Clinical data label correction method based on unsupervised learning and bayesian probability

Abstract

The present application relates to the technical field of clinical big data processing, and particularly discloses a clinical data label correction method based on unsupervised learning and Bayesian probability, which is used for automatic correction of a clinical data set without a pathological gold standard and only relying on clinical symptoms / laboratory indexes to form a non-gold standard noise label, and comprises the following steps executed in sequence: S1, performing a pretreatment operation on a clinical original data set, eliminating irrelevant redundant fields and completing feature standardization. In the present application, unsupervised clustering, Lasso feature screening, Bayesian iteration and parallel computing full-link algorithms are integrated, and the sample label does not need to be manually checked by a clinician piece by piece throughout the whole process, and the whole process of non-gold standard label generation, purification and iterative optimization is automatically completed.

Description

Technical Field

[0001] This invention relates to the field of clinical big data processing technology, and in particular to a clinical data label correction method based on unsupervised learning and Bayesian probability. Background Technology

[0002] Clinical medical cohort studies and real-world diagnostic and treatment data modeling work rely heavily on accurate sample labels. Pathological biopsy and gene testing can be considered as the gold standard for diagnosis. However, in clinical practice, a large number of primary care, chronic disease follow-up, and initial screening population datasets cannot collect gold standard results. Instead, non-gold standard labels can only be formed based on physicians' subjective judgments, single laboratory indicators, and combinations of patients' chief symptoms.

[0003] Existing non-gold standard labels have inherent flaws: subjective biases exist between different physicians' judgment criteria, detection errors exist in single test indicators, and redundant interference features exist in combinations of multiple symptoms. Ultimately, the labels are mixed with a large number of noisy samples, directly compressing the discriminative ability of subsequent prediction models. Currently, there is no automated correction solution in the industry; the mainstream approach is manual review and correction of labels sample by sample. This model has three major shortcomings: First, clinical samples often number in the tens of thousands, and manual verification is extremely time-consuming, with a single batch of datasets taking several weeks to revise, resulting in extremely low efficiency for research and clinical modeling. Second, manual revision relies on the individual clinical experience of physicians, leading to poor consistency in the results among different physicians, introducing secondary subjective noise, and failing to unify label judgment standards. Third, manual screening can only remove obviously abnormal samples, failing to quantify the contribution weight of each clinical symptom and test indicator to label judgment, making it difficult to remove interference features with no discriminative ability, and the redundancy of data dimensions continuously affects the accuracy of label correction. Therefore, to solve these problems, a clinical data label correction method based on unsupervised learning and Bayesian probability is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a clinical data label correction method based on unsupervised learning and Bayesian probability.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A clinical data label correction method based on unsupervised learning and Bayesian probability is used for automated correction of clinical datasets that lack pathological gold standards and rely solely on clinical symptoms / laboratory indicators to form non-gold standard noise labels. The method includes the following steps performed sequentially:

[0007] S1. Perform preprocessing on the original clinical dataset, removing irrelevant and redundant fields and completing feature standardization;

[0008] S2. Based on the unsupervised K-means binary clustering algorithm, generate initial binary labels for the dataset based on all original clinical features;

[0009] S3. Use Lasso regression combined with resampling stability selection to purify high-discrimination feature variables and construct a simplified feature subset;

[0010] S4. Iteratively update the labels for each sample based on the leave-one-out Bayesian posterior probability, determine the label convergence state by the cross-information entropy fluctuation amplitude, terminate the iteration and output the corrected label.

[0011] S5. Employs multi-class machine learning models such as SVM, XGBoost, and Random Forest for training and verification, and quantizes the performance improvement effect brought by label correction.

[0012] S6. Calculate the Bayesian posterior probability of all samples in a unified manner, and automatically generate a label distribution comparison chart and model ROC curve to complete the visualization output of the correction results.

