)∈L 2(R), L 2 (R) is expressed as Hilbert space, and the wavelet decomposition coefficients can be obtained as follows:

[0046]

[0047] in, is the mother wavelet; parameter a represents the scale index, τ represents the time shift; the recursive formula of the wavelet packet decomposition and reconstruction algorithm is as follows:

[0048] Wavelet packet decomposition is performed by W l j,n Get W l j+1,2n and W l j+1,2n+1 :

[0049]

[0050] Wavelet packet reconstruction is based on W l j+1,2n and W l j+1,2n+1 Get W l j,n :

[0051]

[0052] Among them, l, k represent the number of decomposition layers, j, n represent the node numbers, so the original spectrum signal W1 will be decomposed into node signals of different layers and different frequency spans by combining the high-pass filter h and the low-pass filter g; the width of the frequency span is:

[0053]

[0054] where f s is the sampling frequency. When the number of decomposition layers l is preset, the desired frequency span can be obtained.

[0055] Furthermore, the specific process of the whale optimization algorithm in the optimization module selecting the optimal node combination is as follows:

[0056] Step 1. Encoding: Represent each candidate solution as a binary vector, X = (x1, x2, ..., x n ), where x i ∈{0,1},i=1,2,…,n. '1' means the node is selected, while '0' means the node is not selected; the vector length n is equal to the total number of nodes;

[0057] Step 2. Fitness function: The cross-validation root mean square error (RMSECV) of the hemoglobin quantitative model is used as the fitness function f(x) = RMSECV, and the goal is to minimize it;

[0058] Step 3. Whale predation behavior simulation: (1) Surrounding the prey: where t is the current iteration, and is the coefficient vector, is the position vector, is the current optimal solution; (2) Swimming towards the prey in a spiral: Where b is a constant and l is a random number; (3) Random search: in is a random solution;

[0059] Step 4. Termination condition: When the maximum number of iterations is reached, the optimal node combination solution is output.

[0060] Furthermore, the original blood spectrum signal in the wavelet packet decomposition module comes from subjects with similar social and economic backgrounds, and the interval between blood collection and near-infrared spectrum collection does not exceed 24 hours.

[0061] The present invention also provides a processing device, comprising at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above method by calling the program instructions.

[0062] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the above method.

[0063] The advantages of the present invention are:

[0064] The present invention proposes a method for extracting characteristic bands of hemoglobin based on wavelet packet transform (WPT), fuzzy shrinkage and whale optimization algorithm (WOA). First, WPT is used to perform multi-scale decomposition of blood NIRS to obtain wavelet packet node coefficients at different frequency scales; then, fuzzy shrinkage technology is used to adaptively denoise the wavelet packet node coefficients to effectively suppress noise interference; finally, WOA is introduced to perform combined optimization of wavelet packet nodes of different frequencies, and the prediction performance of the partial least squares (PLS) hemoglobin quantitative model is taken as the optimization target to achieve adaptive extraction of hemoglobin characteristic bands. Experimental results show that compared with traditional feature extraction methods, this method can more effectively mine the spectral feature information of hemoglobin, significantly improve the measurement accuracy of hemoglobin NIRS quantitative analysis, and provide a new idea for realizing non-invasive and rapid detection of hemoglobin. The specific advantages are as follows:

[0065] 1. Use WPT to perform multi-scale decomposition of blood NIRS to obtain wavelet packet node coefficients at different frequency scales, laying the foundation for subsequent feature extraction

[0066] 2. An adaptive wavelet packet node shrinkage method (FuzzyShrink) based on fuzzy analysis is proposed, which can better retain the effective signal characteristics while suppressing noise interference and avoiding the excessive shrinkage problem of traditional threshold denoising.

[0067] 3. The WOA algorithm is introduced to optimize the combination of wavelet packet nodes, and the most relevant characteristic information of hemoglobin is screened out at various scales and frequencies, which greatly improves the predictive performance of quantitative analysis.

