SERS (Surface Enhanced Raman Scattering) spectrum quantitative detection method and system based on interpretable stacked ensemble learning

Abstract

The invention relates to the technical field of spectral analysis and biomedical detection, and discloses an SERS (Surface Enhanced Raman Scattering) spectrum quantitative detection method and system based on interpretable stacked ensemble learning. The method comprises the following steps: acquiring SERS spectral data of serum tumor marker standard substances with different concentration gradients; performing baseline correction and normalization preprocessing on the data, performing sparse feature selection by using an LASSO algorithm, and screening out key spectral features to construct a sample data set; constructing an interpretable stacking integration model, wherein the model adopts a base learner layer and a meta learner layer; training the model by using the training set, optimizing model hyper-parameters by using a cross validation strategy, and establishing a mapping relationship between spectral features and tumor marker concentrations; and collecting SERS spectral data of a to-be-detected serum sample, extracting key spectral features, inputting the key spectral features into the trained interpretable stacked integrated model, and outputting a concentration predicted value of the tumor marker in the to-be-detected serum sample. The method has the advantages of high precision, universality and molecular level interpretability.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis and biomedical detection technology, specifically to a quantitative SERS spectral detection method and system based on interpretable stacked ensemble learning. Background Technology

[0002] Precise quantification of serum tumor markers is crucial for early cancer screening, treatment efficacy assessment, and prognostic monitoring. As measurable indicators of normal biological processes, pathological states, or treatment responses, serum biomarkers require highly accurate and reliable detection methods. Studies have shown that serum serological markers (SERS) have the potential to detect these biomarkers.

[0003] Serum biomarker detection technology has shown promise due to its high sensitivity and molecular fingerprinting capabilities. However, serum composition is complex, presenting challenges such as high background interference, spectral overlap, and matrix competitive adsorption. To address this challenge, machine learning methods are increasingly being used in spectral analysis. However, traditional single machine learning models struggle to achieve high-precision quantitative analysis of multiple biomarkers. Existing quantitative models remain limited to single or a few biomarkers, lacking the versatility for multiplex detection, thus requiring further solutions. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a quantitative detection method and system for SERS spectra based on interpretable stacked ensemble learning, which has the advantages of high precision, versatility and molecular-level interpretability.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention discloses a quantitative detection method for SERS spectra based on interpretable stacked ensemble learning, comprising: S1. Collect SERS spectral data of serum tumor marker standards at different concentration gradients; S2. Baseline correction and normalization preprocessing are performed on the SERS spectral data, and sparsity feature selection is performed using the LASSO algorithm to screen out key spectral features. A sample dataset containing key spectral features and true concentration labels is constructed and divided into training set and validation set. The key spectral features are characteristic peak positions related to the quantitative analysis of serum tumor markers, which are screened by the LASSO sparsity model. S3. Construct an interpretable stacked ensemble model, wherein the interpretable stacked ensemble model adopts a two-layer network architecture: the first layer is a base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is a meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results. S4. Train the interpretable stacked ensemble model using the training set, optimize the model hyperparameters using a cross-validation strategy, and establish a mapping relationship between spectral features and tumor marker concentrations; S5. Collect SERS spectral data of the serum sample to be tested, extract key spectral features as in step S2, input them into the trained interpretable stacked ensemble model, and output the predicted concentration values ​​of tumor markers in the serum sample to be tested.

[0006] As a further improvement to the above scheme, the base learner layer includes a support vector regression model, a partial least squares regression model, and a limit gradient boosting model. The support vector regression model is configured with a radial basis kernel function to capture the nonlinear mapping relationship between spectral features and tumor marker concentrations, thereby resolving the nonlinear features in the spectral response. The partial least squares regression model is used to extract latent variables from the spectral data, eliminate multicollinearity among features, and establish a linear response relationship between spectral intensity and marker concentration. The limit gradient boosting model utilizes its gradient boosting tree structure to mine the nonlinear interactions between each input spectral feature and achieve robust optimization for spectral outliers. The predicted concentration values ​​output by the three base learners are concatenated into a feature vector and input to the meta-learner layer. The meta-learner layer uses elastic network regression, and adaptively determines the fusion weights of the prediction results of the three machine learning regression models through hyperparameters optimized by grid search and cross-validation.

[0007] As a further improvement to the above scheme, step S4 specifically includes the training process of the interpretable stacked ensemble model, which includes: The training set is divided into K subsets for K-fold cross-validation; In each fold of training, the base learner is trained using K-1 subsets, and predictions are made on the remaining 1 subset to generate out-of-bag predictions. The out-of-bag predictions of all subsets are concatenated to form a new feature matrix with the same sample size as the original training set. The new feature matrix is ​​concatenated with the original key spectral features after feature selection, and used as input data to train the elastic network of the meta-learner layer. The hyperparameters of the elastic network are optimized using grid search to determine the optimal fusion weights for each base learner.

