A blood spectrum classification method and system fusing sers and machine learning

Abstract

A blood spectrum classification method and system fusing SERS and machine learning, mixing a human peripheral blood sample with a SERS enhancement reagent to obtain a to-be-tested mixed solution; irradiating a SERS substrate by a laser to generate a local photo-thermal effect, so as to perform a photo-thermal driven micro-convection enrichment treatment on the to-be-tested mixed solution; acquiring original spectrum data of the to-be-tested mixed solution in a preset wave number range by a Raman spectrum acquisition device; preprocessing the original spectrum data to obtain a standardized spectrum; the preprocessing includes adaptive window Savitzky-Golay filtering processing, adaptive baseline correction processing based on wavelet transform, and normalization processing; inputting the standardized spectrum into a principal component analysis model to extract a principal component score vector of the spectrum; inputting the principal component score vector into a linear discriminant analysis model to obtain a class identification of the human peripheral blood sample; the class identification includes a first spectrum class, a second spectrum class and a third spectrum class.

Description

Technical Field

[0001] This invention relates to the field of biomedical detection technology, and in particular to a blood spectral classification method and system that integrates SERS and machine learning. Background Technology

[0002] Raman spectroscopy, based on the properties of molecular vibrations, enables non-destructive detection of changes in the molecular composition of biological samples. Surface-enhanced Raman scattering (SERS) technology can significantly improve detection sensitivity, laying the technological foundation for the analysis of complex samples such as blood.

[0003] However, applying Raman spectroscopy to blood sample analysis still faces numerous technical challenges. Blood is a complex substance containing various biomolecules, whose Raman spectral signals overlap, resulting in significant background interference. Existing techniques lack standardized preprocessing methods for blood samples, making it difficult to effectively remove noise interference and retain characteristic peak information, leading to poor stability of analytical results.

[0004] Furthermore, in terms of data processing, traditional feature extraction and classification methods do not adequately consider the specificity of blood Raman spectra. Conventional algorithms struggle to accurately identify key features when processing blood spectra with complex backgrounds, affecting the accuracy and stability of classification.

[0005] In summary, existing technologies have limitations in effectively handling complex background interference, resulting in low accuracy and poor stability of blood sample analysis results. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a blood spectral classification method that integrates SERS and machine learning, comprising the following steps: Human peripheral blood samples are mixed with SERS enhancement reagent to obtain the test mixture; A local photothermal effect is generated by irradiating the SERS substrate with a laser to perform photothermal-driven micro-convection enrichment processing on the test mixture. The target molecules in the human peripheral blood sample are enriched to the laser hot spot area through thermocapillary convection. The raw spectral data of the mixture to be tested within a preset wavenumber range are obtained using a Raman spectroscopy acquisition device. The original spectral data is preprocessed to obtain a standardized spectrum; the preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The standardized spectrum is input into the principal component analysis model to extract the principal component score vector of the spectrum; The principal component score vector is input into the linear discriminant analysis model to obtain the category identifier of the human peripheral blood sample; the category identifier includes a first spectral category, a second spectral category, and a third spectral category.

[0007] Optionally, the SERS enhancement reagent comprises the SERS substrate and PBS buffer; The SERS substrate uses silver nanoparticles with a particle size of 50 nm or silver-coated gold core-shell structured nanoparticles, with a nanoparticle concentration of 3.5–5 mg / ml. The pH value of the PBS buffer is 7.0-7.2.

[0008] Optionally, the volume ratio of the human peripheral blood sample to the SERS enhancement reagent is 1:2.

[0009] Optionally, the laser is a 785nm light source semiconductor laser; In the photothermal-driven microconvection enrichment process, the temperature field distribution satisfies the heat conduction equation, and the target molecule concentration distribution satisfies the convection-diffusion equation.

[0010] Optionally, the preset wavenumber range is 300. up to 1800 .

[0011] Optionally, the adaptive window Savitzky-Golay filtering process includes the following steps: Based on the original spectral data, its local kurtosis index is calculated; Based on the kurtosis index, an adaptive filter window half-width is dynamically calculated and assigned to each spectral data point; wherein the filter window half-width is negatively correlated with the kurtosis index; For each spectral data point, within its corresponding adaptive filtering window, Savitzky-Golay polynomial fitting is performed on the original spectral data, and the fitted value corresponding to that spectral data point in the fitting result is used as the filtered spectral signal of that spectral data point.