[0068] 4. Modeling verification on an actual blood sample dataset shows that the proposed WPT-FS-WOA hybrid method can more effectively eliminate background interference and highlight the spectral characteristics of hemoglobin compared to traditional preprocessing methods, thereby establishing a quantitative analysis model with higher prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 . A schematic diagram of wavelet packet decomposition in an embodiment of the present invention;

[0070] Figure 2 This is a workflow diagram of WPT-FS-WOA in an embodiment of the present invention;

[0071] Figure 3 8 representative nodes after 6-layer wavelet packet decomposition in the embodiment of the present invention;

[0072] Figure 4 The influence of different fuzzy factors on the shrinkage of the simulated wavelet packet decomposition coefficients in the embodiment of the present invention, noise tolerance coefficient |β=3: (a) coefficient noise fuzzy similarity; (b) coefficient fuzzy shrinkage performance;

[0073] Figure 5 The original spectrum of blood in the embodiment of the present invention;

[0074] Figure 6 This is the node [6, 42] before and after fuzzy shrinkage denoising in an embodiment of the present invention, compared with the traditional threshold processing method;

[0075] Figure 7 is the change of RMSECV with the number of iterations of the hemoglobin quantitative model in the embodiment of the present invention;

[0076] Figure 8 The probability distribution diagram of different frequency nodes in the embodiment of the present invention (a), and the wavelet packet reconstruction of nodes [6, 1] and [6, 7];

[0077] Fig. 9 The original spectrum in the embodiment of the present invention (a); the optimal node combination screened by WOA (b); the modeling accuracy of different preprocessing methods (c); and the prediction effect of WPT-FS-WOA. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0079] This embodiment provides a method for improving the NIRS quantitative accuracy of whole blood hemoglobin, comprising the following steps:

[0080] Step 1: Wavelet packet decomposition and node classification

[0081] Wavelet transform (WT) is a widely used technique in signal processing and has been proven to be a powerful method for improving signal resolution. In discrete wavelet transform (DWT), the signal is decomposed into high-frequency and low-frequency components, called discrete details and discrete approximations, and then the discrete approximations are further decomposed until a preset scale is reached. Wavelet packet transform (WPT) is a development of DWT. The first step of WPT is the same as DWT, but in the next step, WPT decomposes both discrete details and discrete approximations. Therefore, a decomposition tree with two branches at each node can be obtained, such as Figure 1 As shown, more components with higher resolution than DWT are obtained. For the spectral signal f(t)∈L 2 (R), L 2 (R) is expressed as Hilbert space, and the wavelet decomposition coefficients can be obtained as follows:

[0082]

[0083] in, is the mother wavelet; parameter a represents the scale index, and τ represents the time shift. In order to extract the local characteristic information of the blood spectrum signal from finer details and improve the time-frequency resolution of the signal, this paper adopts the wavelet packet denoising method. The recursive formula of the wavelet packet decomposition and reconstruction algorithm is as follows:

[0084] Wavelet packet decomposition is performed by W l j,n Get W l j+1,2n and W l j+1,2n+1 :

[0085]

[0086] Wavelet packet reconstruction is based on W l j+1,2n and W l j+1,2n+1 Get W l j,n :

[0087]

[0088] Among them, l, k represent the number of decomposition layers, j, n represent the node numbers, so the original spectrum signal W1 will be decomposed into node signals of different layers and different frequency spans by combining the high-pass filter h and the low-pass filter g. The width of the frequency span is:

[0089]

[0090] where f s is the sampling frequency. When the decomposition layer number l is preset, the desired frequency span can be obtained. When each frequency span is fine enough, the original signal can separate the background, effective signal and noise of the blood spectrum. From the above wavelet packet transform theory, it can be seen that for the wavelet packet decomposition of the blood spectrum, it is necessary to focus on the decomposition layer number and the mother wavelet function. In the calculation process of this paper, the mother wavelet function uses the Daubechies filter (db4) with a vanishing moment of 4.

[0091] After wavelet packet decomposition, the raw NIRS of blood is decomposed into multiple nodes representing different frequency information. However, there are huge differences in the spectral information contained in different nodes. Generally, low-frequency nodes contain more effective information, while high-frequency nodes contain more noise information. In order to better identify whether the WPT node contains information components or pure noise, the nodes need to be classified in advance. The division method is derived from the node correlation threshold and the level correlation threshold, which is defined as follows:

[0092]

[0093] Among them, N j,n is the node coefficient W j,n The length of j is the current decomposition level, δ j,n is the noise intensity estimate, the size is Mad j,n / 0.6745,Mad j,n is the absolute median of node [j,n]. From formula (6), we can see that only when node W j,n The absolute value of j,n , it will be classified as a signal node. Otherwise, it will be considered a pure noise node and set to zero.