[0013] Preferably, step S2, unsupervised clustering to generate initial labels, specifically includes:

[0014] Extract the pure feature matrix after preprocessing and removing irrelevant variables, set the number of K-means cluster centers to 2 to adapt to the clinical binary classification diagnosis scenario, map the cluster numbers one by one to binary 0 / 1 labels, and generate an uncorrected initial label set;

[0015] The clustering process supports multi-core parallel computing acceleration and outputs an intermediate dataset that binds the initial labels and original features, which serves as the input data source for subsequent variable filtering.

[0016] Preferably, step S3, Lasso variable selection and stability screening, includes:

[0017] A Lasso regression model is constructed using the initial labels output in step S2 as the dependent variable and all clinical features as independent variables.

[0018] The frequency of non-zero regression coefficients of each feature was statistically analyzed using multiple rounds of resampling stability selection, and highly stable clinical features with frequencies higher than a preset threshold were retained.

[0019] After removing low-stability noise features, a simplified feature dataset is reconstructed, feature dimension distillation is completed, the computational cost of subsequent Bayesian iterations is reduced, and the interference of invalid features on label determination is eliminated.

[0020] Preferably, step S4, the Bayesian posterior probability iterative optimization step, includes:

[0021] The mean values ​​of positive and negative labeled samples were calculated separately using the leave-one-out method. When the mean was 0, a minimum smoothing value of 1e-10 was added to avoid logarithmic calculation failure.

[0022] The posterior probability of a single sample belonging to two classes is solved in parallel, and the sample label is updated with a threshold of 0.5.

[0023] The average cross-entropy of the entire sample is calculated round by round. When the difference between the entropy of two consecutive rounds is lower than the preset threshold, the label is considered to have converged and the loop iteration is terminated.

[0024] Preferably, the Bayesian posterior probability calculation process calls on the device's multi-core resources to build a parallel computing cluster, splits the probability calculation task of all samples into multi-threaded synchronous operations, and releases the cluster resources after a single round of iteration, which greatly reduces the overall time consumption of multiple rounds of iteration for large-scale clinical datasets.

[0025] Preferably, in step S5, model validation uses three types of supervised learning models: radial basis kernel SVM, XGBoost, and random forest.

[0026] The dataset is divided into training and testing sets at a fixed ratio of 8:2, and standardized scaling features are used. The label correction gain effect is evaluated by a combination of three indicators: accuracy, sensitivity, and specificity.

[0027] Preferably, the visualization output in step S6 includes two types of core charts:

[0028] First, a histogram comparing the distribution of sample labels before and after correction visually reflects the proportion of noise label correction.

[0029] Second, the ROC curve output by the model validation quantifies the improvement in the model's discrimination ability.

[0030] All charts are automatically generated in batches using the drawing toolkit, eliminating the need for manual adjustment of drawing parameters.

[0031] Preferably, the method for performing the correction method according to any one of claims 1-7 includes six functional modules that sequentially interact with data:

[0032] The system includes a data preprocessing module, an unsupervised clustering label generation module, a Lasso stable feature selection module, a Bayesian label iterative optimization module, a multi-model validation module, and a result visualization output module.

[0033] Preferably, the data preprocessing module is used to read the local clinical dataset, remove redundant and irrelevant fields, and perform feature standardization;

[0034] The unsupervised clustering label generation module has built-in K-means binary clustering operation logic;

[0035] The Lasso stable feature selection module is equipped with Lasso regression and resampling stability selection algorithms;

[0036] The Bayesian label iterative optimization module is configured with a multi-core parallel computing interface and information entropy convergence judgment logic.

[0037] The multi-model validation module integrates three types of machine learning algorithms: SVM, XGBoost, and Random Forest.

[0038] The visualization output module has a built-in plotting toolkit that automatically generates label distribution comparison charts and ROC curves.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. This invention integrates unsupervised clustering, Lasso feature selection, Bayesian iteration, and parallel computing algorithms across the entire process. It eliminates the need for clinicians to manually review sample labels one by one, automating the entire process of non-gold standard label generation, purification, and iterative optimization. Furthermore, the correction logic of this invention relies entirely on objective data distribution and statistical probability calculations, eliminating subjective judgment biases caused by differences in physician clinical experience. Multiple runs on the same dataset consistently output the same labels, resolving the industry pain point of unreproducible results from manual revision. The entire process requires no manual intervention in intermediate steps; only the original clinical dataset needs to be input to automatically output converged, accurate labels. It is suitable for large-scale correction needs of real-world clinical follow-up data and outpatient screening data, significantly reducing the manpower and time costs of data processing in clinical research.