[0012] Optionally, the adaptive baseline correction process based on wavelet transform includes the following steps: The filtered spectral signal is decomposed into multiple detail coefficient components including high-frequency information and approximate coefficient components containing low-frequency baseline information by performing discrete wavelet transform. Calculate the total variation of the filtered spectral signal; Based on the total variation, the wavelet basis function and the number of wavelet decomposition layers are adaptively selected. Set the detailed coefficient components to zero, and then perform an inverse wavelet transform on the approximation coefficient components to reconstruct the baseline signal; Based on the filtered spectral signal and the baseline signal, the baseline-corrected spectral signal is obtained.

[0013] Optionally, the normalization process specifically involves obtaining a standardized spectrum based on the baseline-corrected spectral signal through vector normalization or area normalization.

[0014] Optionally, the linear discriminant analysis model classifies the category identifiers based on the output value of the discriminant function; When the output value of the discrimination function is less than the first preset value, it corresponds to the first spectral category; When the output value of the discrimination function is greater than or equal to the first preset value, but less than or equal to the second preset value, it corresponds to the second spectral category; When the output value of the discrimination function is greater than the second preset value, it corresponds to the third spectral category.

[0015] Corresponding to the blood spectral classification method that integrates SERS and machine learning, the present invention provides a blood spectral classification system that integrates SERS and machine learning, comprising: The sample processing unit is used to mix human peripheral blood samples with SERS enhancement reagents to obtain the test mixture. The spectral acquisition unit is used to generate a local photothermal effect by irradiating the SERS substrate with a laser to perform photothermal-driven micro-convection enrichment treatment on the test mixture, enriching the target molecules in the human peripheral blood sample to the laser hot spot area through thermocapillary convection; and to acquire the raw spectral data of the test mixture in a preset wavenumber range through a Raman spectral acquisition device. The preprocessing unit is used to preprocess the original spectral data to obtain a standardized spectrum; the preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The data processing unit is used to input the standardized spectrum into the principal component analysis model to extract the principal component score vector of the spectrum; The classification unit is used to input the principal component score vector into the linear discriminant analysis model to obtain the category identifier of the human peripheral blood sample; the category identifier includes a first spectral category, a second spectral category, and a third spectral category.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By using photothermal-driven microconvection enrichment processing, the laser-driven local photothermal effect significantly enhances the intensity of characteristic molecule signals, effectively solving the problem of weak and easily interfered target molecule signals caused by the complexity of blood components. A standardized preprocessing procedure is adopted, which greatly improves the spectral quality while retaining characteristic peak information. Furthermore, principal component analysis (PCA) is used to reduce the dimensionality and extract the principal component score vector, eliminate spectral collinearity and compress redundant information, and then linear discriminant analysis (LDA) is used to construct a discriminant function for classification. This fully explores the category-specific characteristics of blood Raman spectra and solves the problem of poor classification accuracy and stability caused by insufficient consideration of the specificity of blood spectra in traditional algorithms. It can be seen that the present invention achieves high accuracy, high stability and high sensitivity of spectral classification of blood samples under complex background interference.

[0017] (2) By using SERS enhancement reagent, a stable and uniformly dispersed SERS enhancement medium can be formed in human peripheral blood samples, thereby obtaining a SERS spectrum with stable enhancement effect and high signal reproducibility, avoiding signal fluctuation and detection inaccuracy caused by unstable enhancement medium.

[0018] (3) The use of a 785nm light source semiconductor laser can minimize the fluorescence background interference of biomolecules in blood samples while ensuring SERS excitation efficiency. During the photothermal driven microconvection enrichment process, the temperature field distribution satisfies the heat conduction equation and the target molecule concentration distribution satisfies the convection-diffusion equation, realizing the quantitative control of the physical field of the thermocapillary convection enrichment process, ensuring the high consistency of enrichment efficiency among different samples, thereby significantly improving the repeatability and quantifiability of detection.

[0019] (4) By limiting the acquisition wavenumber range to 300 up to 1800 This range covers the characteristic vibrational modes of key biomolecules such as proteins, nucleic acids, lipids, and carbohydrates in blood samples, while avoiding Rayleigh scattering interference in the low wavenumber region and weak signal regions in the high wavenumber region. This enables the acquisition of the most discriminative spectral information with the most concise data, improving data processing efficiency and classification accuracy.

