A method and apparatus for the fractionation of urine for protein detection

Abstract

The present application relates to the technical field of urine protein detection, and provides a grading method and device for urine protein detection. The method comprises collecting a urine sample, classifying and labeling the sample; based on the sample classification and labeling, staining the urine sample, preparing a sample image, and constructing a sample dataset; pre-processing the sample dataset to obtain a training set, a validation set and a test set; extracting features from the pre-processed sample data, pre-processing the extracted features, and selecting the 58 most important features; training an enhanced hierarchical adaptive multi-kernel fusion support vector machine (E-HAMKF-SVM) to obtain a final multi-kernel fusion model; and using the multi-kernel fusion model to obtain a graded prediction category. The method improves the accuracy, generalization ability and clinical applicability of urine protein detection.

Description

Technical Field

[0001] This invention relates to the field of urine protein detection technology, and in particular to a grading method and apparatus for urine protein detection. Background Technology

[0002] Proteinuria is a key clinical manifestation of kidney damage, appearing in the early stages of the disease and serving as an independent risk factor for the progression of various primary and secondary glomerular diseases, such as diabetic nephropathy. Qualitative and quantitative detection of proteins in urine has irreplaceable clinical value for the early screening, diagnosis, pathological classification, treatment evaluation, and prognosis of kidney diseases.

[0003] Currently, commonly used clinical methods for detecting urine protein mainly include routine urinalysis and quantitative urine protein measurement, both of which have limitations to varying degrees. Routine urinalysis detects urine protein using urine dry chemistry test strips or instruments, providing a semi-quantitative result and unable to offer precise concentration values. 24-hour urine protein quantification, using a biochemical analyzer to measure total protein content, is the gold standard for assessing proteinuria, but requires strict urine collection and high patient compliance. Both methods rely on specialized laboratory equipment and personnel, failing to meet the needs of immediate home-based guidance. Existing home-based urine protein testing technologies primarily rely on urine test strips, which are semi-quantitative. Results interpretation is easily affected by physicochemical factors in urine and is highly subjective, leading to insufficient accuracy and repeatability. This gap between clinical precision and home convenience constitutes the core limitation of the current technological system, severely hindering the early detection and effective management of kidney diseases.

[0004] Existing computer-aided diagnostic methods mainly employ single-core support vector machines (SVM) or traditional machine learning algorithms, which suffer from problems such as limited accuracy, high sensitivity to class imbalance, difficulty in hyperparameter optimization, and imperfect ensemble decision-making mechanisms. Summary of the Invention

[0005] In view of this, the present invention provides a grading method and apparatus for urine protein detection, which improves the accuracy, generalization ability and clinical applicability of urine protein detection.

[0006] In a first aspect, the present invention provides a grading method for urine protein detection, the method comprising:

[0007] Step 1: Collect urine samples, classify and label the samples;

[0008] Step 2: Based on sample classification and annotation, stain the urine samples, prepare sample images, and construct the sample dataset;

[0009] Step 3: Preprocess the sample dataset to obtain the training set, validation set, and test set;

[0010] Step 4: Extract features from the preprocessed sample data, and preprocess the extracted features to select the 58 most important features;

[0011] Step 5: Train the enhanced hierarchical adaptive multi-core fusion support vector machine E-HAMKF-SVM to obtain the final multi-core fusion model;

[0012] Step 6: Use a multi-core fusion model to obtain hierarchical prediction categories.

[0013] Optionally, step 1 includes:

[0014] Step 11: Collect urine samples: Collect midstream urine samples with different concentrations of urinary protein; no demographic restrictions are set for the urine samples to ensure the diversity of sample sources;

[0015] Step 12, Sample Classification and Labeling: Based on the clinical routine urine test results, the samples were divided into 6 different categories according to the range of urine protein concentration and assigned corresponding category labels; the category label was C0 when the urine protein concentration was <0.1g / L; C1 when the urine protein concentration was 0.1g to 0.2g / L; C2 when the urine protein concentration was 0.2g to 1.0g / L; C3 when the urine protein concentration was 1.0g to 2.0g / L; C4 when the urine protein concentration was 2.0g to 4.0g / L; and C5 when the urine protein concentration was >4.0g / L.

[0016] Optionally, step 2 includes:

[0017] Step 21, Staining: Add Coomassie Brilliant Blue staining solution to the labeled urine sample at a volume ratio of 5:1, and mix thoroughly using a vortex mixer to form a homogeneous staining mixture.

[0018] Step 22, Sample Image Preparation: Use a micropipette to transfer 2 μL of the staining mixture and drop it onto the surface of the nitrocellulose membrane. Let it stand at room temperature for 1 minute to form a circular imprint. Define the circular imprint as a sample image.

[0019] Repeat the above sampling process for all classified and labeled samples to construct the sample dataset.

[0020] Optionally, step 3 includes:

[0021] Step 31, Data Cleaning: Delete sample images that are out of focus, have failed to drop samples, or are contaminated during shooting;

[0022] Step 32, Data Standardization: Adjust all sample images to 256×256 pixels, normalize pixel intensity values ​​to the range of [0,1], and perform color space standardization BGR, HSV, LAB;

[0023] Step 33: Dataset partitioning: Divide the sample dataset into training set, validation set and test set in a ratio of 70:15:15.

[0024] Optionally, the feature extraction of the preprocessed sample data in step 4 includes:

[0025] An enhanced feature extractor is used to extract a 115-dimensional feature vector from each sample image, including 72-dimensional color space features, 15-dimensional region features, 16-dimensional texture features, and 12-dimensional morphological features, denoted as . ;

[0026] Step 41, 72-dimensional color space features include 24-dimensional BGR color space features, 24-dimensional HSV color space features, and 24-dimensional LAB color space features; (Note:) For the BGR color space, the channel identifier is: For the HSV color space, For the LAB color space, For each color space Each channel Extract 8 statistics:

[0027] ;

[0028] in, For the corresponding color space channels The mean of N, where N is the total number of pixels in the current channel; Let be the pixel value of the i-th pixel, which ranges from 0 to 255; Standard deviation, The qth percentile is mathematically defined as the value of all pixel values ​​sorted in ascending order. To ensure that the data in q% is less than or equal to its value; and These are the minimum and maximum values ​​of the pixel values, respectively.