[0008] As a further improvement to the above scheme, in step S5, the formula for calculating the concentration prediction value is as follows: ; In the formula, The predicted concentration of tumor markers in the serum sample to be tested; The predicted concentration value output by the support vector regression model. To improve the predicted concentration values ​​output by the limiting gradient model, The predicted concentration value is output by the partial least squares regression model. For bias terms; , , The weights of each base learner are obtained by minimizing the loss function: ; In the formula, The number of samples; This represents the actual concentration. The concentration of tumor markers predicted by the base learner; Let ω be a vector containing three weights: ω = [ω1, ω2, ω3]. T The superscript T is the transpose symbol; and For hyperparameters; and Let L1 and L2 represent the norm and L2 norm, respectively.

[0009] As a further improvement to the above scheme, the detection method further includes: S6. Based on the SHAP framework, calculate the contribution of key feature bands of the serum sample to the concentration prediction value and generate a visual interpretation result.

[0010] As a further improvement to the above scheme, in step S6, the SHAP framework calculates the SHAP value of the feature using the following formula: ; In the formula, Spectral characteristics SHAP value, For a specific data point in the dataset to be explained, The set of all input spectral features. For set Exclusion features any subset, For the prediction function of the interpretable stacked ensemble model; To include spectral features The input data points of a complete feature subset; Not containing spectral features The input data points are a subset of the features; the SHAP value is used to quantitatively evaluate the impact of a single spectral feature on the model output. If the SHAP value is 0, it means that the spectral features have no effect on the prediction; If the SHAP value of a key spectral feature is positive, then the feature is determined to have a promoting effect on concentration prediction, corresponding to the specific vibrational signal of the tumor marker. If the SHAP value of a key spectral feature is negative, then the feature is determined to have an inhibitory effect on concentration prediction, corresponding to interference signals of the serum matrix or spectral overlap regions.

[0011] As a further improvement to the above scheme, in step S1, SERS spectral data is acquired using a gold nanoparticle substrate. The preparation method of this substrate includes: AuNPs colloids were synthesized using the sodium citrate reduction method; AuNPs colloid was dropped onto the surface of a silicon wafer. Under the conditions of ambient temperature of 25°C and relative humidity of 75%, the silicon wafer was first inverted and kept for a predetermined time, and then dried upright to form a film, thus forming a uniform SERS detection area.

[0012] As a further improvement to the above scheme, in step S2, the LASSO algorithm compresses the coefficients of non-critical features to zero through L1 regularization constraints, thereby screening out the critical spectral features in the full-spectrum data; wherein, the number of retained features accounts for 20%-30% of the total number of original features.

[0013] As a further improvement to the above scheme, the serum tumor markers include one or more of the following: alpha-fetoprotein, carcinoembryonic antigen, carbohydrate antigen 19-9, carbohydrate antigen 15-3, carbohydrate antigen 50, carbohydrate antigen 125, carbohydrate antigen 24-2, carbohydrate antigen 72-4, ferritin, human epididymal protein 4, neuron-specific enolase, and squamous cell carcinoma antigen.

[0014] This invention also discloses a SERS spectral quantitative detection system based on interpretable stacked ensemble learning, which applies the SERS spectral quantitative detection method based on interpretable stacked ensemble learning as described above; the detection system includes: a data acquisition module, a data processing module, a model building module, a model training module, an actual detection module, and a result display module.

[0015] The data acquisition module is used to acquire SERS spectral data of serum tumor marker standards at different concentration gradients; The data processing module is used to perform baseline correction and normalization preprocessing on SERS spectral data, and to use the LASSO algorithm for sparsity feature selection to screen out key spectral features, construct a sample dataset containing key spectral features and true concentration labels, and divide it into training set and validation set; wherein, the key spectral features are characteristic peak positions related to the quantitative analysis of serum tumor markers screened by the LASSO sparsity model. The model building module is used to build an interpretable stacked ensemble model, which adopts a two-layer network architecture: the first layer is the base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is the meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results. The model training module uses the training set to train the interpretable stacked ensemble model, adopts a cross-validation strategy to optimize the model hyperparameters, and establishes a mapping relationship between spectral features and tumor marker concentrations. The actual detection module is used to collect SERS spectral data of the serum sample to be tested. After extracting key spectral features according to step S2, it is input into the trained interpretable stacked ensemble model and outputs the predicted concentration values ​​of tumor markers in the serum sample to be tested. The results display module is used to calculate the contribution of key feature bands of the serum sample to the concentration prediction based on the SHAP framework, and generate visual interpretation results.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention discloses a quantitative detection method for serum tumor markers based on interpretable stacked ensemble learning. By stacking and integrating multiple heterogeneous machine learning regression models and using an elastic network for optimal weight fusion, the interpretable stacked ensemble model can effectively capture linear and nonlinear patterns related to the concentrations of various tumor markers in the complex serum spectrum. Experiments show that its quantitative prediction determination coefficients (R²) for 12 common tumor markers (AFP, CEA, CA19-9, etc.) all exceed 0.9, demonstrating excellent prediction accuracy and broad marker coverage.