[0020] (5) Adaptive window Savitzky-Golay filtering dynamically allocates the half-width of the filter window that is negatively correlated with kurtosis by calculating the local kurtosis index. It can automatically expand the window in the noise-dominant region to enhance the smoothing effect, while automatically shrinking the window in the characteristic peak region to protect the peak shape and peak position information. It achieves an intelligent balance between noise suppression and information preservation, which significantly improves the fidelity and signal-to-noise ratio of the spectral signal compared with fixed window filtering.

[0021] (6) Adaptive baseline correction processing based on wavelet transform: By calculating the total spectral variation, the wavelet basis function and the number of decomposition layers are adaptively selected. The flexibility of baseline fitting can be dynamically adjusted according to the signal complexity. After setting the detail coefficients to zero, the baseline is reconstructed, which effectively removes the complex fluorescence background and nonlinear baseline drift in blood samples, while preserving the sharp Raman characteristic peaks. This significantly improves the baseline flatness and the identification of characteristic peaks in the spectrum.

[0022] (7) By using vector normalization or area normalization, the variation in overall signal intensity caused by external factors such as sample volume error, laser power fluctuation, and focusing depth difference is effectively eliminated, thereby significantly enhancing the robustness of the classification model to non-target factor interference and its generalization ability across samples.

[0023] (8) The linear discriminant analysis model divides the three spectral categories based on the output value of the discriminant function. By setting the first and second preset values ​​as objective thresholds, it realizes the quantitative determination and clear attribution of sample classification, ensuring the objectivity, repeatability and traceability of the classification results, and improving the clinical application feasibility of the classification standard. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a simplified flowchart of an embodiment of the blood spectral classification method integrating SERS and machine learning of the present invention; Figure 2 This is an optical path diagram of a Raman spectroscopy acquisition device according to an embodiment of the blood spectral classification method integrating SERS and machine learning of the present invention. Figure 3 This is a two-dimensional principal component analysis distribution diagram of an embodiment of the blood spectral classification method integrating SERS and machine learning of the present invention; Figure 4 This is a three-dimensional distribution map of principal component analysis of an embodiment of the blood spectral classification method integrating SERS and machine learning of the present invention; Figure 5 This is a framework diagram of an embodiment of the blood spectral classification system that integrates SERS and machine learning according to the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention can be used for blood spectral detection in cases of myocardial infarction, and the following description uses this application scenario as a specific example to illustrate the invention. However, this invention can also be used for blood spectral detection in cases of atherosclerosis and diabetic complications. The following embodiments are merely illustrative and do not constitute an improper limitation of this invention.

[0027] like Figure 1 As shown, the present invention provides a blood spectral classification method that integrates SERS and machine learning, which includes the following steps: Human peripheral blood samples are mixed with SERS enhancement reagent to obtain the test mixture; A local photothermal effect is generated by irradiating the SERS substrate with a laser to perform photothermal-driven micro-convection enrichment of the test mixture. The target molecules in the human peripheral blood sample are enriched to the laser hot spot area through thermocapillary convection. The raw spectral data of the mixture to be tested within a preset wavenumber range are obtained using a Raman spectroscopy acquisition device. The raw spectral data is preprocessed to obtain a standardized spectrum. The preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The standardized spectrum is input into the principal component analysis model to extract the principal component score vector of the spectrum; The principal component score vector is input into the linear discriminant analysis model to obtain the category identifier of human peripheral blood samples; the category identifier includes the first spectral category, the second spectral category, and the third spectral category.

[0028] In other embodiments of the present invention, in addition to LDA, the classification algorithm can also be implemented by support vector machine (SVM), random forest or neural network model, which can be implemented by existing technology and will not be described in detail here.

[0029] Preferably, the human peripheral blood sample is whole blood (approximately 0.5–1 μL). For example, a disposable sterile lancet is used to puncture the fingertip, and the blood is collected naturally using a micropipette or a dedicated blood collection card. The human peripheral blood sample should be tested within 5 minutes of collection using the blood spectral classification method integrating SERS and machine learning described in this invention to avoid spectral drift caused by red blood cell rupture or metabolite degradation. However, it should be noted that the method and system described in this invention only detect and classify the test mixture of isolated human peripheral blood samples and SERS enhancement reagent, and are not directly applied to living human bodies.