[0029] Therefore, color feature vector The dimension is 3 color spaces × 3 channels × 8 statistics = 72 dimensions;

[0030] Step 42, 15-dimensional region features include 3D central region color mean features, 3D edge region color mean features, 3D background region color mean features, 3D center-background relative difference features, and 3D edge-background relative difference features;

[0031] Let the center of the circle be... , radius is , Define the central region for the coordinates of any point in the sample image. Edge areas and background area Their expressions are as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] Calculate the color mean feature of the central region Color mean characteristics of edge regions and background area color mean characteristics Their expressions are as follows:

[0036] ;

[0037] ;

[0038] ;

[0039] in, The total number of pixels in the central region. It is the sum of all pixel values ​​within the central region; This represents the total number of pixels in the edge region. It is the sum of all pixel values ​​within the edge region; This represents the total number of pixels in the background area. This is the sum of all pixel values ​​within the background area;

[0040] Center-background relative difference features The expression is:

[0041] ;

[0042] Edge-background relative difference features The expression is:

[0043] ;

[0044] in This is used to avoid the denominator being zero;

[0045] Therefore, the eigenvector of the 15-dimensional region is: ;

[0046] Step 43: The 16-dimensional texture features include 4-dimensional Sobel gradient features, 1-dimensional Canny edge features, 3-dimensional Laplacian features, 4-dimensional circular region gradient features, and 4-dimensional color gradient features.

[0047] Sobel gradient eigenvectors The expression is:

[0048] ;

[0049] in, , , , Gradient magnitude Mean, standard deviation, maximum value, 90th percentile; gradient magnitude Defined as , The gradient is in the horizontal direction. The gradient is in the vertical direction. Represents the image matrix, which is the set of pixel values ​​for the entire sample image;

[0050] The Canny edge feature vector expression is:

[0051] ;

[0052] in, The number of edge pixels. This represents the total number of pixels in the current channel.

[0053] The Laplacian eigenvector expression is: ;

[0054] in, , and They are respectively The mean, standard deviation, and maximum value; The expression is: ,in Let I be the second derivative of the image matrix I along the x-direction. Let be the second derivative of the image matrix I along the y-direction;

[0055] The expression for the gradient feature vector of the circular region is:

[0056] ;

[0057] in, , , and The gradient sets within the circular regions are respectively. The mean, standard deviation, maximum value, and 90th percentile; The expression is:

[0058] ;

[0059] in, As the central area, For point Gradient magnitude at;

[0060] The expression for the color gradient feature vector is:

[0061] ;

[0062] in, , , and These represent the gradients of the BGR channels;

[0063] Therefore, the 16-dimensional texture feature vector is: ;

[0064] Step 44: 12-dimensional morphological features include 2-dimensional erosion features, 2-dimensional dilation features, 2-dimensional opening operation features, 2-dimensional closing operation features, and 4-dimensional morphological difference features.

[0065] Morphological eigenvectors are denoted as:

[0066] ;

[0067] in, The average brightness after erosion is N, and N is the total number of pixels in the current channel. The standard deviation after erosion represents the degree of fluctuation in the brightness of the eroded image. and These are the average brightness and standard deviation after dilation, respectively; and These are the average brightness and the standard deviation after the opening operation, respectively. and These are the average brightness and standard deviation after the closing operation, respectively; The difference between the original and corroded layers is used to reflect the boundary thickness of protein precipitation. The difference between the expanded and original regions is used to reflect the protein signal intensity in the surrounding area; The difference between the original and opening operations is used to reflect the noise intensity in the image; The difference between the closing operation and the original value is represented by the pixel value added during the closing operation.

[0068] The expressions for erosion, dilation, opening, and closing operations are as follows:

[0069] Corrosion operation for:

[0070] ;

[0071] in, Let B be the image matrix and B be the structuring element, which is:

[0072] ;

[0073] expansion operation for:

[0074] ;

[0075] Opening operation is ;

[0076] Closing operation is .

[0077] Optionally, in step 4, the extracted features are preprocessed, and the 58 most important features are selected, including:

[0078] The extracted features are preprocessed by standardizing them with StandardScaler to scale them to zero mean and unit variance; then, they are filtered by VarianceThreshold to remove features with variance of 0 or close to 0; finally, the 58 most important features are selected by SelectKBest.

[0079] The standard scaler expression is:

[0080] ;

[0081] in, This is the standardized feature vector, used to transform features to a uniform scale; The original feature vector, and These are the mean vector and standard deviation vector of the training set;

[0082] The expression for VarianceThreshold variance filtering is:

[0083] ;

[0084] in, This is the feature vector after variance filtering, used to remove features with variance of 0; Let be the variance of the i-th feature;

[0085] SelectKBest selects the 58 most important features:

[0086] ;

[0087] in, This is the final 58-dimensional feature vector selected for model training.

[0088] Optionally, step 5 includes:

[0089] The training process of E-HAMKF-SVM is divided into four stages: adaptive sample weight calculation, multi-kernel fusion classifier construction, hyperparameter optimization, and model retraining. The entire process uses a hierarchical structure and a multi-kernel fusion strategy to classify urinary protein concentration.

[0090] Phase 1: Adaptive Sample Weight Calculation;

[0091] remember Let c be the number of samples. Given the total number of samples, the weight of each category is... for:

[0092] ;

[0093] Phase 2: Construction of a multi-core fusion classifier; Five SVM classifiers with different configurations are fused, as follows:

[0094] I. SVM1: RBF kernel, standard weights, with a value of 1.0, used to capture non-linear features in the data;

[0095] II. SVM2: Linear kernel, standard weights, with a value of 1.0, used to process linear features in the data;

[0096] III. SVM3: Polynomial kernel, standard weights, with a value of 1.0, used to capture higher-order interactions between features;

[0097] IV. SVM4: RBF kernel, enhanced weights, with a value of 1.5, used to strengthen the influence of the RBF classifier in ensemble voting;

[0098] V. SVM5: Linear kernel, enhances weights, with a value of 1.5, used to strengthen the influence of the Linear classifier in ensemble voting;

[0099] Phase 3: Hyperparameter optimization;

[0100] The goal of hyperparameter optimization is to find the combination of hyperparameters that best performs the model on the validation set within a given search space.

[0101] (a) Hyperparameter search space;

[0102] Hyperparameter spaces of SVM1 and SVM4:

[0103] Regularization parameters ;

[0104] Kernel function parameters The number of parameter combinations is 49.

[0105] Hyperparameter space of SVM2 and SVM5:

[0106] Regularization parameters The number of parameter combinations is 7.

[0107] SVM3's hyperparameter space:

[0108] Regularization parameters ;

[0109] polynomial degree ;

[0110] Kernel function parameters The number of parameter combinations is 196.