[0017] 2. This invention uses LASSO for feature selection, which can automatically filter out hundreds of key features from thousands of spectral features. While retaining most of the effective information, it reduces the data dimensionality by about 75.3%, which significantly improves the computational efficiency of subsequent model training and prediction, and makes the model more focused on key signals, thus enhancing robustness.

[0018] 3. This invention deeply integrates the SHAP interpretability framework, which not only provides feature importance ranking but also quantifies the direction and magnitude of each Raman shift feature's contribution to a specific prediction result through SHAP values. This allows researchers and clinicians to understand which spectral features (such as specific protein peaks or interfered sugar signals) the model is based on to make judgments, transforming "black box" predictions into "white box" analysis based on molecular spectral evidence, greatly improving the scientific rigor and clinical credibility of the results.

[0019] 4. This invention establishes a complete framework from high-quality SERS data acquisition to automated, high-precision quantification and molecular-level interpretation. This framework can not only be used for early cancer screening and auxiliary diagnosis, but the biomarker-specific spectral interference patterns it reveals also provide a new spectroscopic perspective for studying biological issues such as the glycosylation state and conformational changes of tumor biomarkers, and has significant scientific research and clinical translational value. Attached Figure Description

[0020] Figure 1 This is a flowchart of the SERS spectral quantitative detection method based on interpretable stacked ensemble learning in an embodiment of the present invention.

[0021] Figure 2 The image shows a scanning electron microscope image of the assembled gold nanoparticle substrate in an embodiment of the present invention (the interpolation D=49.50±0.27 nm in the figure indicates that the average particle size is 49.50 nm and the standard deviation is ±0.27 nm).

[0022] Figure 3 The 10 samples obtained from the surface-enhanced Raman scattering substrate of gold nanoparticles in this embodiment of the invention are... -7 M CV at 1620cm -1 SERS spectrum at the location.

[0023] Figure 4 The 10 samples obtained from the surface-enhanced Raman scattering substrate of gold nanoparticles in this embodiment of the invention are... -7 M CV at 1620cm -1 The SERS spectrum intensity at the location.

[0024] Figure 5 The average spectra of 12 tumor biomarkers at their highest concentrations and the average spectra of healthy serum are shown in the embodiments of the present invention.

[0025] Figure 6 The AFP spectra of 15 concentration gradients before and after data preprocessing in this embodiment of the invention are shown. Figure 6 In this context, A represents the area before preprocessing, and B represents the area after preprocessing.

[0026] Figure 7 This is a performance evaluation of four feature selection methods (CARS, RFE, SPA, and LASSO) in this embodiment of the invention, compared with the original data without any feature selection, in terms of quantification of tumor markers based on four regression indicators. Figure 7 A in R 2 The indicators are: B for RMSE, C for CCC, and D for RPD.

[0027] Figure 8This invention provides a comparison of four models for different regression indices across all tumor biomarker quantification tasks. Figure 8 A in R 2 The indicators are: B for RMSE, C for CCC, and D for RPD.

[0028] Figure 9 The weight contribution results of the three base learners in the embodiments of the present invention after completing the prediction of tumor biomarker concentration are shown.

[0029] Figure 10 This is a regression diagram showing the prediction results of four tumor biomarkers, AFP, CEA, CA19-9, and CA125, in an embodiment of the present invention. Figure 10 In the diagram, A represents AFP, B represents CEA, C represents CA19-9, and D represents CA125.

[0030] Figure 11 This is a regression diagram showing the prediction results of four tumor biomarkers, CA50, CA15-3, CA72-4, and NSE, in an embodiment of the present invention. Figure 11 In this context, E represents CA50, F represents CA15-3, G represents CA72-4, and H represents NSE.

[0031] Figure 12 This is a regression diagram showing the prediction results of four tumor biomarkers—CA24-2, Ferritin, HE4, and SCCA—in embodiments of the present invention. Figure 12 In this context, I stands for CA24-2, J for Ferritin, K for HE4, and L for SCCA.

[0032] Figure 13 The top ten spectral features contributing to the prediction of CEA concentration in the embodiments of the present invention ( Figure 13 (A) and the distribution of SHAP values ​​of each feature in the CEA concentration prediction process ( Figure 13 (B in the middle).

[0033] Figure 14 The top ten spectral features contributing to the prediction of tumor marker CA19-9 concentration in the embodiments of the present invention are ( Figure 14 The distribution of SHAP values ​​for each feature in the CA19-9 concentration prediction process (A) and CA19-9 concentration prediction process. Figure 14 (B in the middle).

[0034] Figure 15 The top ten spectral features contributing to the prediction of AFP concentration in the embodiments of the present invention ( Figure 15 (A) and the distribution of SHAP values ​​of each feature in the AFP concentration prediction process ( Figure 15 (B in the middle).

[0035] Figure 16 The top ten spectral features contributing to the prediction of tumor marker CA72-4 concentration in the embodiments of this invention are ( Figure 16 The distribution of SHAP values ​​for each feature in the CA72-4 concentration prediction process (A) and CA72-4 concentration prediction process. Figure 16 (B in the middle). Detailed Implementation

[0036] 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, and not all embodiments. 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.