[0030] This invention utilizes photothermal-driven microconvection enrichment processing, leveraging the localized photothermal effect of lasers to drive thermocapillary convection, significantly enhancing the signal intensity of characteristic molecules. This effectively addresses the problem of weak target molecule signals and susceptibility to interference caused by the complexity of blood components. A standardized preprocessing workflow is employed, greatly improving spectral quality while preserving characteristic peak information. Furthermore, principal component analysis (PCA) is used for dimensionality reduction to extract principal component score vectors, eliminating spectral collinearity and compressing redundant information. Linear discriminant analysis (LDA) is then used to construct a discriminant function for classification, fully exploiting the category-specific characteristics of blood Raman spectra. This solves the problem of poor classification accuracy and stability caused by insufficient consideration of blood spectral specificity in traditional algorithms. Therefore, this invention achieves high accuracy, high stability, and high sensitivity spectral classification of blood samples under complex background interference.

[0031] In this embodiment, the SERS enhancement reagent includes a SERS substrate and a PBS buffer; the pH of the PBS buffer is 7.0-7.2, which is used to maintain the physiological environment.

[0032] The SERS substrate uses silver nanoparticles (Au Nanostars) with a particle size of 50 nm or silver-coated gold core-shell structure (Ag@Au) nanoparticles at a concentration of 3.5-5 mg / ml (based on silver). Its surface localized plasmon resonance (LSPR) peak is located in the 633–785 nm laser excitation wavelength range, which can significantly enhance the Raman signal (enhancement factor up to 10). 6 -10 8 ).

[0033] Preferably, the volume ratio of human peripheral blood sample to SERS enhancement reagent is 1:2. The mixture is stirred by ultrasonic vibration and left to stand at room temperature for 1 minute to allow the target molecules to be fully adsorbed onto the surface of the nanoparticles.

[0034] This invention utilizes SERS enhancement reagents to form a stable and uniformly dispersed SERS enhancement medium in human peripheral blood samples, thereby obtaining SERS spectra with stable enhancement effects and high signal reproducibility, avoiding signal fluctuations and detection inaccuracies caused by unstable enhancement media.

[0035] In this embodiment, the instrument configuration of the Raman spectroscopy acquisition device is as follows: Excitation source: The laser is a 785nm semiconductor laser (100mW power, to avoid fluorescence interference and sample light damage); Spectrometer: S11639CCD detector, spectral resolution ≤8 ; Sampling mode: A focused Raman system is used to focus on the central region of the droplet to avoid edge effects; Integration time: 3 seconds / point. Nine replicate points are collected for each sample and the average value is taken to improve the signal-to-noise ratio.

[0036] Output data format: Raw spectrum is wavenumber (unit: – A two-dimensional array of Raman intensities (in au), typically ranging from 200 to 100 wavenumbers. Up to 3500 ; Please refer to Figure 2 This invention uses a dedicated Raman spectroscopy optical path system to acquire the raw spectral data of the mixture to be tested. The specific process is as follows: Excitation and enrichment: A 785nm semiconductor laser 4 is used as the excitation source. After being collimated by the first lens 2, the laser beam is reflected by the dichroic mirror 3 and focused onto the detection chip 1 containing the test mixture. The local photothermal effect generated by the laser irradiation of the SERS substrate drives photothermal micro-convection in the test mixture, and through thermocapillary effect, the target molecules in the blood (such as proteins, nucleic acids, lipids, and carbohydrates) are efficiently enriched to the hot spot area where the laser spot is located.

[0037] Scattered light collection and filtering: The laser is focused onto the detection chip 1 to induce the molecules to generate Raman scattered light. After passing through the dichroic mirror 3 (which effectively separates the excitation light and the scattered light), the Raman scattered light passes sequentially through the second lens 6 (for convergence) and a long-pass filter 5 (to filter out strong Rayleigh scattered light and retain the effective Raman signal).

[0038] Dispersion and Detection: The filtered and purified Raman signal, after its optical path is adjusted by the first reflecting mirror 7 and the second reflecting mirror 9, is incident on the grating 8 for dispersion, separating light of different wavelengths in space. The dispersed light is then focused by the cylindrical lens 10, finally forming a clear Raman spectral image on the sensor 11 (such as a CCD detector). The sensor 11 converts the optical signal into an electrical signal, with an output wavenumber range of 300. up to 1800 The original spectral data within.