[0111] (b) GridSearchCV optimization process;

[0112] Step 1: Parameter combination generation;

[0113] Generate all parameter combinations based on the search space;

[0114] Step 2: 5-fold stratified cross-validation;

[0115] For each parameter combination, perform 5-fold hierarchical cross-validation: divide the training set into 5 folds of equal size, each fold maintaining the original class distribution; select the 1st, 2nd, 3rd, 4th, and 5th folds as the validation set, and the remaining 4 folds as the training set; perform a total of 5 training and validation iterations.

[0116] Step 3: Calculation of evaluation indicators;

[0117] Calculate on the validation set for each fold Score the results, then average them as the cross-validation score for the current parameter combination:

[0118] ;

[0119] in, These are parameter constraints. Let be the parameter search space, representing the set of all possible parameter combinations; the cross-validation score represents the sampled parameters. The trained model on the i-th fold validation set Score; Indicates the weight of the F1 score;

[0120] Step 4: Optimal parameter selection;

[0121] Iterate through all parameter combinations and select the parameter combination that yields the highest cross-validation score as the optimal parameters:

[0122] ;

[0123] in, The parameter maximization operator;

[0124] Phase 4: Model retraining;

[0125] The training and validation sets are merged, and five SVM classifiers are retrained on the merged dataset using the optimal hyperparameters to obtain the final multi-core fusion model.

[0126] Optionally, step 6 includes:

[0127] The prediction results of five SVM classifiers with different configurations are combined, and a confidence-weighted voting mechanism is used to obtain the final predicted category. The expression for the confidence-weighted voting mechanism is as follows:

[0128] ;

[0129] in, The classifier weights are... ; Let be the predicted probability of classifier i for class c, with a value range of [0,1], and the constraint condition is: , This indicates six categories of urine protein concentration levels.

[0130] In a second aspect, the present invention provides a grading device for urine protein detection, the device being used to implement the grading method for urine protein detection in the first aspect or any possible implementation of the first aspect, the device comprising: a base, upper and lower sliders, a test strip holder, a trapezoidal holder, a SIM card slot, a right light source clamp, and a left light source clamp;

[0131] The base and the upper and lower sliders are fitted with through holes. The upper and lower sliders have two pairs of fine holes that engage with the through holes in the base to achieve vertical sliding. The upper and lower sliders and the test strip holder are fitted with four through holes. The test strip holder has two pairs of fine holes that engage with the through holes in the upper and lower sliders to achieve horizontal sliding. The test strip holder has a slot to hold the nitrocellulose membrane test strip with the added dye mixture in place. The trapezoidal holder and the test strip holder are fitted with through holes. The trapezoidal holder has multiple fine holes that engage with the through holes in the base to achieve vertical sliding and adjust the height. (Mobile phone slot included.) The trapezoidal holder uses through holes for connection. The phone card slot has multiple small holes that engage with the through holes in the trapezoidal holder to allow for back-and-forth sliding. The phone card slot is used to hold the shooting equipment. The right light source clamp and the base are connected by a slot. The right light source clamp has a protrusion that engages with the groove in the base to allow for left-right sliding. The left light source clamp and the base are connected by a slot. The left light source clamp has a protrusion that engages with the groove in the base to allow for left-right sliding. The left light source clamp and the right light source clamp are connected by a clamp. The right light source clamp and the left light source clamp are used to fix the light source equipment.

[0132] The technical solution provided by this invention includes the following steps: collecting urine samples, classifying and labeling the samples; staining the urine samples based on the classification and labeling to prepare sample images and construct a sample dataset; preprocessing the sample dataset to obtain a training set, a validation set, and a test set; extracting features from the preprocessed sample data and preprocessing the extracted features to select the 58 most important features; training an enhanced hierarchical adaptive multi-kernel fusion support vector machine (E-HAMKF-SVM) to obtain the final multi-kernel fusion model; and using the multi-kernel fusion model to obtain hierarchical prediction categories. This method improves the accuracy, generalization ability, and clinical applicability of urine protein detection. Attached Figure Description

[0133] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0134] Figure 1 A flowchart of a grading method for urine protein detection provided in an embodiment of the present invention;

[0135] Figure 2 This is a preprocessed C0 class sample image provided in an embodiment of the present invention;

[0136] Figure 3 The preprocessed C1 class sample image provided in this embodiment of the invention;

[0137] Figure 4The preprocessed C2 class sample image provided in this embodiment of the invention;

[0138] Figure 5 The preprocessed C3 class sample image provided in this embodiment of the invention;

[0139] Figure 6 The preprocessed C4 class sample image provided in this embodiment of the invention;

[0140] Figure 7 This is a preprocessed C5 class sample image provided in an embodiment of the present invention;

[0141] Figure 8 The confusion matrix diagram of the E-HAMKF-SVM algorithm provided in the embodiments of the present invention;

[0142] Figure 9 The confusion matrix diagram of the SVM-RBF algorithm provided in the embodiments of the present invention;

[0143] Figure 10 The SVM-Linear algorithm confusion matrix diagram provided in the embodiments of the present invention;

[0144] Figure 11 The confusion matrix diagram of the Random Forest algorithm provided in the embodiments of the present invention;

[0145] Figure 12 The confusion matrix diagram of the Gradient Boosting algorithm provided in the embodiments of the present invention;

[0146] Figure 13 The confusion matrix diagram of the KNN algorithm provided in the embodiments of the present invention;

[0147] Figure 14 A confusion matrix diagram of the Logistic Regression algorithm provided in this embodiment of the invention;

[0148] Figure 15a A comparison chart of accuracy performance indicators provided in the embodiments of the present invention;

[0149] Figure 15b A comparison chart of accuracy performance indicators provided in embodiments of the present invention;

[0150] Figure 15c A comparison chart of recall performance metrics provided in embodiments of the present invention;

[0151] Figure 15d A comparison chart of F1 score performance metrics provided in embodiments of the present invention;

[0152] Figure 15eA comparison chart of Cohen's Kappa performance metrics provided in embodiments of the present invention;

[0153] Figure 16 A schematic diagram of a grading device for urine protein detection provided in an embodiment of the present invention;

[0154] Figure 17 A schematic diagram of the base provided in an embodiment of the present invention;

[0155] Figure 18 This is a schematic diagram of the upper and lower sliders provided in an embodiment of the present invention;

[0156] Figure 19 This is a schematic diagram of the test strip holder provided in an embodiment of the present invention;

[0157] Figure 20 This is a schematic diagram of a trapezoidal support provided in an embodiment of the present invention;

[0158] Figure 21 This is a schematic diagram of a mobile phone SIM card slot provided in an embodiment of the present invention;

[0159] Figure 22 This is a schematic diagram of the right light source fixture provided in an embodiment of the present invention;

[0160] Figure 23 This is a schematic diagram of the left light source fixture provided in an embodiment of the present invention.