[0037] Example 1 This embodiment focuses on 12 specific tumor protein biomarkers dissolved in the serum matrix of healthy individuals, including alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), carbohydrate antigen 125 (CA125), carbohydrate antigen 50 (CA50), carbohydrate antigen 15-3 (CA15-3), carbohydrate antigen 72-4 (CA72-4), neuron-specific enolase (NSE), carbohydrate antigen 24-2 (CA24-2), ferritin, human epididymal protein 4 (HE4), and squamous cell carcinoma antigen (SCCA). This embodiment illustrates how to achieve high-precision, simultaneous quantitative analysis of multiple components using an interpretable stacked integrated model (ISEM) and the SHAP interpretation framework in complex serum matrices with severe spectral interference and the presence of multiple tumor biomarkers.

[0038] Please see Figure 1 The SERS spectral quantitative detection method based on interpretable stacked ensemble learning of the present invention includes steps S1 to S6.

[0039] S1. Collect SERS spectral data of serum tumor marker standards at different concentration gradients.

[0040] First, serum samples and SERS substrates are prepared to ensure the reliability of data acquisition.

[0041] (1) Serum sample preparation: Blood samples were collected from healthy individuals and centrifuged at 10,000 rpm for 10 minutes. The supernatant was collected as a blank serum matrix. Subsequently, high-concentration standard solutions of the above 12 tumor markers were serially diluted in serum. To cover the lower threshold to significantly elevated levels in clinical testing, 15 concentration gradients were prepared for each marker. Mechanical stirring was used in each dilution step to ensure solution homogeneity, and 12 sets of tumor marker standards at different concentrations were prepared.

[0042] (2) SERS substrate preparation: Gold nanoparticles (AuNPs) were synthesized using the sodium citrate reduction method. 198 mL of ultrapure water was mixed with 2 mL of 1 wt% chloroauric acid (HAuCl4) solution and boiled. Then, 2 mL of 1 wt% sodium citrate solution was added, and the mixture was refluxed for 40 minutes until the solution turned wine red. The synthesized AuNPs were centrifuged at 8000 rpm for 10 minutes, and the supernatant was discarded for concentration.

[0043] (3) Inverted Drying Film Formation Process: To address the spectral inhomogeneity caused by the "coffee ring effect" in traditional droplet drying, this embodiment employs a modified inverted drying strategy. Under constant ambient temperature of 25°C and relative humidity of 75%, 1 μL of concentrated AuNPs colloid is dropped onto the silicon wafer surface. The wafer is then inverted and held for 15 minutes to utilize gravity to counteract surface tension. Subsequently, it is flipped back to its upright position until completely dried and formed a film. Finally, a relatively uniform region is formed at the center of the droplet, and this region is selected for spectral measurement.

[0044] like Figure 2 As shown, the AuNPs prepared by this method have a uniform particle size distribution (average particle size of approximately 49.5 nm); Figure 3 and Figure 4 As shown, verification was performed using crystal violet (CV) probe molecules at a density of 108 × 86 μm. 2 The relative standard deviation (RSD) of the SERS signal within the region was only 15.93%, demonstrating that the substrate has excellent homogeneity and repeatability.

[0045] (4) Spectral Acquisition: 1 μL of the serum sample to be tested was dropped onto the prepared AuNPs substrate. A portable Raman spectrometer was used for detection. The excitation wavelength was selected at 785 nm to suppress the fluorescence background, the laser power was set to 100 mW, and the integration time was 10 seconds. Fifty points were randomly selected from each sample for acquisition, and the spectral scanning range was 600–1800 cm⁻¹. -1 .like Figure 5 As shown, the original SERS spectra of 12 markers at different concentrations were collected.

[0046] S2. Baseline correction and normalization preprocessing are performed on the SERS spectral data, and sparsity feature selection is performed using the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to screen out key spectral features. A sample dataset containing key spectral features and true concentration labels is constructed and divided into training set and validation set. The key spectral features are characteristic peaks related to the quantitative analysis of serum tumor markers, which are selected by the LASSO sparsity model.

[0047] The LASSO algorithm compresses the coefficients of non-critical features to zero through L1 regularization constraints, thereby screening out the critical spectral features from the full-spectrum data; wherein, the number of retained features accounts for 20%-30% of the total number of original features.

[0048] Because serum contains high concentrations of proteins, lipids, and inorganic salts, the raw spectrum contains a large amount of fluorescence background noise and non-specific signals (such as...). Figure 6 As shown in A), preprocessing and feature selection are necessary, specifically including: (1) Preprocessing: The original spectrum was baseline corrected and fluorescence background was subtracted using a linear polynomial fitting algorithm. The linear polynomial fitting algorithm has two working modes: full-spectrum fitting and specified background region fitting, which can effectively remove low-frequency background interference. Subsequently, a maximum value normalization algorithm (mapping the intensity to the 0-100 range) was used to eliminate fluctuations in absolute intensity and unify the data scale. For example... Figure 6 As shown in Figure B, the preprocessed spectral signal-to-noise ratio (SNR) increased from an average of 5.5 to 51.3, significantly enhancing the clarity of the characteristic peaks and confirming the effectiveness of data preprocessing in eliminating background interference and position-dependent variations in absolute SERS intensity. This provides a robust foundation for subsequent feature selection.