[0094] Step 2: Node fuzzy shrinkage denoising

[0095] The traditional threshold function denoising method easily leads to excessive shrinkage of effective signals, resulting in information loss. In order to overcome this problem, this paper is inspired by the data membership of fuzzy entropy and proposes an adaptive wavelet packet node shrinkage method based on fuzzy partitioning, namely fuzzy shrinkage. The core idea of ​​the fuzzy shrinkage method is to quantify the similarity between the wavelet packet node coefficients and the noise by introducing fuzzy membership evaluation. Specifically, the shrinkage factor is defined as follows:

[0096] θ(W j,n ,β j,n ,m)=1-ε(W j,n ,β j,n ,m) (7)

[0097]

[0098] Among them, W j,n represents the coefficient of the wavelet packet node [j,n], m represents the fuzzy factor, β j,n is the tolerance factor (in this paper, it is assumed to be the noise threshold in equation (6)). Therefore, equation (8) actually represents w j,n With noise β j,n The fuzzy similarity between them, and formula (7) represents W j,nis the probability of a valid signal. Compared with the traditional threshold method, fuzzy shrinkage can dynamically adjust the shrinkage factor according to the fuzzy relationship between the node coefficient and the noise, so as to better retain the potential signal components in the wavelet packet node coefficients while suppressing the noise. This adaptive shrinkage strategy based on fuzzy analysis can more accurately extract the effective components in the spectral signal and avoid the information loss caused by excessive shrinkage. Node coefficients after fuzzy shrinkage denoising

[0099]

[0100] Step 3: Whale optimization algorithm spectral feature extraction

[0101] In the above steps, we have denoised the blood spectral signal by wavelet packet decomposition and fuzzy shrinkage to obtain the denoised signal. However, in the entire wavelet packet transform (WPT) process, the contribution of different nodes to the model is not equal, and the nodes containing background noise have almost no contribution to the model. Therefore, it is necessary to select the optimal node combination and extract effective spectral feature information to achieve more accurate hemoglobin quantitative analysis. To this end, this paper uses the Whale Optimization Algorithm (WOA) to optimize the node combination. WOA is an emerging swarm intelligence algorithm that simulates the natural behavior of whale predation and has strong global search capabilities and fast convergence speed [reference]. In WOA, each candidate solution represents a combination of nodes containing specific frequencies. The goal of the algorithm is to find the optimal node combination that makes the hemoglobin quantitative model most accurate. The specific mathematical model steps are as follows:

[0102] Step 1. Encoding: Represent each candidate solution as a binary vector, X = (x1, x2, ..., x n ), where x i ∈{0,1},i=1,2,…,n. '1' means the node is selected, while '0' means the node is not selected. The length of the vector n is equal to the total number of nodes.

[0103] Step 2. Fitness function: The root mean square error of cross validation (RMSECV) of the hemoglobin quantitative model is used as the fitness function f(x) = RMSECV, and the goal is to minimize it.

[0104] Step 3. Whale predation behavior simulation: (1) Surrounding the prey: where t is the current iteration, and is the coefficient vector, is the position vector, is the current optimal solution. (2) Swimming towards the prey in a spiral: Where b is a constant and l is a random number. (3) Random search: in is a random solution.

[0105] Step 4. Termination condition: When the maximum number of iterations is reached, the optimal node combination solution is output.

[0106] Figure 2 The workflow of the proposed denoising hybrid algorithm (WPT-FS-WOA) is demonstrated.

[0107] The following is a specific example to illustrate the above method in detail:

[0108] 1. Acquisition of blood spectral dataset

[0109] The 67 blood samples in this embodiment are all from the hematology department of a cancer hospital. The peripheral blood samples (about 3 ml) of the subjects are quickly collected and stored in anticoagulation blood vessels, and then immediately transferred to a -20 ° C refrigerator for storage to maintain the chemical composition and integrity of the blood samples. All peripheral blood samples were collected by near-infrared spectroscopy on the day of collection. Among the subjects, 49.25% were male and 50.75% were female, with an average age of 63 years old. All subjects had similar socioeconomic backgrounds. This study has been approved by the hospital ethics committee, and all subjects were informed of the details of the research plan.