[0161] The numbers in the diagram are: 1-base, 2-upper and lower sliders, 3-test paper holder, 4-trapezoidal holder, 5-SIM card slot, 6-right light source clamp, 7-left light source clamp. Detailed Implementation

[0162] 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. 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.

[0163] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0164] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0165] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0166] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0167] Figure 1 The flowchart of the grading method for urine protein detection provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method includes:

[0168] Step 1: Collect urine samples, classify and label the samples.

[0169] In this embodiment of the invention, step 1 includes:

[0170] Step 11: Collect urine samples: Collect midstream urine samples with different concentrations of urinary protein; no demographic restrictions (age, gender, etc.) are set for urine samples to ensure the diversity of sample sources;

[0171] Step 12, Sample Classification and Labeling: Based on the clinical routine urine test results, the samples were divided into 6 different categories according to the range of urine protein concentration, and corresponding category labels were assigned. The urine sample category labels are shown in Table 1. When the urine protein concentration is <0.1g / L, the category label is C0 (or —); when the urine protein concentration is 0.1g~0.2g / L, the category label is C1 (or ±); when the urine protein concentration is 0.2g~1.0g / L, the category label is C2 (or +); when the urine protein concentration is 1.0g~2.0g / L, the category label is C3 (or ++); when the urine protein concentration is 2.0g~4.0g / L, the category label is C4 (or +++); when the urine protein concentration is >4.0g / L, the category label is C5 (or +++++).

[0172] Table 1. Labels for Urine Sample Categories

[0173] .

[0174] Step 2: Based on sample classification and annotation, stain the urine samples, prepare sample images, and construct the sample dataset.

[0175] In this embodiment of the invention, step 2 includes:

[0176] Step 21, Staining: Add Coomassie Brilliant Blue staining solution to the labeled urine sample at a volume ratio of 5:1, and mix thoroughly using a vortex mixer to form a homogeneous staining mixture.

[0177] Step 22, Sample Image Preparation: Use a micropipette to transfer 2 μL of the staining mixture and drop it onto the surface of the nitrocellulose membrane. Let it stand at room temperature for 1 minute to form a circular imprint. Define the circular imprint as a sample image.

[0178] Repeat the above sampling process for all classified and labeled samples to construct a sample dataset of no less than 1800 sample images, with no less than 300 images for each urinary protein concentration level, to ensure a balanced distribution of samples in each category.

[0179] Step 3: Preprocess the sample dataset to obtain the training set, validation set, and test set.

[0180] In this embodiment of the invention, step 3 includes:

[0181] Step 31, Data Cleaning: Delete sample images that are out of focus, have failed to drop (irregular shape, severe diffusion), or were contaminated during shooting;

[0182] Step 32, Data Standardization: Adjust all sample images to 256×256 pixels, normalize pixel intensity values ​​to the range of [0,1], and perform color space standardization BGR, HSV, LAB;

[0183] Step 33: Dataset partitioning: Divide the sample dataset into a training set (1260 samples), a validation set (270 samples), and a test set (270 samples) in a ratio of 70:15:15.

[0184] Step 4: Extract features from the preprocessed sample data and preprocess the extracted features, selecting the 58 most important features.

[0185] In this embodiment of the invention, step 4, feature extraction of the preprocessed sample data, includes:

[0186] An enhanced feature extractor is used to extract a 115-dimensional feature vector from each sample image, including 72-dimensional color space features, 15-dimensional region features, 16-dimensional texture features, and 12-dimensional morphological features, denoted as . ;

[0187] Step 41, 72-dimensional color space features include 24-dimensional BGR color space features, 24-dimensional HSV color space features, and 24-dimensional LAB color space features; (Note:) For the BGR color space, the channel identifier is: For the HSV color space, For the LAB color space, For each color space Each channel Extract 8 statistics:

[0188] ;

[0189] in, For the corresponding color space channels The mean of is expressed as . N is the total number of pixels in the current channel; Let be the pixel value of the i-th pixel, which ranges from 0 to 255; Let $\begin{pmatrix}$ be the standard deviation, and its expression is: , The qth percentile is mathematically defined as the value of all pixel values ​​sorted in ascending order. To ensure that q% of the data is less than or equal to its value; for example, for the pixel value sequence [50, 100, 150, 200], To ensure that the data in q% are less than or equal to its value, then That is, take the first value of the pixel value sequence. ; and These are the minimum and maximum values ​​of the pixel values, respectively.

[0190] Therefore, color feature vector The dimension is 3 color spaces × 3 channels × 8 statistics = 72 dimensions;

[0191] Step 42, 15-dimensional region features include 3D central region color mean features, 3D edge region color mean features, 3D background region color mean features, 3D center-background relative difference features, and 3D edge-background relative difference features;

[0192] Let the center of the circle be... , radius is , Define the central region for the coordinates of any point in the sample image. Edge areas and background area Their expressions are as follows:

[0193] ;

[0194] ;

[0195] ;

[0196] Calculate the color mean feature of the central region Color mean characteristics of edge regions and background area color mean characteristics Their expressions are as follows:

[0197] ;

[0198] ;

[0199] ;

[0200] in, The total number of pixels in the central region. It is the sum of all pixel values ​​within the central region; This represents the total number of pixels in the edge region. It is the sum of all pixel values ​​within the edge region; This represents the total number of pixels in the background area. This is the sum of all pixel values ​​within the background area;

[0201] Center-background relative difference features The expression is:

[0202] ;

[0203] Edge-background relative difference features The expression is:

[0204] ;

[0205] in This is used to avoid the denominator being zero;

[0206] Therefore, the eigenvector of the 15-dimensional region is: ;

[0207] Step 43: The 16-dimensional texture features include 4-dimensional Sobel gradient features, 1-dimensional Canny edge features, 3-dimensional Laplacian features, 4-dimensional circular region gradient features, and 4-dimensional color gradient features.

[0208] Sobel gradient eigenvectors The expression is:

[0209] ;

[0210] in, , , , Gradient magnitude Mean, standard deviation, maximum value, 90th percentile; gradient magnitude Defined as , The gradient is in the horizontal direction. The gradient is in the vertical direction. Represents the image matrix, which is the set of pixel values ​​for the entire sample image;

[0211] The Canny edge feature vector expression is:

[0212] ;

[0213] in, The number of edge pixels. This represents the total number of pixels in the current channel.