[0049] (2) LASSO feature selection: In order to remove redundant information from high-dimensional spectral data (originally containing 1201 spectral shift points), this embodiment compares four methods: competitive adaptive reweighted sampling (CARS), continuous projection algorithm (SPA), recursive feature elimination (RFE) and minimum absolute shrinkage and selection operator (LASSO).

[0050] The 9121 preprocessed SERS spectral data points were randomly divided into a training set (70%) and an independent test set (30%) in a 7:3 ratio. Figure 7 As shown, the LASSO method has a high coefficient of determination (R²). 2 It performs best in metrics such as root mean square error (RMSE).

[0051] (3) Screening Results: Utilizing the L1 regularization property of the LASSO algorithm, the coefficients of non-critical features were compressed to zero. Ultimately, the LASSO method retained an average of 296 key spectral features, achieving approximately 75.3% dimensionality reduction compared to the original 1201 features. Based on LASSO feature selection, seven biomarkers demonstrated excellent predictive performance, with R... 2 Values ​​exceeding 0.95. These retained features (points with non-zero regression coefficients) are the key spectral features that have a statistically strong correlation with tumor marker concentrations.

[0052] S3. Construct an interpretable stacked ensemble model (ISEM), which adopts a two-layer network architecture: the first layer is a base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is a meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results.

[0053] The base learner layer includes support vector regression (SVR) models, partial least squares regression (PLSR) models, and extreme gradient boosting (XGBoost) models.

[0054] The support vector regression model is configured with a radial basis function (RBF) kernel to model the nonlinear mapping between spectral features and tumor marker concentrations, in order to resolve the nonlinear features in the spectral response.

[0055] Partial least squares regression models are used to extract latent variables from spectral data to eliminate multicollinearity among spectral features and to model the linear response relationship between spectral intensity and tumor marker concentration.

[0056] The extreme gradient boosting model utilizes its gradient boosting tree structure to explore the nonlinear interactions between various input spectral features and achieve robust optimization for spectral outliers.

[0057] The output of the base learner layer is formed by concatenating the predicted concentration values ​​from the three base learners mentioned above, and is used as a new feature vector input to the meta learner layer.

[0058] The meta-learner layer uses ElasticNet regression, and the hyperparameters obtained through grid search and cross-validation are adaptively used to determine the fusion weights of the prediction results of the three machine learning regression models.

[0059] To comprehensively evaluate the superiority of the described interpretable stacked ensemble model, this embodiment systematically compares its predictive performance with three single machine learning models that constitute its base learners: SVR, XGBoost, and PLSR. Evaluation metrics include the coefficient of determination (R²). 2The parameters are: root mean square error (RMSE), consistency correlation coefficient (CCC), and performance deviation ratio (RPD).

[0060] On the overall test set containing 12 biomarkers, ISEM demonstrated significantly better predictive power than any single model. Average R0 2 The ISEM score reached 0.955, while the best-performing single model, SVR, had a score of 0.950, and PLSR only 0.892. The average RMSE for ISEM was 32.235, lower than SVR's 34.042 and PLSR's 48.276. The average CCC and RPD for ISEM were 0.977 and 5.506, respectively, both the highest values. The RPD values ​​for all 12 biomarkers exceeded 3.2, indicating that ISEM possesses excellent robustness and versatility in predicting different concentrations and biomarkers.

[0061] ISEM showed a high linear correlation between the predicted and actual concentrations of each biomarker, with the fitted curve closely approximating the ideal diagonal (slope = 1). Specific R² values ​​are as follows: Ferritin: 0.981; Squamous cell carcinoma antigen (SCCA): 0.988; Carbohydrate antigen 125 (CA125): 0.974; Carbohydrate antigen 19-9 (CA19-9): 0.971; Carcinoembryonic antigen (CEA): 0.963; Carbohydrate antigen 24-2 (CA24-2): 0.963; Neuron-specific enolase (NSE): 0.943; Carbohydrate antigen 50 (CA50): 0.939; Carbohydrate antigen 15-3 (CA15-3): 0.937; Alpha-fetoprotein (AFP): 0.906; Carbohydrate antigen 72-4 (CA72-4): 0.905; Human epididymal protein 4 antigen (HE4): 0.986. ISEM successfully achieved simultaneous, high-precision quantification of 12 tumor markers, with R values ​​for all markers. 2 >0.9, with R values ​​for 8 markers 2 >0.95.

[0062] S4. The interpretable stacked ensemble model is trained using the training set, and the model hyperparameters are optimized using a cross-validation strategy to establish a mapping relationship between spectral features and tumor marker concentrations.