[0110] All blood near-infrared spectra were obtained using an MB3600 spectrometer (ABB Ltd., Canada). In order to better obtain the spectral information of hemoglobin and avoid interference from other substances, 200 μl of each blood sample was taken and 600 μl of red blood cell lysis solution was added at a ratio of 1:3. After mixing evenly, the mixture was placed in a PBS tube. During the experiment, a quartz cuvette liquid pool with an optical path length of 1 mm was used, and the spectrum collected by the empty cuvette was used as the background spectrum. The spectrum acquisition mode was transmission mode, and the instrument resolution was set to 4 cm -1 The scanning number is 16 times, and the spectrum acquisition range is controlled at 4000-11000cm -1 .

[0111] 2. Determination of the number of wavelet packet decomposition layers

[0112] The number of wavelet packet decomposition layers is a key parameter for NIRS denoising and hemoglobin information feature extraction in this paper. The choice of the number of decomposition layers directly affects the effect of feature extraction and the computational complexity, and it is necessary to find a balance between noise suppression and information retention. If the number of decomposition layers is too small, the wavelet packet transform cannot give full play to the advantages of multi-scale and multi-frequency analysis. At this time, background noise and effective signals may be mixed in the same node, making it difficult to achieve effective separation, which will seriously affect the accuracy of subsequent feature information extraction. On the other hand, if the number of decomposition layers is too large, the wavelet packet tree will become very large, introducing too much redundant information and irrelevant noise components. This not only increases the computational burden, but may also mask key features, leading to the "dimensionality disaster" problem, which is also not conducive to the extraction of key features.

[0113] This study uses 3634 blood spectrum data points. Considering the decomposition accuracy and computational efficiency, this paper uses 6-layer wavelet packet decomposition to obtain 64 different frequency nodes, each of which has 63 data points, which can well balance noise suppression and information retention. We selected several representative nodes for key analysis, and the results are shown in the figure. Among them, the lowest frequency background node [6, 0] ( Figure 3 a) is basically the same as the absorption shape of water in NIRS, indicating that this method can well separate background information from effective information. By comparing the signals of different nodes with the absorption characteristics of hemoglobin, we found that several other intermediate frequency nodes (Figure bg) can effectively extract the characteristic absorption information of the component information, confirming the effectiveness of wavelet packet decomposition in blood spectral feature extraction. At the same time, these higher frequency nodes also highlight some noise information (f), which provides important clues for subsequent node fuzzy shrinkage denoising. Therefore, the 6-layer wavelet packet decomposition shows excellent performance in background noise removal, hemoglobin feature extraction and computational efficiency, and is a good choice for NIRS hemoglobin information analysis.

[0114] 3. Reasonable selection of fuzzy factors

[0115] In the process of hemoglobin NIRS feature extraction based on wavelet packet decomposition, information node fuzzy shrinkage denoising is a key step, in which the selection of fuzzy factors determines the degree of fuzzification in the denoising process, which is crucial for suppressing noise components and retaining effective hemoglobin feature information. In order to illustrate the reasonable selection of fuzzy factors, this paper provides a simulated node coefficient example, and the noise tolerance coefficient |β is set to 3. Figure 4As shown in (a), different fuzzy factor selections will lead to different degrees of fuzzy shrinkage of node coefficients. As the fuzzy factor increases, the coefficient noise fuzzy similarity changes more slowly with the coefficient, and the fuzziness increases. As can be seen from Figure (b), the shrinkage inhibitory effect of the node coefficient on coefficients below the threshold is also enhanced. After comprehensive analysis of the two, this embodiment sets the value of the fuzzy factor m to 1.5. In this case, fuzzy shrinkage of node coefficients can better retain coefficients greater than the threshold and avoid losing valid feature information. At the same time, it also has a good inhibitory effect on coefficients less than the threshold, effectively removing the interference of noise.

[0116] 4. Results

[0117] Figure 5 The original spectrum of blood NIRS is shown. It can be clearly observed that the 6900 cm -1 、5200cm -1 and 4950cm -1 There is a broad absorption band near 6900cm -1 The broad peak near 5200cm is caused by the first harmonic absorption of the OH stretching vibration of water molecules; -1 The peaks near 4950cm correspond to the combined absorption of bending and stretching vibrations (v2+v3) of water molecules; -1 The large peaks that exceed the spectral detection limit are caused by the combined frequency absorption of the antisymmetric and bending OH of water molecules. These strong water absorption characteristic bands mask the absorption signals of the target blood components, making it difficult to directly analyze the spectral data and establish an accurate correction model. Therefore, before establishing a quantitative model, it is necessary to eliminate the spectral interference factors (including noise) in the data as much as possible to highlight the characteristic information of the target substance.