[0039] The principle of photothermal microvortex enrichment is as follows: by irradiating the SERS substrate with a laser, a local photothermal effect is generated, which triggers micro-scale convection and "pumps" the surrounding molecules into the hot spot area; that is, the target molecules in human peripheral blood samples are enriched into the laser hot spot area through thermocapillary convection. In the photothermal-driven microconvection enrichment process, the temperature field distribution satisfies the heat conduction equation, and the target molecule concentration distribution satisfies the convection-diffusion equation.

[0040] Specifically, the temperature field is described by the heat conduction equation, as follows: In the formula, For fluid density, For isobaric specific heat capacity, For temperature, Thermal conductivity, This refers to the photothermal absorption power per unit volume.

[0041] in, In the formula, The laser angular frequency, The vacuum permittivity, The effective dielectric function of the nanostructure. Let represent the local electric field strength, and Im denote the mathematical operation: taking the imaginary part of the complex number.

[0042] The velocity field of the marangoni flow induced by the temperature gradient satisfies the following equation: In the formula, For the fluid velocity field, For pressure, For dynamic viscosity, It is the acceleration due to gravity. The coefficient of thermal expansion is... This is the ambient reference temperature.

[0043] The convection-diffusion equation for the target molecule is as follows: In the formula, For the target molecule concentration, is the molecular diffusion coefficient.

[0044] Under steady state, the concentration enhancement factor in the hotspot region can be estimated as: In the formula, As a concentration-enhancing factor in hotspot areas, The characteristic flow rate driven by the thermocapillary effect, The enrichment time.

[0045] This invention employs a 785nm semiconductor laser source, which can minimize the fluorescence background interference of biomolecules in blood samples while ensuring SERS excitation efficiency. During the photothermal-driven microconvection enrichment process, the temperature field distribution satisfies the heat conduction equation, and the target molecule concentration distribution satisfies the convection-diffusion equation. This enables quantitative control of the physical field of the thermocapillary convection enrichment process, ensuring a high degree of consistency in enrichment efficiency among different samples, thereby significantly improving the repeatability and quantifiability of the detection.

[0046] In this embodiment, the preset wavenumber range is 300. up to 1800 This range encompasses key vibrational modes of biomolecules.

[0047] Specifically, 500 Up to 800 C–S and C–C skeletal vibrations (including cysteine, glutathione, etc.); 900 Up to 1100 C–C and C–O stretching vibrations (carbohydrates, phosphorylated metabolites); 1200 Up to 1500 Amide III, CH2 / CH3 bending vibration (proteins, lipids); 1600 Up to 1700 C=O stretching, amide I band (characteristic peaks of myocardial infarction-related inflammatory factors and troponin degradation products).

[0048] Exclusion zone: <300 (Due to instrument noise, the values ​​have been reduced) >1800 (The signal is weak and easily affected by water peaks, so it has been cropped).

[0049] This invention limits the acquisition wavenumber range to 300. up to 1800 This range covers the characteristic vibrational modes of key biomolecules such as proteins, nucleic acids, lipids, and carbohydrates in blood samples, while avoiding Rayleigh scattering interference in the low wavenumber region and weak signal regions in the high wavenumber region. This enables the acquisition of the most discriminative spectral information with the most concise data, improving data processing efficiency and classification accuracy.

[0050] Traditional Savitzky-Golay (SG) filtering uses a fixed window length and polynomial order to smooth the entire spectrum, making it difficult to balance peak shape preservation with filtering effects far from the peaks. To preserve characteristic peak shapes, this invention proposes a local window adaptive strategy: using a smaller window in the spectral peak region to avoid over-smoothing, and using a larger window in flat or background regions to enhance denoising capabilities.