[0041] 2. This invention adds a Lasso regression combined with a resampling stability selection step after clustering generates initial labels. Unlike existing technologies that directly use all original features for computation, this approach quantifies the contribution of each clinical symptom and test indicator to the sample label, automatically eliminating low-stability, diagnostically worthless, and redundant noise variables. Clinical raw datasets often contain dozens of irrelevant signs and duplicate test indicators; redundant features directly interfere with K-means clustering boundaries, causing initial label clustering shifts. This invention reconstructs the dataset by resampling multiple times to statistically analyze the frequency of regression coefficients, retaining only highly stable core diagnostic features, significantly reducing feature dimensionality. The simplified feature set greatly reduces the computational cost of Bayesian posterior probability iteration, while eliminating interference from irrelevant variables in probability calculations. This improves the accuracy of label iteration updates from the underlying data level, preventing invalid features from continuously amplifying label noise and providing a high-quality input data source for subsequent Bayesian convergence optimization.

[0042] 3. This invention employs a leave-one-out Bayesian posterior probability method to update labels for each sample, coupled with a cross-information entropy convergence criterion mechanism, to overcome the shortcomings of existing single-clustering methods where labels remain fixed and lack iterative optimization capabilities. In each iteration, the algorithm recalculates the objective probabilities of all samples being positive or negative, dynamically correcting erroneous labels with a 0.5 probability threshold. Simultaneously, it calculates the average cross-information entropy of all samples as a quantitative indicator of label stability. Iteration automatically terminates when the entropy fluctuation falls below the threshold, avoiding infinite loops or premature label output before convergence. To address the issue of probability calculation failure caused by feature values ​​of 0 in clinical data, a minimum value smoothing logic is added to ensure that the posterior probability can be calculated correctly for all samples. The entire iterative mechanism continuously corrects noisy labels based on the data's own distribution patterns, requiring no external standard reference. It is compatible with all clinical datasets without a gold standard, the convergence criterion is objectively quantified, and the output labels possess statistical stability, eliminating label oscillations and incomplete correction issues.

[0043] 4. This invention adds a joint verification step using three classic machine learning models: SVM, XGBoost, and Random Forest, along with an automatic visualization and plotting module. This addresses the shortcomings of existing calibration techniques, which cannot objectively demonstrate the effectiveness of label optimization and whose results are difficult to interpret intuitively. The pre- and post-calibration datasets are input into the three models for training, and the accuracy, sensitivity, and specificity are compared to accurately quantify the performance improvement brought about by label revision. In the example, after calibration of the symptom dataset, the overall accuracy of the model improved by 15%, and the sensitivity of the test dataset improved by 10%, with the data gain supported by objective experimental data. The visualization module automatically generates label distribution comparison charts and model ROC curves, transforming abstract probability and classification indicators into intuitive charts. Researchers can visually distinguish the differences in data before and after revision without manual plotting. Attached Figure Description

[0044] Figure 1 This is a flowchart of the clinical data label correction method based on unsupervised learning and Bayesian probability proposed in this invention;

[0045] Figure 2 This is a schematic diagram of the system structure of the clinical data label correction method based on unsupervised learning and Bayesian probability proposed in this invention.

[0046] Figure 3 This diagram illustrates a comparison of the label correction effects of the clinical data label correction method based on unsupervised learning and Bayesian probability proposed in this invention. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0048] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0049] Reference Figure 1-3 A clinical data label correction method based on unsupervised learning and Bayesian probability.

[0050] Example 1: Self-learning revision of labels for a dataset of 54 clinical symptoms.