[0118] 4.1 Wavelet Packet Node Fuzzy Shrinkage Denoising of Blood Spectrum

[0119] After determining the number of wavelet packet decomposition layers and the fuzzy shrinkage factor, the blood near-infrared spectrum is decomposed by 6 layers of wavelet packets. Among them, a small number of high-frequency nodes are directly set to zero because they only contain noise, while most nodes are retained because they contain valid information. However, there is still a lot of noise interference in the retained nodes, which requires fuzzy shrinkage denoising. Taking the 42nd node [6, 42] of the 6th layer as a representative example, and comparing it with the traditional hard and soft threshold denoising methods, the results before and after denoising are shown in the figure. Figure 5 shown.

[0120] from Figure 5It can be clearly seen that although the traditional threshold denoising can remove noise to a certain extent, there are still many residues. After fuzzy shrinkage processing, the effective signal features are retained, and the noise is basically completely removed. This is mainly because the denoising performance of the traditional threshold processing method depends largely on the selection of the threshold. It is usually necessary to manually adjust the threshold according to the characteristics of the signal and the noise level, which increases the complexity and uncertainty of the denoising process. In contrast, the fuzzy shrinkage denoising method can adaptively adjust the shrinkage intensity through fuzzy rules, reducing the sensitivity of threshold selection. In addition, compared with the traditional threshold method, fuzzy shrinkage denoising can shrink the coefficients continuously, maintaining the continuity of the processed coefficients, and effectively avoiding the additional resonance in the processed data caused by the discontinuous coefficient thresholding of hard threshold denoising, while the soft threshold denoising brings information loss due to the excessive shrinkage of the signal correlation coefficient, which helps to reduce the pseudo-Gibbs phenomenon in the reconstructed signal.

[0121] 4.2WOA hemoglobin feature extraction

[0122] After wavelet packet decomposition and node fuzzy shrinkage preprocessing, a series of wavelet packet node coefficients were obtained from the near-infrared spectrum of blood. Due to the inherent multi-scale decomposition characteristics of WPT, the node coefficients of these different frequency bands contain the characteristic information of different components in the original NIRS, and therefore there are significant differences. Some nodes may be rich in characteristic information of hemoglobin, while other nodes may be mainly composed of noise or redundant information. In order to select the subset that contributes most to the quantitative model of hemoglobin concentration from many nodes and highlight the characteristic information related to hemoglobin, this study used inverse wavelet packet transform (IWPT) to reconstruct the different frequency nodes after fuzzy shrinkage denoising, and selected 4100-4920cm, which is richer in protein information. -1 The bands are analyzed in depth. On this basis, the whale optimization algorithm (WOA) is introduced to iteratively optimize the node combination. In the iterative process, the minimum root mean square error (RMSECV) of the partial least squares (PLS) hemoglobin content model of the cross-validation set is used as the objective function of the current iteration.

[0123] 4.2.1 Performance of Whale Iterations

[0124] In the WOA algorithm, the number of iterations t is a key parameter that affects the performance of the optimal node combination search. Models with too small iterations have poor prediction accuracy because they do not find the optimal variable combination. Too many iterations will increase the complexity of the model and make the calculation time-consuming. Therefore, it is necessary to find a suitable t value. This paper studies the establishment of a hemoglobin quantitative model for each t when t changes from 1 to 240, and calculates the current fitness function RMSECV, as shown in Figure 6As shown in the figure. When t is small, RMSECV is relatively large and unstable. As t increases, RMSECV decreases rapidly and tends to be stable. When t is in the range of 0-100, as t increases, RMSECV decreases and the fluctuation range is large, indicating that the differences between individual whales are large at the beginning, and then the individual differences gradually decrease. When t exceeds 50, RMSECV is almost stable, indicating that the differences between individual whales are weak at this time, and a good convergence effect is achieved. Considering the prediction ability, t is set to 200 in the subsequent calculations.