[0214] The Laplacian eigenvector expression is: ;

[0215] in, , and They are respectively The mean, standard deviation, and maximum value; The expression is: ,in Let I be the second derivative of the image matrix I along the x-direction. Let be the second derivative of the image matrix I along the y-direction;

[0216] The expression for the gradient feature vector of the circular region is:

[0217] ;

[0218] in, , , and The gradient sets within the circular regions are respectively. The mean, standard deviation, maximum value, and 90th percentile; The expression is:

[0219] ;

[0220] in, As the central area, For point Gradient magnitude at;

[0221] The expression for the color gradient feature vector is:

[0222] ;

[0223] in, , , and These represent the gradients of the BGR channel, respectively, and are expressed as follows: , , ;

[0224] Therefore, the 16-dimensional texture feature vector is: ;

[0225] Step 44: 12-dimensional morphological features include 2-dimensional erosion features, 2-dimensional dilation features, 2-dimensional opening operation features, 2-dimensional closing operation features, and 4-dimensional morphological difference features.

[0226] Morphological eigenvectors are denoted as:

[0227] ;

[0228] in, The average brightness after erosion is N, and N is the total number of pixels in the current channel. The standard deviation after erosion represents the degree of fluctuation in the brightness of the eroded image. and These are the average brightness and standard deviation after dilation, respectively; and These are the average brightness and the standard deviation after the opening operation, respectively. and These are the average brightness and standard deviation after the closing operation, respectively; The difference between the original and corroded layers is used to reflect the boundary thickness of protein precipitation. The difference between the expanded and original regions is used to reflect the protein signal intensity in the surrounding area; The difference between the original and opening operations is used to reflect the noise intensity in the image; The difference between the closing operation and the original value is represented by the pixel value added during the closing operation.

[0229] The expressions for erosion, dilation, opening, and closing operations are as follows:

[0230] Corrosion operation for:

[0231] ;

[0232] in, Let B be the image matrix and B be the structuring element, which is:

[0233] ;

[0234] expansion operation for:

[0235] ;

[0236] Opening operation is ;

[0237] Closing operation is .

[0238] In this embodiment of the invention, step 4 involves preprocessing the extracted features, selecting the 58 most important features, including:

[0239] The extracted features are preprocessed by standardizing them with StandardScaler to scale them to zero mean and unit variance; then, they are filtered by VarianceThreshold to remove features with variance of 0 or close to 0; finally, the 58 most important features are selected by SelectKBest.

[0240] The standard scaler expression is:

[0241] ;

[0242] in, This is a standardized feature vector used to transform features to a uniform scale and eliminate the dimensional differences between different features; The original feature vector, and These are the mean vector and standard deviation vector of the training set;

[0243] The expression for VarianceThreshold variance filtering is:

[0244] ;

[0245] in, This is the feature vector after variance filtering, used to remove features with zero variance while retaining features with information content; Let be the variance of the i-th feature;

[0246] SelectKBest selects the 58 most important features:

[0247] ;

[0248] in, The final 58-dimensional feature vector selected for model training aims to balance information preservation and dimensionality reduction.

[0249] Step 5: Train the enhanced hierarchical adaptive multi-core fusion support vector machine E-HAMKF-SVM to obtain the final multi-core fusion model.

[0250] In this embodiment of the invention, step 5 includes:

[0251] The Enhanced Hierarchical Adaptive Multi-Kernel Fusion Support Vector Machine (E-HAMKF-SVM) is used for automatic classification of urine protein concentration images. The training process of E-HAMKF-SVM consists of four stages: adaptive sample weight calculation, multi-kernel fusion classifier construction, hyperparameter optimization, and model retraining. The entire process uses a hierarchical structure and a multi-kernel fusion strategy to achieve classification of urine protein concentration.

[0252] Phase 1: Adaptive Sample Weight Calculation;

[0253] remember Let c be the number of samples. Given the total number of samples, the weight of each category is... for:

[0254] ;

[0255] Phase 2: Construction of a multi-core fusion classifier; Five SVM classifiers with different configurations are fused, as follows:

[0256] I. SVM1: RBF kernel, standard weights, with a value of 1.0, used to capture non-linear features in the data;

[0257] II. SVM2: Linear kernel, standard weights, with a value of 1.0, used to process linear features in the data;

[0258] III. SVM3: Polynomial kernel, standard weights, with a value of 1.0, used to capture higher-order interactions between features;

[0259] IV. SVM4: RBF kernel, enhanced weights, with a value of 1.5, used to strengthen the influence of the RBF classifier in ensemble voting;

[0260] V. SVM5: Linear kernel, enhances weights, with a value of 1.5, used to strengthen the influence of the Linear classifier in ensemble voting;

[0261] Phase 3: Hyperparameter optimization;

[0262] The goal of hyperparameter optimization is to find the combination of hyperparameters that best performs the model on the validation set within a given search space.

[0263] (a) Hyperparameter search space;

[0264] Hyperparameter spaces of SVM1 and SVM4 (RBF SVM):

[0265] Regularization parameters ;

[0266] Kernel function parameters The number of parameter combinations is 49.

[0267] Hyperparameter space of SVM2 and SVM5 (Linear SVM):

[0268] Regularization parameters The number of parameter combinations is 7.

[0269] The hyperparameter space of SVM3 (Polynomial SVM):

[0270] Regularization parameters ;

[0271] polynomial degree ;

[0272] Kernel function parameters The number of parameter combinations is 196.

[0273] (b) GridSearchCV optimization process;

[0274] Step 1: Parameter combination generation;

[0275] Generate all parameter combinations based on the search space;

[0276] Step 2: 5-fold stratified cross-validation;

[0277] For each parameter combination, perform 5-fold hierarchical cross-validation: divide the training set into 5 folds of equal size, each fold maintaining the original class distribution; select the 1st, 2nd, 3rd, 4th, and 5th folds as the validation set, and the remaining 4 folds as the training set; perform a total of 5 training and validation iterations.

[0278] Step 3: Calculation of evaluation indicators;

[0279] Calculate on the validation set for each fold Score the results, then average them as the cross-validation score for the current parameter combination:

[0280] ;

[0281] in, These are parameter constraints. Let be the parameter search space, representing the set of all possible parameter combinations (e.g., ...). Cross-validation score represents the sampling parameters. The trained model on the i-th fold validation set Score; The weights representing the F1 score are expressed as follows:

[0282] ;

[0283] in, The score is the F1 score. Let c be the number of samples. This represents the total number of samples;

[0284] Step 4: Optimal parameter selection;

[0285] Iterate through all parameter combinations and select the parameter combination that yields the highest cross-validation score as the optimal parameters:

[0286] ;

[0287] in, The parameter maximization operator;

[0288] Phase 4: Model retraining;

[0289] The training and validation sets are merged, and five SVM classifiers are retrained on the merged dataset using the optimal hyperparameters to obtain the final multi-core fusion model.