[0051] This embodiment uses a chronic disease follow-up dataset from the respiratory outpatient department of a tertiary hospital as its data source. The original dataset contains 3260 outpatients, collecting data on 54 respiratory-related subjective symptoms. There is no gold standard for pathological biopsy; the original non-gold standard labels are formed solely based on physician consultation records. The original labels contain numerous samples with contradictory subjective physician judgments. The complete workflow of this invention is used to correct the labels, with the following step-by-step details: Step 1: Data Preprocessing. The original dataset d:\Data Blue Ocean Project\test.RDS is read from the local path. The original data frame contains 15 columns of irrelevant patient basic information fields. The program automatically executes df<-df[,-c(2:15)] to remove redundant fields, retaining only the 54 symptom features and the original label columns. The dplyr package's standardization function is called to center and scale all symptom features, eliminating differences in the weighting of different symptoms, and outputting a standardized clean dataset work-df. This step differs from traditional manual data cleaning; it eliminates the need for manual field deletion, as the code automatically identifies and removes irrelevant variables. The preprocessing of 3260 samples takes less than 30 seconds. Step 2: K-means unsupervised clustering generates initial binary labels. A custom `kmeans-make-label` function is called to extract all 54 features from the data frame (excluding the label column) to form a feature matrix `mat`. The K-means cluster centers are set to `centers=2`. After clustering, `cluster=1` is mapped to label 1, and `cluster=2` is mapped to label 0, generating initial binary labels. These are then recombine to form a new data frame. The first column is the initial label generated by clustering, and the subsequent 54 columns are the original symptom features. In this example, approximately 18% of the initial labels output by clustering conflict with the physician's original manual labels, indicating a high proportion of noise in the initial non-gold standard labels and significant room for correction. Step 3: Lasso regression + stability selection filters highly stable core features. Using the initial labels generated by clustering as the dependent variable and 54 symptoms as independent variables, the glmnet package was loaded to perform Lasso regression. A stability selection process was adopted using 100 rounds of resampling, counting the frequency of non-zero coefficients for each feature. Highly stable variables with a frequency ≥ 0.7 were selected. In this embodiment, 10 core distinguishing symptoms were retained: persistent cough, nocturnal wheezing, shortness of breath after activity, chest tightness, hemoptysis, recurrent low-grade fever, chest pain, sleep apnea, decreased blood oxygen, and long-term smoking history. The remaining 44 low-stability symptoms with no distinguishing value were directly removed, and a simplified dataset containing only the labels and the 10 core features was reconstructed. After feature simplification, the computation time of each subsequent Bayesian iteration was reduced by 62%, while completely eliminating the interference of irrelevant symptoms on probability calculations and solving the problem of cluster boundary shift caused by multiple mixed features. The fourth step: Bayesian posterior probability parallel iterative optimization of labels, with convergence determined by cross-information entropy.The system employs a loop-based iterative logic. In each iteration, all remaining CPU cores (excluding the main core) are used to build a parallel cluster, splitting the task of calculating the posterior probability of 3260 samples into multi-threaded synchronous computations. For single-sample computation, a leave-one-out method is used: after removing the current sample, the mean values ​​of 10 features for both label 1 and label 0 are calculated. If the mean value of a feature is 0, a minimum smoothing value of 1e-10 is added to avoid meaningless logarithmic calculations. A custom Bayesian-posterior function is used to calculate the posterior probability of a single sample being positive or negative, generating a new posterior label with a 0.5 threshold. After each iteration, the mean cross-information entropy (mean-ent) of all samples is calculated, and the number of samples matching the label from the previous iteration (sn) is recorded. This loop continues until the difference between the mean-ent values ​​of two consecutive iterations is less than 0.001, at which point the label convergence is considered complete, and the iteration terminates. In this embodiment, a total of 17 iterations were performed to reach the convergence criterion. After the iteration, the converged labels were compared with the initial clustering labels, and 412 samples underwent label correction, achieving a noise label correction rate of 12.6%. Step 5: Joint Validation of Gain Correction Using Multiple Machine Learning Models. The converged and optimized simplified dataset is divided into training and test sets in an 8:2 ratio, with a fixed random seed of 123 to ensure consistent grouping. The fastml package is used to simultaneously construct three models: Radial Basis Functional VM (RBF) SVM, XGBoost, and Random Forest, with feature standardization and automatic encoding of categorical variables enabled uniformly. Confusion matrices are output on both the training and test sets, and model accuracy is calculated. A parallel control experiment is conducted using the uncorrected original labeled dataset. The average accuracy of the three models in the control group is 79%. After label correction using this invention, the average accuracy of the three models increases to 94%, an overall improvement of 15%, directly demonstrating that the label correction of this invention can significantly optimize model prediction performance. Simultaneously, model sensitivity and specificity data are extracted. After correction, sensitivity increases by 13%, specificity increases by 16%, and there is no overfitting. Step 6: Unified Calculation and Visualization of Posterior Probability Results. Based on the converged stable labels, the Bayesian posterior probabilities of all samples were recalculated. The ggplot2 plotting toolkit was used to execute the Comparison-charts-two-groups-distributions function, automatically generating two output charts: the first is a histogram of the label sample distribution before and after revision, visually demonstrating that the distribution of positive and negative samples after correction better reflects the actual clinical disease proportions; the second is an ROC curve of the three-class model test set, automatically annotating the AUC values, with the average AUC improving from 0.81 to 0.93 after correction. All charts are automatically saved to the local output folder, requiring no manual adjustment of plotting parameters, and can be directly used in clinical research reports.