[0051] In this embodiment, the adaptive window Savitzky-Golay filtering process includes the following steps: Based on the original spectral data, the local kurtosis index is calculated; specifically, the first derivative is obtained through preliminary smoothing (such as a fixed window SG). Calculate its second derivative. Define local kurtosis index : In the formula, N is the total number of spectral data points, and q is the index of the spectral data point. ={q The range r, ..., q+r} refers to a local neighborhood centered on the spectral data point q (e.g., r=5), where r is the radius of the neighborhood. = A constant used to prevent division by zero; Based on the kurtosis index, an adaptive filter window half-width is dynamically calculated and assigned to each spectral data point; the filter window half-width is negatively correlated with the kurtosis index; the filter window half-width is specifically calculated according to the following formula: In the formula, The filter window half-width, The minimum half-window for the peak region, The maximum half-window size for the background area is denoted by α, which is a window decay control parameter used to control the window decay rate, and α > 0; preferably, =3, =12; α=5; The final local window length is an odd number; In the formula, This represents the total length of the local filtering window; For each spectral data point, within its corresponding adaptive filtering window, a Savitzky-Golay polynomial fit is performed on the original spectral data; that is, for each spectral data point q, a window [k] is used to fit the original spectral data using a Savitzky-Golay polynomial. ,k+ Fit a d-th order polynomial (usually d=2) to the data within the range, and solve the following least squares problem: In the formula, For the set of real numbers, This is the vector of polynomial fitting coefficients. The order of the Savitzky–Golay polynomial fitting is given. These are the polynomial fitting coefficients; The fitted value corresponding to the spectral data point in the fitting result is taken as the filtered spectral signal of that spectral data point; the filtered spectral signal of that spectral data point is the constant term estimate, i.e.: .

[0052] The adaptive window Savitzky-Golay filtering process of this invention dynamically allocates the half-width of the filtering window that is negatively correlated with kurtosis by calculating the local kurtosis index. It can automatically expand the window in the noise-dominant region to enhance the smoothing effect, while automatically shrinking the window in the characteristic peak region to protect the peak shape and peak position information. This achieves an intelligent balance between noise suppression and information preservation, and significantly improves the fidelity and signal-to-noise ratio of the spectral signal compared with fixed window filtering.

[0053] To overcome the problem that the global parameters in the Asymmetric Least Squares (ALS) method are difficult to adapt to different spectral backgrounds, this invention uses wavelet transform to achieve data-driven baseline estimation.

[0054] In this embodiment, the adaptive baseline correction process based on wavelet transform includes the following steps: Spectral signal after filtering Perform j-level discrete wavelet transform: Among them, W j ∈ , which is the detail coefficient (high-frequency component) of the j-th layer. ∈ For the first Layer approximation coefficients (low-frequency components, including baseline information); for the filtered spectral signal It is decomposed into multiple detail coefficient components that include high-frequency information and approximate coefficient components that include low-frequency baseline information; Calculate the total variation of the filtered spectral signal. The calculation formula is as follows: ; Based on the total variation, the wavelet basis function ψ and the number of wavelet decomposition levels J are adaptively selected: In the formula, The minimum baseline feature scale;

[0055] In the formula, It is a Symlets wavelet, order 4 (an approximately symmetrical orthogonal wavelet suitable for signal smoothing and baseline separation). It is a Daubechies wavelet of order 6 (with tight support and high vanishing moments, suitable for capturing spectral details). The wavelet is a Coiflets wavelet of order 5 (approximately symmetric, simultaneously optimizing the vanishing moments of the scaling and wavelet functions, suitable for baseline estimation); θ1 and θ2 are the first and second thresholds of total variation, respectively, preferably θ1=50, θ2=150, L min =64; Set the detail coefficient components to zero, retain only the J-th level approximation coefficients for wavelet reconstruction, and then reconstruct the baseline signal using the inverse wavelet transform of the approximation coefficient components. (k): ;in, It is the inverse discrete wavelet transform operator; Based on the filtered spectral signal With baseline signal The baseline-corrected spectral signal is obtained. , .

[0056] The adaptive baseline correction processing based on wavelet transform of this invention adaptively selects the wavelet basis function and the number of decomposition layers by calculating the total spectral variation. It can dynamically adjust the flexibility of baseline fitting according to the signal complexity. After reconstructing the baseline by setting the detail coefficients to zero, it effectively removes the complex fluorescence background and nonlinear baseline drift in blood samples, while completely preserving the sharp Raman characteristic peaks, thereby significantly improving the baseline flatness and the identification of characteristic peaks in the spectrum.

[0057] In this embodiment, the normalization process specifically involves: based on the baseline-corrected spectral signal, vector normalization or area normalization is performed to obtain a standardized spectrum with approximately 1500 dimensions (corresponding to 300–1800). Step size 1 ).

[0058] This invention effectively eliminates the variation in overall signal intensity caused by external factors such as sample volume error, laser power fluctuation, and focusing depth difference through vector normalization or area normalization, thereby significantly enhancing the robustness of the classification model to interference from non-target factors and its generalization ability across samples.