[0290] Step 6: Use a multi-core fusion model to obtain hierarchical prediction categories.

[0291] In this embodiment of the invention, step 6 includes:

[0292] The prediction results of five SVM classifiers with different configurations are combined, and a confidence-weighted voting mechanism is used to obtain the final predicted category. The expression for the confidence-weighted voting mechanism is as follows:

[0293] ;

[0294] in, The classifier weights are... ; Let be the predicted probability of classifier i for class c, with a value range of [0,1], and the constraint condition is: , This indicates six categories of urine protein concentration levels.

[0295] This invention provides performance evaluation by performing predictions on a test set, calculating performance metrics such as accuracy, precision, recall, F1 score, and Cohen's Kappa score, generating a confusion matrix, and then conducting performance comparison analysis. The confusion matrix used is a... The matrix has the following structure:

[0296] ;

[0297] In this context, TP stands for True Positive and FP stands for False Positive.

[0298] Accuracy is defined as the number of correctly predicted samples divided by the total number of samples. Precision is defined as the number of samples correctly predicted as c divided by the total number of samples predicted as c. Recall is defined as the number of samples correctly predicted as c divided by the total number of samples that actually returned c.

[0299] For class c, the F1 score is defined as:

[0300] ;

[0301] Among them, accuracy Recall represents the proportion of samples that actually belong to class c out of all samples predicted as class c; This represents the proportion of samples that are correctly predicted as class c out of all samples that actually belong to class c.

[0302] Cohen's Kappa coefficient is defined as follows:

[0303] ;

[0304] in, For observed consistency (accuracy). Expected consistency (accuracy of random guesses) measures how much the model's performance exceeds that of random guesses.

[0305] Experimental results and analysis of this invention:

[0306] Experimental setup and dataset: After sample collection, classification and annotation, sample staining and image preparation, and sample preprocessing, a total of 1800 urinary protein concentration sample images were collected, covering 6 concentration levels (C0~C5), with a perfectly balanced number of samples per class (300 images per class). The dataset was split into a training set (1260 images, 210 images per class), a validation set (270 images, 45 images per class), and a test set (270 images, 45 images per class) in a 70:15:15 ratio, as follows: Figures 2 to 7 The images shown are partial sample images, with 6 sample images shown for each concentration level, namely sample 1 to sample 6.

[0307] In embodiments of the present invention, such as Figure 8 As shown, confusion matrix analysis reveals the classification performance characteristics of E-HAMKF-SVM. In the test set of 270 samples, E-HAMKF-SVM correctly classified 266 samples and misclassified only 4 samples, with an error rate of 1.48%. Compared with other algorithms, E-HAMKF-SVM has the fewest misclassified samples; Figures 9 to 14 As shown, SVM-RBF had 6 misclassifications with an error rate of 2.22%; SVM-Linear had 8 misclassifications with an error rate of 2.96%; Random Forest had 9 misclassifications with an error rate of 3.33%; Gradient Boosting had 10 misclassifications with an error rate of 3.70%; LogisticRegression had 13 misclassifications with an error rate of 4.81%; and KNN had 23 misclassifications with an error rate of 8.52%. The above data indicate that E-HAMKF-SVM has the strongest robustness in handling the classification of urine protein concentration.

[0308] The E-HAMKF-SVM algorithm proposed in this invention also achieved excellent performance on the test set, such as... Figures 15a to 15e As shown, in the comparison of five key performance indicators (accuracy, precision, recall, F1 score, and Cohen's Kappa), the E-HAMKF-SVM algorithm achieved an accuracy of 98.52% (correctly classifying 266 out of 270 samples), a precision of 98.54%, a recall of 98.52%, and an F1 score of 98.52%. Notably, the Cohen's Kappa coefficient reached 0.9822. E-HAMKF-SVM outperformed traditional machine learning algorithms in all metrics.

[0309] Under a rigorous three-part data evaluation framework, the E-HAMKF-SVM algorithm achieved an accuracy of 98.52% on the test set, representing a 0.74 percentage point improvement over the closest benchmark method, SVM-RBF, and a relative improvement of 33%. It also achieved perfect classification across all concentration levels. These results not only significantly outperform other comparable algorithms, including random forest, gradient boosting, and logistic regression, but also establish a new performance benchmark for the automated diagnosis of urinary protein concentration. The high accuracy, perfect classification consistency, and balanced performance across all concentration levels of the E-HAMKF-SVM algorithm give it significant potential for clinical applications and can provide strong support for the development of medical diagnostic systems.

[0310] The above results demonstrate that the organic combination of innovative technologies proposed in this invention, such as multi-core fusion strategy, adaptive sample weighting, aggressive hyperparameter optimization, confidence-weighted voting mechanism, and correct data segmentation, can effectively improve classification performance and achieve breakthrough performance in the urine protein concentration image classification task.

[0311] Compared with the prior art, the present invention has the following advantages:

[0312] It boasts high accuracy, achieving 98.52% accuracy on the test set, outperforming existing algorithms. Its classification performance is excellent, achieving zero misclassifications on the test set, correctly classifying all six concentration levels. It exhibits good consistency across metrics, with accuracy, precision, recall, and F1 score all reaching 98.52%, and a Cohen's Kappa coefficient of 0.9822, indicating excellent consistency. It demonstrates strong generalization ability, showing stable performance across all concentration levels without any class-specific bias. It offers high interpretability, with confidence-weighted voting providing transparency in decision-making, facilitating understanding and trust among clinicians. It has high application value and broad application prospects, applicable to hospital automated urine analysis systems, bedside diagnostic equipment, telemedicine platforms, home health monitoring, and can also be extended to other medical image classification tasks.