[0063] In this embodiment, the training process of the interpretable stacked ensemble model adopts an OOF stacking strategy: 5-fold cross-validation is used. The training set is divided into 5 parts, and 4 parts are used in turn to train the base model. "Out-of-Fold (OOF) predictions" are generated for the remaining part. The 5 generated OOF predictions are concatenated to form a new feature matrix of the same length as the original training set (containing the preliminary prediction results of SVR, XGBoost, and PLSR), which is used as input to train the second layer of ElasticNet.

[0064] Parameter optimization: The hyperparameters (α and λ) of ElasticNet are optimized using grid search to adaptively determine the optimal weight combination for each base model.

[0065] S5. Collect SERS spectral data of the serum sample to be tested, extract key spectral features as in step S2, input them into the trained interpretable stacked ensemble model, and output the predicted concentration values ​​of tumor markers in the serum sample to be tested.

[0066] The formula for calculating the predicted concentration value is as follows: ; In the formula, The predicted concentration of tumor markers in the serum sample to be tested; The predicted concentration value output by the support vector regression model. To improve the predicted concentration values ​​output by the limiting gradient model, The predicted concentration value is output by the partial least squares regression model. For bias terms; , , The weights of each base learner are obtained by minimizing the loss function: ; In the formula, The number of samples; This represents the actual concentration. The concentration of tumor markers predicted by the base learner; Let ω be a vector containing three weights: ω = [ω1, ω2, ω3]. T The superscript T is the transpose symbol; and For hyperparameters; and Let L1 and L2 represent the norm and L2 norm, respectively.

[0067] like Figures 8 to 9As shown, experimental results demonstrate that the interpretable stacked ensemble model of this invention exhibits superior performance in the quantitative detection of multiple components: 1. Overall Accuracy: On the test set, the interpretable average prediction coefficient of determination (R²) of the stacked ensemble model for 12 markers is [value missing]. 2 The R-value is as high as 0.955, and the average performance deviation ratio (RPD) is 5.506, both significantly better than the single model (R-value of SVR). 2 =0.950, R of PLSR 2 =0.892).

[0068] 2. Specific marker results (e.g.) Figures 10 to 12 As shown): the predicted R values ​​for all 12 markers 2 All exceeded 0.9. Among them, the R of squamous cell carcinoma antigen (SCCA) was... 2 The R value of ferritin is as high as 0.988. 2 The value was 0.981, demonstrating extremely high quantitative accuracy.

[0069] 3. Model weight analysis (e.g.) Figure 9 As shown in the figure, the stacked ensemble model can be explained by its ability to adaptively allocate weights based on the spectral characteristics of different biomarkers. For example, in AFP detection, SVR was assigned 91.5% of the weight, demonstrating the advantage of the nonlinear model in handling low-concentration AFP signals; while in CEA and Ferritin detection, the weight of PLSR was increased (to 40.7% and 25.7%, respectively), indicating that these biomarkers have good linear responses within specific concentration ranges.

[0070] S6. Based on the SHAP framework, calculate the contribution of key characteristic bands of the serum sample to the concentration prediction value and generate a visual interpretation result. The purpose of this step is to transform the "black box" model into a "white box" analysis that is visible at the molecular level.

[0071] SHAP value calculation: The marginal contribution of each Raman shift feature to the prediction result is calculated using the SHAP formula. A positive SHAP value indicates that the feature promotes concentration prediction (positive correlation), while a negative value indicates inhibition (interference). The SHAP formula is: ; In the formula, Spectral characteristics SHAP value, For a specific data point in the dataset to be explained, The set of all input spectral features. For set Exclusion features any subset, For the prediction function of the interpretable stacked ensemble model; To include spectral features The input data points of a complete feature subset; Not containing spectral features The input data points are a subset of the features; the SHAP value is used to quantitatively evaluate the impact of a single spectral feature on the model output. If the SHAP value is 0, it means that the spectral features have no effect on the prediction; If the SHAP value of a key spectral feature is positive, then the feature is determined to have a promoting effect on concentration prediction, corresponding to the specific vibrational signal of the tumor marker. If the SHAP value of a key spectral feature is negative, then the feature is determined to have an inhibitory effect on concentration prediction, corresponding to interference signals of the serum matrix or spectral overlap regions.

[0072] Please see Figures 13 to 16 Color indicates feature intensity (blue for low, red for high). The CEA abstract plot shows 732, 1333, and 1299 cm⁻¹. -1 The characteristic peak intensity at 732 cm⁻¹ shows a stable positive correlation with the predicted value. -1 The stronger the Raman signal at a given location, the higher the CEA concentration predicted by the model; while at 1261 and 1400 cm⁻¹... -1 The characteristic at this point shows a negative correlation, which may represent interference signals. The most important characteristic peak in the CEA prediction is located at 732 cm⁻¹. -1 (Corresponding to the circular respiratory vibration of nucleic acids) and 1333 cm -1 1299cm -1 (Corresponding to the protein amide band). The SHAP values ​​of these peaks are all positive and relatively high, indicating that the model accurately captures the specific fingerprint of the CEA molecule, and the stable N-glycan structure of CEA makes it less susceptible to interference from the serum matrix.