[0059] In this embodiment, principal component analysis (PCA) is used to extract spectral principal components and screen data (the specific implementation can be achieved through existing technologies, which will not be elaborated here, but only briefly described, and does not constitute an improper limitation of the present invention). Data filtering: False positives and false negatives are identified based on the first two and third principal components. The two-dimensional and three-dimensional images generated by the first two and third principal components of the PCA analysis are used to find false positives mixed with positives and false negatives mixed with negatives. Figure 3 , Figure 4 As shown, Figure 3 Principal component analysis two-dimensional distribution plot, Figure 4 Principal component analysis three-dimensional distribution plot; Dimensionality reduction: Construct a spectral matrix for all samples in the training set (including healthy controls, myocardial infarction patients, and high-risk groups), calculate the covariance matrix, and extract the first v principal components (PCs). Typically, v = 10–30 can explain >95% of the total variance. Output: Each sample is mapped to a v-dimensional principal component score vector, which serves as the input feature for subsequent classifiers.

[0060] A spectral matrix is ​​constructed for all samples in the training set (including healthy controls, patients with myocardial infarction, and high-risk groups), the covariance matrix is ​​calculated, and the first v principal components (PCs) are extracted. Typically, v = 10–30 can explain >95% of the total variance. Output: Each sample is mapped to a v-dimensional principal component score vector, which serves as the input feature for subsequent classifiers.

[0061] In this embodiment, the linear discriminant analysis model classifies categories based on the output value of the discriminant function; When the output value of the discrimination function is less than the first preset value, it corresponds to the first spectral category; When the output value of the discrimination function is greater than or equal to the first preset value, but less than or equal to the second preset value, it corresponds to the second spectral category; When the output value of the discrimination function is greater than the second preset value, it corresponds to the third spectral category.

[0062] Preferably, classification and identification are performed using a Linear Discriminant Analysis (LDA) model. Specifically, the model is constructed as follows: The LDA classifier is trained on the PCA-reduced data using a labeled training set (labels: low risk, medium risk, and high risk, corresponding to the three levels of myocardial infarction risk, respectively). In this embodiment, the first spectral category is high risk, the second spectral category is medium risk, and the third spectral category is low risk. LDA constructs the optimal linear discriminant hyperplane by maximizing inter-class divergence and minimizing intra-class divergence; The model performance was evaluated using leave-one-out cross-validation (LOOCV) or 5-fold cross-validation.

[0063] Classification threshold setting: Based on clinical guidelines and ROC curve analysis, the boundary point of the discriminant function output value is set; preferably, the first preset value is -0.8 and the second preset value is 0.5.

[0064] The linear discriminant analysis model of this invention divides three spectral categories based on the output value of the discriminant function. By setting a first preset value and a second preset value as objective thresholds, it realizes the quantitative determination and clear attribution of sample classification, ensuring the objectivity, repeatability and traceability of classification results, and improving the clinical application feasibility of classification standards.

[0065] like Figure 5 As shown, the present invention also provides a blood spectral classification system that integrates SERS and machine learning, comprising: The sample processing unit 10 is used to mix human peripheral blood samples with SERS enhancement reagents to obtain a test mixture. The spectral acquisition unit 20 is used to generate a local photothermal effect by irradiating the SERS substrate with a laser, so as to perform photothermal driven micro-convection enrichment treatment on the mixture to be tested, and enrich the target molecules in the human peripheral blood sample to the laser hot spot area through thermocapillary convection; and to acquire the original spectral data of the mixture to be tested in a preset wavenumber range through a Raman spectral acquisition device. The preprocessing unit 30 is used to preprocess the raw spectral data to obtain a standardized spectrum. The preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The data processing unit 40 is used to input the standardized spectrum into the principal component analysis model to extract the principal component score vector of the spectrum; Classification unit 50 is used to input the principal component score vector into the linear discriminant analysis model to obtain the category identifier of human peripheral blood samples; the category identifier includes a first spectral category, a second spectral category, and a third spectral category.