[0313] This invention provides a grading device for urine protein detection, such as... Figures 16 to 23 As shown, the device is used to implement the above-mentioned grading method for urine protein detection; the device includes: a base 1, upper and lower sliders 2, test strip holder 3, trapezoidal holder 4, mobile phone card slot 5, right light source clamp 6 and left light source clamp 7;

[0314] The base 1 and the upper and lower sliders 2 are fitted with through holes. The upper and lower sliders 2 have two pairs of fine holes that engage with the through holes in the base 1 to achieve vertical sliding. The upper and lower sliders 2 and the test strip holder 3 are fitted with four through holes. The test strip holder 3 has two pairs of fine holes that engage with the through holes in the upper and lower sliders 2 to achieve horizontal sliding. The test strip holder 3 has a slot for securing the nitrocellulose membrane test strip with the added dye mixture. The trapezoidal holder 4 and the test strip holder 3 are fitted with through holes. The trapezoidal holder 4 has multiple fine holes that engage with the through holes in the base 1 to achieve vertical sliding and adjust the height. The mobile phone SIM card slot 5 and the trapezoidal holder... 4. A through-hole fit is adopted. The mobile phone card slot 5 has multiple small holes, which fit with the through holes of the trapezoidal bracket 4 to achieve back-and-forth sliding. The mobile phone card slot 5 is used to place shooting equipment (mobile phone, camera, etc.). The right light source clamp 6 and the base 1 are fitted with a card slot. The right light source clamp 6 has a protrusion that fits with the groove of the base 1 to achieve left-right sliding. The left light source clamp 7 and the base 1 are fitted with a card slot. The left light source clamp 7 has a protrusion that fits with the groove of the base 1 to achieve left-right sliding. The left light source clamp 7 and the right light source clamp 6 are fitted with a clamp. The right light source clamp 6 and the left light source clamp 7 are used to fix the light source equipment.

[0315] The device of the present invention is relatively flexible. The mobile phone card slot 5 can slide on the trapezoidal support 4, the trapezoidal support 4 can move up and down, the upper and lower sliders 2 can move up and down, the test paper holder 3 can move horizontally, the right light source clamp 6 and the left light source clamp 7 can move horizontally, and the holes on the device are used to fix the various structures with screws.

[0316] The technical solution provided by this invention includes the following steps: collecting urine samples, classifying and labeling the samples; staining the urine samples based on the classification and labeling to prepare sample images and construct a sample dataset; preprocessing the sample dataset to obtain a training set, a validation set, and a test set; extracting features from the preprocessed sample data and preprocessing the extracted features to select the 58 most important features; training an enhanced hierarchical adaptive multi-kernel fusion support vector machine (E-HAMKF-SVM) to obtain the final multi-kernel fusion model; and using the multi-kernel fusion model to obtain hierarchical prediction categories. This method improves the accuracy, generalization ability, and clinical applicability of urine protein detection.

[0317] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fractionation method for urinary protein detection, characterized in that, The method includes: Step 1: Collect urine samples, classify and label the samples; Step 2: Based on sample classification and annotation, stain the urine samples, prepare sample images, and construct the sample dataset; Step 3: Preprocess the sample dataset to obtain the training set, validation set, and test set; Step 4: Extract features from the preprocessed sample data, and preprocess the extracted features to select the 58 most important features; Step 5: Train the enhanced hierarchical adaptive multi-core fusion support vector machine E-HAMKF-SVM to obtain the final multi-core fusion model; Step 6: Use a multi-core fusion model to obtain hierarchical prediction categories; Step 4, which involves feature extraction from the preprocessed sample data, includes: A 115-dimensional feature vector is extracted from each sample image using the enhanced feature extractor, including 72-dimensional color space features, 15-dimensional number region features, 16-dimensional texture features, and 12-dimensional morphological features, denoted as ; Step 41, 72-dimensional color space features include 24-dimensional BGR color space features, 24-dimensional HSV color space features, and 24-dimensional LAB color space features; (Note:) For the BGR color space, the channel identifier is: For the HSV color space, For the LAB color space, For each color space Each channel Extract 8 statistics: ； in, For the corresponding color space channels The mean of N, where N is the total number of pixels in the current channel; Let be the pixel value of the i-th pixel, which ranges from 0 to 255; Standard deviation, The qth percentile is mathematically defined as the value of all pixel values ​​sorted in ascending order. To ensure that the data in q% is less than or equal to its value; and These are the minimum and maximum values ​​of the pixel values, respectively. Therefore, color feature vector The dimension is 3 color spaces × 3 channels × 8 statistics = 72 dimensions; Step 42, 15-dimensional region features include 3D central region color mean features, 3D edge region color mean features, 3D background region color mean features, 3D center-background relative difference features, and 3D edge-background relative difference features; Let the center of the circle be... , radius is , Define the central region for the coordinates of any point in the sample image. Edge area and background area Their expressions are as follows: ； ； ； Calculate the color mean feature of the central region Color mean characteristics of edge regions and background area color mean characteristics Their expressions are as follows: ； ； ； in, The total number of pixels in the central region. It is the sum of all pixel values ​​within the central region; This represents the total number of pixels in the edge region. It is the sum of all pixel values ​​within the edge region; This represents the total number of pixels in the background area. This is the sum of all pixel values ​​within the background area; Center-background relative difference features The expression is: ； Edge-background relative difference features The expression is: ； in This is used to avoid the denominator being zero; Therefore, the eigenvector of the 15-dimensional region is: ; Step 43: The 16-dimensional texture features include 4-dimensional Sobel gradient features, 1-dimensional Canny edge features, 3-dimensional Laplacian features, 4-dimensional circular region gradient features, and 4-dimensional color gradient features. Sobel gradient eigenvectors The expression is: ； in, , , , Gradient magnitude Mean, standard deviation, maximum value, 90th percentile; gradient magnitude Defined as , The gradient is in the horizontal direction. The gradient is in the vertical direction. Represents the image matrix, which is the set of pixel values ​​for the entire sample image; The Canny edge feature vector expression is: ； in, The number of edge pixels. This represents the total number of pixels in the current channel. The Laplacian eigenvector expression is: ; in, , and They are respectively The mean, standard deviation, and maximum value; The expression is: ,in Let I be the second derivative of the image matrix I along the x-direction. Let be the second derivative of the image matrix I along the y-direction; The expression for the gradient eigenvector of the circular region is: ； in, , , and The gradient sets within the circular regions are respectively. The mean, standard deviation, maximum value, and 90th percentile; The expression is: ； in, As the central area, For point Gradient magnitude at; The expression for the color gradient feature vector is: ； in, , , and These represent the gradients of the BGR channels; Therefore, the 16-dimensional texture feature vector is: ; Step 44: 12-dimensional morphological features include 2-dimensional erosion features, 2-dimensional dilation features, 2-dimensional opening operation features, 2-dimensional closing operation features, and 4-dimensional morphological difference features. Morphological eigenvectors are denoted as: ； in, The average brightness after erosion is N, and N is the total number of pixels in the current channel. The standard deviation after erosion represents the degree of fluctuation in the brightness of the eroded image. and These are the average brightness and standard deviation after dilation, respectively; and These are the average brightness and the standard deviation after the opening operation, respectively. and These are the average brightness and standard deviation after the closing operation, respectively; The difference between the original and corroded layers is used to reflect the boundary thickness of protein precipitation. The difference between the expanded and original regions is used to reflect the protein signal intensity in the surrounding area; The difference between the original and opening operations is used to reflect the noise intensity in the image; The difference between the closing operation and the original value is represented by the pixel value added during the closing operation. The expressions for erosion, dilation, opening, and closing operations are as follows: Corrosion operation for: ； in, Let B be the image matrix and B be the structuring element, which is: ； expansion operation for: ； Opening operation is ; Closing operation is ; The E-HAMKF-SVM is constructed using a multi-core fusion classifier; it employs five SVM classifiers with different configurations, as follows: I. SVM1: RBF kernel, standard weights, with a value of 1.0, used to capture non-linear features in the data; II. SVM2: Linear kernel, standard weights, with a value of 1.0, used to process linear features in the data; III. SVM3: Polynomial kernel, standard weights, with a value of 1.0, used to capture higher-order interactions between features; IV. SVM4: RBF kernel, enhanced weights, with a value of 1.5, used to strengthen the influence of the RBF classifier in ensemble voting; V. SVM5: Linear kernel, enhanced weights, with a value of 1.5, used to strengthen the influence of the Linear classifier in ensemble voting.