[0125] 4.2.2 Spectral feature extraction results

[0126] The WOA algorithm is used to optimize different wavelet packet node combination schemes. In order to reduce the randomness of the calculation results, we run the algorithm 500 times, and each run will give a node combination candidate solution. By counting the frequency of each node being selected, the relative contribution of the node to the model can be evaluated. Figure 8 As shown in (a), some important nodes, such as [6, 1], [6, 2], [6, 3], [6, 4], [6, 5], etc., are significantly more likely to be selected by WOA than other nodes, indicating that these nodes have the same meaning in explaining the model and contribute much more than other nodes. At the same time, these high-probability nodes are mainly distributed in the low-frequency and medium-frequency bands, indicating that WPT can effectively divide the spectral information into multiple scales, and WOA can filter out nodes with stronger correlation with hemoglobin, while the lowest-frequency background and high-frequency noise nodes are discarded because they contain a lot of irrelevant redundant information, achieving the effect of removing background and further denoising.

[0127] In order to further demonstrate the feature extraction effect of wavelet packet fuzzy shrinkage denoising combined with WOA algorithm, Figure 8 (b) and (c) show the wavelet packet reconstruction images of low-frequency nodes [6, 1] and medium-frequency nodes [6, 7] with higher selection frequencies. Table 1 also allocates the bands of the extracted characteristic spectra. Figure 8 As shown in Table 1, the high-probability nodes selected by WOA contain obvious hemoglobin characteristic information [ ], such as 4110, 4260, 4.365, 4915cm -1 , indicating the effectiveness of the WOA algorithm. In addition, these nodes also have 4810cm -1 、4281cm -1 The characteristic information related to urea and glucose is shown in Figure 2, which indicates that after WPT decomposition, the absorption frequencies of these blood components are similar to those of hemoglobin and the characteristic peak positions overlap. They also exist in the nodes and are therefore selected.

[0128] TableⅠ.Near-infrared spectral band attribution of hemoglobin in wholeblood

[0129]

[0130]

[0131] 4.3 Model Performance Evaluation

[0132] In order to verify the effectiveness and robustness of the hybrid method (WPT-FS-WOA), the blood dataset was modeled and the hemoglobin content was used as the prediction component. First, the KS (Kennard-Stone) algorithm was used to divide the spectral dataset into a calibration set (54 samples) and a prediction set (13 samples), with a ratio of 8:2. Then, the hybrid method proposed in this paper was used to process the original spectrum to remove the interference information of noise and background, and the partial least squares (PLS) regression model was established after extracting the characteristic information of hemoglobin. During the modeling process, in order to prevent overfitting and underfitting of the model, the 5-fold cross-validation method was used to select the number of latent variables (LV) of the PLS, and the cross-validation root mean square error (RMSEC) of the calibration set was used. V ) and the root mean square error of the prediction set (RMSEP) are used to evaluate the performance of the correction model. The specific formula is as follows:

[0133]

[0134] Where n is the number of samples in the cross-calibration set or prediction set, y is the true value of the sample, and y i is the predicted value of the model. A good model usually has a smaller root mean square error and a larger coefficient of determination. In order to better evaluate the performance of the WPT-FS-WOA hybrid algorithm, a model is built based on the original spectrum and the currently popular preprocessing methods for removing the background and highlighting the spectral feature information (i.e., derivative method (Der), continuous wavelet transform (CWT), orthogonal signal correction (OSC), and adaptive iterative reweighted penalized least squares method (airPLS)) to compare the results. The results are shown in Table 2 and Figure 8 shown.

[0135] contrast Fig. 9 The original blood spectrum in (a) shows the spectral data after being processed by the WPT-FS-WOA hybrid algorithm. Fig. 9 (b), its absorption peak shape is sharper and more prominent than the original spectrum, which effectively highlights the characteristic absorption information of hemoglobin and filters out most of the noise and background interference. Fig. 9(c) The modeling effects of several common preprocessing methods are compared. It can be seen that the RMSEP value of the WPT-FS-WOA method is the smallest, and its prediction performance is significantly better than other methods. Fig. 9 (d) The calibration curve and prediction results of the WPT-FS-WOA-PLS model are also given. The calibration curve basically coincides with the diagonal line, and the correlation coefficient R is above 0.99, which further confirms that the model has excellent quantitative analysis capabilities and can accurately predict the hemoglobin content of unknown samples.