[0066] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0067] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0068] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A blood spectral classification method integrating SERS and machine learning, characterized in that, Includes the following steps: Human peripheral blood samples are mixed with SERS enhancement reagent to obtain the test mixture; A local photothermal effect is generated by irradiating the SERS substrate with a laser to perform photothermal-driven micro-convection enrichment processing on the test mixture. The target molecules in the human peripheral blood sample are enriched to the laser hot spot area through thermocapillary convection. The raw spectral data of the mixture to be tested within a preset wavenumber range are obtained using a Raman spectroscopy acquisition device. The original spectral data is preprocessed to obtain a standardized spectrum; the preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The standardized spectrum is input into the principal component analysis model to extract the principal component score vector of the spectrum; The principal component score vector is input into the linear discriminant analysis model to obtain the category identifier of the human peripheral blood sample; the category identifier includes a first spectral category, a second spectral category, and a third spectral category.

2. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The SERS enhancement reagent comprises the SERS substrate and PBS buffer; The SERS substrate uses silver nanoparticles with a particle size of 50 nm or silver-coated gold core-shell structured nanoparticles, with a nanoparticle concentration of 3.5-5 mg / ml. The pH value of the PBS buffer is 7.0-7.

2.

3. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The volume ratio of the human peripheral blood sample to the SERS enhancement reagent was 1:

2.

4. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The laser is a 785nm light source semiconductor laser; In the photothermal-driven microconvection enrichment process, the temperature field distribution satisfies the heat conduction equation, and the target molecule concentration distribution satisfies the convection-diffusion equation.

5. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The preset wavenumber range is 300. up to 1800 .

6. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The adaptive window Savitzky-Golay filtering process includes the following steps: Based on the original spectral data, its local kurtosis index is calculated; Based on the kurtosis index, an adaptive filter window half-width is dynamically calculated and assigned to each spectral data point; wherein the filter window half-width is negatively correlated with the kurtosis index; For each spectral data point, within its corresponding adaptive filtering window, Savitzky-Golay polynomial fitting is performed on the original spectral data, and the fitted value corresponding to that spectral data point in the fitting result is used as the filtered spectral signal of that spectral data point.

7. The blood spectral classification method integrating SERS and machine learning according to claim 6, characterized in that, The adaptive baseline correction process based on wavelet transform includes the following steps: The filtered spectral signal is decomposed into multiple detail coefficient components including high-frequency information and approximate coefficient components containing low-frequency baseline information by performing discrete wavelet transform. Calculate the total variation of the filtered spectral signal; Based on the total variation, the wavelet basis function and the number of wavelet decomposition layers are adaptively selected. Set the detailed coefficient components to zero, and then perform an inverse wavelet transform on the approximation coefficient components to reconstruct the baseline signal; Based on the filtered spectral signal and the baseline signal, the baseline-corrected spectral signal is obtained.

8. The blood spectral classification method integrating SERS and machine learning according to claim 7, characterized in that, The normalization process specifically involves obtaining a standardized spectrum based on the baseline-corrected spectral signal through vector normalization or area normalization.

9. The blood spectral classification method integrating SERS and machine learning according to claim 1, characterized in that, The linear discriminant analysis model classifies the category identifiers based on the output value of the discriminant function. When the output value of the discrimination function is less than the first preset value, it corresponds to the first spectral category; When the output value of the discrimination function is greater than or equal to the first preset value, but less than or equal to the second preset value, it corresponds to the second spectral category; When the output value of the discrimination function is greater than the second preset value, it corresponds to the third spectral category.

10. A blood spectral classification system integrating SERS and machine learning, characterized in that, include: The sample processing unit is used to mix human peripheral blood samples with SERS enhancement reagents to obtain the test mixture. The spectral acquisition unit is used to generate a local photothermal effect by irradiating the SERS substrate with a laser to perform photothermal-driven micro-convection enrichment processing on the mixture to be tested, and to enrich the target molecules in the human peripheral blood sample to the laser hot spot area through thermocapillary convection. And the raw spectral data of the mixture to be tested within a preset wavenumber range are obtained by a Raman spectroscopy acquisition device; The preprocessing unit is used to preprocess the original spectral data to obtain a standardized spectrum; the preprocessing includes adaptive window Savitzky-Golay filtering, adaptive baseline correction based on wavelet transform, and normalization. The data processing unit is used to input the standardized spectrum into the principal component analysis model to extract the principal component score vector of the spectrum; The classification unit is used to input the principal component score vector into the linear discriminant analysis model to obtain the category identifier of the human peripheral blood sample; the category identifier includes a first spectral category, a second spectral category, and a third spectral category.

Images