[0052] Example 2: Label correction of a dataset of 20 laboratory biochemical test indicators.

[0053] The data source for this embodiment is a dataset of 1740 individuals from a diabetes screening population at a health checkup center. Twenty biochemical indicators, including fasting blood glucose, glycated hemoglobin, insulin, and blood lipids, were collected. There is no gold standard for pancreatic function pathology; the original non-gold standard labels are determined solely by single blood glucose thresholds. Due to fluctuations in single tests, the original labels contain a large number of misclassified samples. The complete implementation process is consistent with the framework of Embodiment 1, with differences only in the number of features, screening results, and model indicators. The key differences are detailed below: In the data preprocessing stage, the dataset is read, and irrelevant fields such as patient name and examination date are removed, retaining only the 20 continuous indicators. K-means binary clustering is used to generate initial labels. The original blood glucose threshold label conflicted with the initial clustering label in 21% of the samples. After Lasso stability screening, only 5 highly stable indicators were retained: glycated hemoglobin, fasting insulin, 2-hour postprandial blood glucose, triglycerides, and high-density lipoprotein cholesterol. The remaining 15 indicators with low correlation were removed. A total of 12 rounds of Bayesian parallel iteration reached the information entropy convergence standard, and 227 samples completed label correction. Model validation used only the XGBoost algorithm for control testing. The sensitivity of the original labeled model was 76%, which increased to 86% after correction, an improvement of 10%. The visualization module prominently outputs ROC curve comparison charts, clearly demonstrating the significant improvement in the model's discriminative ability after correction. This embodiment verifies that the present invention is not only applicable to discrete symptom data but also adaptable to continuous laboratory test indicators, demonstrating strong versatility and covering two major categories of mainstream non-gold standard data scenarios in clinical practice.

[0054] 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 clinical data label correction method based on unsupervised learning and Bayesian probability, characterized in that, This correction method is used for the automated correction of clinical datasets that lack pathological gold standards and rely solely on clinical symptoms / laboratory indicators to form non-gold standard noise labels. It includes the following steps performed sequentially: S1. Perform preprocessing on the original clinical dataset, removing irrelevant and redundant fields and completing feature standardization; S2. Based on the unsupervised K-means binary clustering algorithm, generate initial binary labels for the dataset based on all original clinical features; S3. Use Lasso regression combined with resampling stability selection to purify high-discrimination feature variables and construct a simplified feature subset; S4. Iteratively update the labels for each sample based on the leave-one-out Bayesian posterior probability, determine the label convergence state by the cross-information entropy fluctuation amplitude, terminate the iteration and output the corrected label. S5. Employs multi-class machine learning models such as SVM, XGBoost, and Random Forest for training and verification, and quantizes the performance improvement effect brought by label correction. S6. Calculate the Bayesian posterior probability of all samples in a unified manner, and automatically generate a label distribution comparison chart and model ROC curve to complete the visualization output of the correction results.

2. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, Step S2, unsupervised clustering to generate initial labels, specifically includes: Extract the pure feature matrix after preprocessing and removing irrelevant variables, set the number of K-means cluster centers to 2 to adapt to the clinical binary classification diagnosis scenario, map the cluster numbers one by one to binary 0 / 1 labels, and generate an uncorrected initial label set; The clustering process supports multi-core parallel computing acceleration and outputs an intermediate dataset that binds the initial labels and original features, which serves as the input data source for subsequent variable filtering.

3. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, Step S3, Lasso variable selection and stability screening, includes: A Lasso regression model is constructed using the initial labels output in step S2 as the dependent variable and all clinical features as independent variables. The frequency of non-zero regression coefficients of each feature was statistically analyzed using multiple rounds of resampling stability selection, and highly stable clinical features with frequencies higher than a preset threshold were retained. After removing low-stability noise features, a simplified feature dataset is reconstructed, feature dimension distillation is completed, the computational cost of subsequent Bayesian iterations is reduced, and the interference of invalid features on label determination is eliminated.

4. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, Step S4, the Bayesian posterior probability iterative optimization step, includes: The mean values ​​of positive and negative labeled samples were calculated separately using the leave-one-out method. When the mean was 0, a minimum smoothing value of 1e-10 was added to avoid logarithmic calculation failure. The posterior probability of a single sample belonging to two classes is solved in parallel, and the sample label is updated with a threshold of 0.

5. The average cross-entropy of the entire sample is calculated round by round. When the difference between the entropy of two consecutive rounds is lower than the preset threshold, the label is considered to have converged and the loop iteration is terminated.

5. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, The Bayesian posterior probability calculation process utilizes the device's multi-core resources to build a parallel computing cluster, splitting the probability calculation task of traversing all samples into multi-threaded synchronous operations. After each round of iteration, cluster resources are released, significantly reducing the overall time consumption of multiple rounds of iteration on large-scale clinical datasets.

6. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, Step S5 Model Validation uses three types of supervised learning models: Radial Basis Kernel SVM, XGBoost, and Random Forest. The dataset is divided into training and testing sets at a fixed ratio of 8:2, and standardized scaling features are used. The label correction gain effect is evaluated by a combination of three indicators: accuracy, sensitivity, and specificity.

7. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, Step S6 visualization output includes two core types of charts: First, a histogram comparing the distribution of sample labels before and after correction visually reflects the proportion of noise label correction. Second, the ROC curve output by the model validation quantifies the improvement in the model's discrimination ability. All charts are automatically generated in batches using the drawing toolkit, eliminating the need for manual adjustment of drawing parameters.

8. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, The method for performing the correction method according to any one of claims 1-7 comprises six functional modules that sequentially interact with data: The system includes a data preprocessing module, an unsupervised clustering label generation module, a Lasso stable feature selection module, a Bayesian label iterative optimization module, a multi-model validation module, and a result visualization output module.

9. The clinical data label correction method based on unsupervised learning and Bayesian probability according to claim 1, characterized in that, The data preprocessing module is used to read local clinical datasets, remove redundant and irrelevant fields, and perform feature standardization. The unsupervised clustering label generation module has built-in K-means binary clustering operation logic; The Lasso stable feature selection module is equipped with Lasso regression and resampling stability selection algorithms; The Bayesian label iterative optimization module is configured with a multi-core parallel computing interface and information entropy convergence judgment logic. The multi-model validation module integrates three types of machine learning algorithms: SVM, XGBoost, and Random Forest. The visualization output module has a built-in plotting toolkit that automatically generates label distribution comparison charts and ROC curves.

Images