[0136] TableⅠ.Modeling effects of different preprocessing methods

[0137]

[0138]

[0139] 5 Conclusion

[0140] This example aims to improve the accuracy of NIRS quantitative analysis of hemoglobin in whole blood by using a hybrid method based on wavelet packet transform (WPT), fuzzy shrinkage and whale optimization algorithm (WOA) to solve the background noise interference problem in hemoglobin quantitative analysis of blood near infrared spectroscopy (NIRS). The main conclusions are as follows:

[0141] 1. Use WPT to perform multi-scale decomposition of blood NIRS to obtain wavelet packet node coefficients at different frequency scales, laying the foundation for subsequent feature extraction

[0142] 2. An adaptive wavelet packet node shrinkage method (FuzzyShrink) based on fuzzy analysis is proposed, which can better retain the effective signal characteristics while suppressing noise interference and avoiding the excessive shrinkage problem of traditional threshold denoising.

[0143] 3. The WOA algorithm is introduced to optimize the combination of wavelet packet nodes, and the most relevant characteristic information of hemoglobin is screened out at various scales and frequencies, which greatly improves the predictive performance of quantitative analysis.

[0144] 4. Modeling verification on an actual blood sample dataset shows that the proposed WPT-FS-WOA hybrid method can more effectively eliminate background interference and highlight the spectral characteristics of hemoglobin compared to traditional preprocessing methods, thereby establishing a quantitative analysis model with higher prediction accuracy.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for improving the NIRS quantitative accuracy of whole blood hemoglobin, characterized in that: The following steps are involved: S1. Decompose the original blood spectrum signal into multiple nodes with different frequency information by wavelet packet decomposition, and classify the nodes. The classification method is as follows: Among them, N j,n is the coefficient W of node [j,n] j,n The length of j is the current decomposition level, δ j,n is the noise intensity estimate, Mad j,n is the absolute median of node [j,n], when node W j,n The absolute value of j,n When , it is classified as a signal node, otherwise it is identified as a pure noise node and set to zero; S2. Node fuzzy shrinkage denoising: Introducing fuzzy membership evaluation and quantizing the wavelet packet node coefficient W j,n With the noise intensity estimate δ j,n The similarity between them; specifically, the shrinkage factor is defined as follows: θ(W j,n ,b j,n ,m)=1-ε(W j,n ,b j,n ,m) Among them, m represents the fuzzy factor, β j,n is the tolerance coefficient, dynamically adjusts the shrinkage factor, so as to better retain the potential signal components in the wavelet packet node coefficients and suppress noise; the node coefficients after fuzzy shrinkage denoising S3. Based on the denoised blood spectrum model in step S2, the whale optimization algorithm is used to select the optimal node combination, extract effective spectral feature information, and achieve accurate hemoglobin quantitative analysis.

2. A method for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 1, characterized in that: The specific method of performing wavelet packet decomposition of the original blood spectrum signal into multiple nodes of different frequency information in step S1 is: For the original blood spectrum signal f(t)∈L 2 (R), L 2 (R) is expressed as Hilbert space, and the wavelet decomposition coefficients can be obtained as follows: in, is the mother wavelet; parameter a represents the scale index, τ represents the time shift; the recursive formula of the wavelet packet decomposition and reconstruction algorithm is as follows: Wavelet packet decomposition is performed by W l j,n Get W l j+1,2n and W l j+1,2n+1 : Wavelet packet reconstruction is based on W l j+1,2n and W l j+1,2n+1 Get W l j,n : Among them, l, k represent the number of decomposition layers, j, n represent the node numbers, so the original spectrum signal W1 will be decomposed into node signals of different layers and different frequency spans by combining the high-pass filter h and the low-pass filter g; the width of the frequency span is: where f s is the sampling frequency. When the number of decomposition layers l is preset, the desired frequency span can be obtained.

3. A method for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 1, characterized in that: The specific process of the whale optimization algorithm selecting the optimal node combination in step S3 is: Step 1. Encoding: Represent each candidate solution as a binary vector, X = (x1, x2, ..., x n ), where x i ∈{0,1},i=1,2,…,n. '1' means the node is selected, while '0' means the node is not selected; the vector length n is equal to the total number of nodes; Step 2. Fitness function: The cross-validation root mean square error (RMSECV) of the hemoglobin quantitative model is used as the fitness function f(x) = RMSECV, and the goal is to minimize it; Step 3. Whale predation behavior simulation: (1) Surrounding the prey: where t is the current iteration, and is the coefficient vector, is the position vector, is the current optimal solution; (2) Swimming towards the prey in a spiral: Where b is a constant and l is a random number; (3) Random search: in is a random solution; Step 4. Termination condition: When the maximum number of iterations is reached, the optimal node combination solution is output.