2. The method according to claim 1, characterized in that, Step 1 includes: Step 11: Collect urine samples: Collect midstream urine samples with different concentrations of urinary protein; no demographic restrictions are set for the urine samples to ensure the diversity of sample sources; Step 12, Sample Classification and Labeling: Based on the clinical routine urine test results, the samples were divided into 6 different categories according to the range of urine protein concentration and assigned corresponding category labels; the category label was C0 when the urine protein concentration was <0.1g / L; C1 when the urine protein concentration was 0.1g to 0.2g / L; C2 when the urine protein concentration was 0.2g to 1.0g / L; C3 when the urine protein concentration was 1.0g to 2.0g / L; C4 when the urine protein concentration was 2.0g to 4.0g / L; and C5 when the urine protein concentration was >4.0g / L.

3. The method according to claim 2, characterized in that, Step 2 includes: Step 21, Staining: Add Coomassie Brilliant Blue staining solution to the labeled urine sample at a volume ratio of 5:1, and mix thoroughly using a vortex mixer to form a homogeneous staining mixture. Step 22, Sample Image Preparation: Use a micropipette to transfer 2 μL of the staining mixture and drop it onto the surface of the nitrocellulose membrane. Let it stand at room temperature for 1 minute to form a circular imprint. Define the circular imprint as a sample image. Repeat the above sampling process for all classified and labeled samples to construct the sample dataset.

4. The method according to claim 3, characterized in that, Step 3 includes: Step 31, Data Cleaning: Delete sample images that are out of focus, have failed to drop samples, or are contaminated during shooting; Step 32, Data Standardization: Adjust all sample images to 256×256 pixels, normalize pixel intensity values ​​to the range of [0,1], and perform color space standardization BGR, HSV, LAB; Step 33: Dataset partitioning: Divide the sample dataset into training set, validation set and test set in a ratio of 70:15:

15.

5. The method according to claim 1, characterized in that, In step 4, the extracted features are preprocessed, and the 58 most important features are selected, including: The extracted features are preprocessed by standardizing them with StandardScaler to scale them to zero mean and unit variance; then, they are filtered by VarianceThreshold to remove features with variance of 0 or close to 0; finally, the 58 most important features are selected by SelectKBest. The standard scaler expression is: ； in, This is the standardized feature vector, used to transform features to a uniform scale; The original feature vector, and These are the mean vector and standard deviation vector of the training set; The expression for VarianceThreshold variance filtering is: ； in, This is the feature vector after variance filtering, used to remove features with variance of 0; Let be the variance of the i-th feature; SelectKBest selects the 58 most important features: ； in, This is the final 58-dimensional feature vector selected for model training.

6. The method according to claim 5, characterized in that, Step 5 includes: The training process of E-HAMKF-SVM is divided into four stages: adaptive sample weight calculation, multi-kernel fusion classifier construction, hyperparameter optimization, and model retraining. The entire process uses a hierarchical structure and a multi-kernel fusion strategy to classify urinary protein concentration. Phase 1: Adaptive Sample Weight Calculation; remember Let c be the number of samples. Given the total number of samples, the weight of each category is... for: ； Phase 2: Construction of a multi-core fusion classifier; 5 SVM classifiers with different configurations are fused together; Phase 3: Hyperparameter optimization; The goal of hyperparameter optimization is to find the combination of hyperparameters that best performs the model on the validation set within a given search space. (a) Hyperparameter search space; Hyperparameter spaces of SVM1 and SVM4: Regularization parameters ; Kernel function parameters The number of parameter combinations is 49. Hyperparameter space of SVM2 and SVM5: Regularization parameters The number of parameter combinations is 7. SVM3's hyperparameter space: Regularization parameters ; polynomial degree ; Kernel function parameters The number of parameter combinations is 196. (b) GridSearchCV optimization process; Step 1: Parameter combination generation; Generate all parameter combinations based on the search space; Step 2: 5-fold stratified cross-validation; For each parameter combination, perform 5-fold hierarchical cross-validation: divide the training set into 5 folds of equal size, each fold maintaining the original class distribution; select the 1st, 2nd, 3rd, 4th, and 5th folds as the validation set, and the remaining 4 folds as the training set; perform a total of 5 training and validation iterations. Step 3: Calculation of evaluation indicators; Calculate on the validation set for each fold Score the results, then average them as the cross-validation score for the current parameter combination: ； in, These are parameter constraints. Let be the parameter search space, representing the set of all possible parameter combinations; the cross-validation score represents the sampled parameters. The trained model on the i-th fold validation set Score; Indicates the weight of the F1 score; Step 4: Optimal parameter selection; Iterate through all parameter combinations and select the parameter combination that yields the highest cross-validation score as the optimal parameters: ； in, The parameter maximization operator; Phase 4: Model retraining; The training and validation sets are merged, and five SVM classifiers are retrained on the merged dataset using the optimal hyperparameters to obtain the final multi-core fusion model.

7. The method according to claim 6, characterized in that, Step 6 includes: The prediction results of five SVM classifiers with different configurations are combined, and a confidence-weighted voting mechanism is used to obtain the final predicted category. The expression for the confidence-weighted voting mechanism is as follows: ； in, The classifier weights are... ; Let be the predicted probability of classifier i for class c, with a value range of [0,1], and the constraint condition is: , This indicates six categories of urine protein concentration levels.

Images