[0073] like Figure 15 As shown, although the predicted R² (0.906) for AFP met the target, it was relatively low. SHAP analysis revealed the reason: the positive characteristic peak at 639 cm⁻¹ contributed the most to the AFP prediction. - ¹(L-tyrosine) significantly overlaps with the high abundance of albumin signal in serum; simultaneously, 646 cm -1 (Carbohydrate vibration) and 1477 cm -1 A significant negative SHAP value was observed, quantifying the interference of glycosylation and serum matrix on AFP signaling. This analysis demonstrates that the interpretable stacked ensemble model not only measures accurately but also explains why certain samples are difficult to predict through SHAP mapping, providing a molecular-level basis for clinical analysis.

[0074] Molecular-level interpretability analysis using SHAP successfully translated the high-dimensional, abstract mathematical calculations of the ISEM model into "spectral language" and "molecular stories" that biologists and chemists can understand. It not only validated the rationality of the model's decisions but, more importantly, diagnosed the difficulties in prediction (such as spectral overlap and glycosylation interference), achieving a complete closed loop from prediction results to prediction basis to sources of interference, greatly enhancing the scientific credibility and depth of the entire quantitative analysis framework.

[0075] Example 2 This embodiment provides a SERS spectral quantitative detection system based on interpretable stacked ensemble learning, applying the SERS spectral quantitative detection method based on interpretable stacked ensemble learning as described in Embodiment 1; the detection system includes: a data acquisition module, a data processing module, a model building module, a model training module, an actual detection module, and a result display module.

[0076] The data acquisition module is used to acquire SERS spectral data of serum tumor marker standards at different concentration gradients; The data processing module is used to perform baseline correction and normalization preprocessing on SERS spectral data, and to use the LASSO algorithm for sparsity feature selection to screen out key spectral features, construct a sample dataset containing key spectral features and true concentration labels, and divide it into training set and validation set; wherein, the key spectral features are characteristic peak positions related to the quantitative analysis of serum tumor markers screened by the LASSO sparsity model. The model building module is used to build an interpretable stacked ensemble model, which adopts a two-layer network architecture: the first layer is the base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is the meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results. The model training module uses the training set to train the interpretable stacked ensemble model, adopts a cross-validation strategy to optimize the model hyperparameters, and establishes a mapping relationship between spectral features and tumor marker concentrations. The actual detection module is used to collect SERS spectral data of the serum sample to be tested. After extracting key spectral features according to step S2, it is input into the trained interpretable stacked ensemble model and outputs the predicted concentration values ​​of tumor markers in the serum sample to be tested. The results display module is used to calculate the contribution of key feature bands of the serum sample to the concentration prediction based on the SHAP framework, and generate visual interpretation results.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A quantitative detection method for SERS spectra based on interpretable stacked ensemble learning, characterized in that, include: S1. Collect SERS spectral data of serum tumor marker standards at different concentration gradients; S2. Baseline correction and normalization preprocessing are performed on the SERS spectral data, and sparsity feature selection is performed using the LASSO algorithm to screen out key spectral features. A sample dataset containing key spectral features and true concentration labels is constructed and divided into training set and validation set. The key spectral features are characteristic peak positions related to the quantitative analysis of serum tumor markers, which are selected by the LASSO sparsity model. S3. Construct an interpretable stacked ensemble model, wherein the interpretable stacked ensemble model adopts a two-layer network architecture: the first layer is a base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is a meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results. S4. Train the interpretable stacked ensemble model using the training set, optimize the model hyperparameters using a cross-validation strategy, and establish a mapping relationship between spectral features and tumor marker concentrations; S5. Collect SERS spectral data of the serum sample to be tested, extract key spectral features as in step S2, input them into the trained interpretable stacked ensemble model, and output the predicted concentration values ​​of tumor markers in the serum sample to be tested.

2. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 1, characterized in that, The base learner layer includes a support vector regression model, a partial least squares regression model, and a limit gradient boosting model. The support vector regression model is configured with a radial basis kernel function to capture the nonlinear mapping relationship between spectral features and tumor marker concentrations, thereby resolving the nonlinear features in the spectral response. The partial least squares regression model is used to extract latent variables from the spectral data, eliminate multicollinearity among features, and establish a linear response relationship between spectral intensity and marker concentration. The limit gradient boosting model utilizes its gradient boosting tree structure to mine the nonlinear interactions between various input spectral features and achieve robust optimization for spectral outliers. The predicted concentration values ​​output by the three base learners are concatenated into a feature vector and input to the meta-learner layer. The meta-learner layer uses elastic network regression, and through grid search and cross-validation optimization of hyperparameters, adaptively determines the fusion weights of the prediction results of the three machine learning regression models.

3. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 2, characterized in that, In step S4, the training process of the interpretable stacked ensemble model specifically includes: The training set is divided into K subsets for K-fold cross-validation; In each fold of training, the base learner is trained using K-1 subsets, and predictions are made on the remaining 1 subset to generate out-of-bag predictions. The out-of-bag predictions of all subsets are concatenated to form a new feature matrix with the same sample size as the original training set. The new feature matrix is ​​concatenated with the original key spectral features after feature selection, and used as input data to train the elastic network of the meta-learner layer. The hyperparameters of the elastic network are optimized using grid search to determine the optimal fusion weights for each base learner.

4. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 2, characterized in that, In step S5, the formula for calculating the concentration prediction value is as follows: In the formula, The predicted concentration of tumor markers in the serum sample to be tested; The predicted concentration value output by the support vector regression model. To improve the predicted concentration values ​​output by the limiting gradient model, The predicted concentration value is output by the partial least squares regression model. For bias terms; , , The weights of each base learner are obtained by minimizing the loss function: In the formula, The number of samples; This represents the actual concentration. The concentration of tumor markers predicted by the base learner; Let ω be a vector containing three weights: ω = [ω1, ω2, ω3]. T The superscript T is the transpose symbol; and For hyperparameters; and Let L1 and L2 represent the norm and L2 norm, respectively.

5. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 1, characterized in that, Also includes: S6. Based on the SHAP framework, calculate the contribution of key feature bands of the serum sample to the concentration prediction value and generate a visual interpretation result.

6. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 5, characterized in that, In step S6, the SHAP framework calculates the SHAP value of the feature using the following formula: In the formula, Spectral characteristics SHAP value, For a specific data point in the dataset to be explained, The set of all input spectral features. For set Exclusion features any subset, For the prediction function of the interpretable stacked ensemble model; To include spectral features The input data points of a complete feature subset; Not containing spectral features The input data points are a subset of the features; the SHAP value is used to quantitatively evaluate the impact of a single spectral feature on the model output. If the SHAP value is 0, it means that the spectral features have no effect on the prediction; If the SHAP value of a key spectral feature is positive, then the feature is determined to have a promoting effect on concentration prediction, corresponding to the specific vibrational signal of tumor markers. If the SHAP value of a key spectral feature is negative, then the feature is determined to have an inhibitory effect on concentration prediction, corresponding to interference signals of the serum matrix or spectral overlap regions.

7. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 1, characterized in that, In step S1, SERS spectral data are acquired using a gold nanoparticle substrate. The substrate is prepared by the following method: AuNPs colloids were synthesized using the sodium citrate reduction method; AuNPs colloid was dropped onto the surface of a silicon wafer. Under the conditions of ambient temperature of 25°C and relative humidity of 75%, the silicon wafer was first inverted and kept for a predetermined time, and then dried upright to form a film, thus forming a uniform SERS detection area.

8. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 1, characterized in that, In step S2, the LASSO algorithm compresses the coefficients of non-critical features to zero through L1 regularization constraints, thereby screening out the critical spectral features from the full-spectrum data; wherein, the number of retained features accounts for 20%-30% of the total number of original features.

9. The SERS spectral quantitative detection method based on interpretable stacked ensemble learning according to claim 1, characterized in that, The serum tumor markers include one or more of the following: alpha-fetoprotein, carcinoembryonic antigen, carbohydrate antigen 19-9, carbohydrate antigen 15-3, carbohydrate antigen 50, carbohydrate antigen 125, carbohydrate antigen 24-2, carbohydrate antigen 72-4, ferritin, human epididymal protein 4, neuron-specific enolase, and squamous cell carcinoma antigen.

10. A quantitative SERS spectral detection system based on interpretable stacked ensemble learning, characterized in that, The method for quantitative detection of SERS spectra based on interpretable stacked ensemble learning as described in any one of claims 1 to 9 is applied; the detection system comprises: The data acquisition module is used to acquire SERS spectral data of serum tumor marker standards at different concentration gradients; The data processing module is used to perform baseline correction and normalization preprocessing on SERS spectral data, and to use the LASSO algorithm for sparsity feature selection to screen out key spectral features, construct a sample dataset containing key spectral features and true concentration labels, and divide it into training set and validation set; wherein, the key spectral features are characteristic peak positions related to the quantitative analysis of serum tumor markers screened by the LASSO sparsity model. The model building module is used to build an interpretable stacked ensemble model. The interpretable stacked ensemble model adopts a two-layer network architecture: the first layer is the base learner layer, which uses multiple heterogeneous machine learning regression models as base learners to generate preliminary predictions from key spectral features; the second layer is the meta-learner layer, which is used to adaptively weight and fuse the outputs of the base learner layer to generate concentration prediction results. The model training module uses the training set to train the interpretable stacked ensemble model, optimizes the model hyperparameters using a cross-validation strategy, and establishes a mapping relationship between spectral features and tumor marker concentrations. The actual detection module is used to collect SERS spectral data of the serum sample to be tested. After extracting key spectral features according to step S2, it is input into the trained interpretable stacked ensemble model and outputs the predicted concentration values ​​of tumor markers in the serum sample to be tested. The results display module is used to calculate the contribution of key feature bands of the serum sample to the concentration prediction based on the SHAP framework and generate visual interpretation results.

Images