4. A method for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 1, characterized in that: The original blood spectrum signal in step S1 comes from subjects with similar social and economic backgrounds, and the interval between blood collection and near-infrared spectrum collection does not exceed 24 hours.

5. A system for improving the NIRS quantitative accuracy of whole blood hemoglobin, characterized in that: The following steps are involved: The wavelet packet decomposition module is used to decompose the original blood spectrum signal into multiple nodes with different frequency information through wavelet packets, and classify the nodes. The classification method is as follows: Among them, N j,n is the coefficient W of node [j,n] j,n The length of j is the current decomposition level, δ j,n is the noise intensity estimate, Mad j,n is the absolute median of node [j,n], when node W j,n The absolute value of j,n When , it is classified as a signal node, otherwise it is identified as a pure noise node and set to zero; Node fuzzy shrinkage denoising module, introduces fuzzy membership evaluation, quantizes wavelet packet node coefficients W j,n With the noise intensity estimate δ j,n The similarity between them; specifically, the shrinkage factor is defined as follows: θ(W j,n ,b j,n ,m)=1-ε(W j,n ,b j,n ,m) Among them, m represents the fuzzy factor, β j,n is the tolerance coefficient, dynamically adjusts the shrinkage factor, so as to better retain the potential signal components in the wavelet packet node coefficients and suppress noise; the node coefficients after fuzzy shrinkage denoising The optimal selection module uses the whale optimization algorithm to select the optimal node combination based on the denoised blood spectrum model, extract effective spectral feature information, and achieve accurate hemoglobin quantitative analysis.

6. A system for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 5, characterized in that: The specific method of performing wavelet packet decomposition on the original blood spectrum signal into multiple nodes with different frequency information in the wavelet packet decomposition module is: For the original blood spectrum signal f(t)∈L 2 (R), L 2 (R) is expressed as Hilbert space, and the wavelet decomposition coefficients can be obtained as follows: in, is the mother wavelet; parameter a represents the scale index, τ represents the time shift; the recursive formula of the wavelet packet decomposition and reconstruction algorithm is as follows: Wavelet packet decomposition is performed by W l j,n Get W l j+1,2n and W l j+1,2n+1 : Wavelet packet reconstruction is based on W l j+1,2n and W l j+1,2n+1 Get W l j,n : Among them, l, k represent the number of decomposition layers, j, n represent the node numbers, so the original spectrum signal W1 will be decomposed into node signals of different layers and different frequency spans by combining the high-pass filter h and the low-pass filter g; the width of the frequency span is: where f s is the sampling frequency. When the number of decomposition layers l is preset, the desired frequency span can be obtained.

7. A system for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 1, characterized in that: The specific process of the whale optimization algorithm in the optimization module to select the optimal node combination is as follows: Step 1. Encoding: Represent each candidate solution as a binary vector, X = (x1, x2, ..., x n ), where x i ∈{0,1},i=1,2,…,n. '1' means the node is selected, while '0' means the node is not selected; the vector length n is equal to the total number of nodes; Step 2. Fitness function: The cross-validation root mean square error (RMSECV) of the hemoglobin quantitative model is used as the fitness function f(x) = RMSECV, and the goal is to minimize it; Step 3. Whale predation behavior simulation: (1) Surrounding the prey: where t is the current iteration, and is the coefficient vector, is the position vector, is the current optimal solution; (2) Swimming towards the prey in a spiral: Where b is a constant and l is a random number; (3) Random search: in is a random solution; Step 4. Termination condition: When the maximum number of iterations is reached, the optimal node combination solution is output.

8. A system for improving the NIRS quantitative accuracy of whole blood hemoglobin according to claim 1, characterized in that: The original blood spectrum signal in the wavelet packet decomposition module comes from subjects with similar social and economic backgrounds, and the interval between blood collection and near-infrared spectrum collection does not exceed 24 hours.

9. A processing device, characterized in that: It comprises at least one processor and at least one memory in communication with the processor, wherein the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute any method according to claims 1 to 4.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the method according to any one of claims 1 to 4.