Method and system for predicting chronic spontaneous urticarial connective tissue disease association

Abstract

The application relates to a machine learning-based chronic spontaneous urticaria connective tissue disease association prediction method and system, which comprises the following steps: acquiring multi-dimensional laboratory detection index data of a chronic spontaneous urticaria patient, performing missing value filling and standardization processing on the data, generating a derivative variable to form a multi-dimensional feature vector, inputting the multi-dimensional feature vector into a deep neural network model, extracting a high-order interaction relationship between indexes to generate a fusion feature, calculating and outputting a risk probability value of the patient combining a connective tissue disease based on the fusion feature, performing grade division according to a preset risk threshold, generating a prediction result and outputting a differentiated clinical screening suggestion. The method can accurately predict the risk of the patient combining a connective tissue disease and provide a scientific basis for clinical screening.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of medicine and information technology, and in particular to a method and system for predicting the association between chronic spontaneous urticaria and connective tissue disease. Background Technology

[0002] In the healthcare field, chronic spontaneous urticaria (CSU) is a common and complex skin disease whose pathogenesis involves abnormal activation of the immune system. Connective tissue diseases (CTDs), as a type of autoimmune disease, are closely related to immune system disorders. Clinically, it is not uncommon for CSU patients to also have CTDs, but currently, there is a lack of effective methods to predict the association between the two.

[0003] Existing predictive methods are primarily based on physicians' clinical experience and the analysis of a few single indicators. Physicians will make a preliminary assessment of the possibility of coexisting CTD based on the patient's symptoms and a few laboratory test indicators, such as antinuclear antibodies (ANA). However, this method has significant limitations. Analysis of a single indicator cannot comprehensively reflect the patient's overall immune status, easily leading to missed diagnoses or misdiagnoses.

[0004] The main problem with current solutions is the lack of comprehensive analysis of multi-dimensional laboratory testing indicators. Fluctuations in a single indicator may be influenced by multiple factors, failing to accurately reflect the intrinsic relationship between CSU and CTD, thus failing to provide accurate evidence for clinical diagnosis and treatment. Summary of the Invention

[0005] The main objective of this application is to provide a machine learning-based method and system for predicting the association between chronic spontaneous urticaria and connective tissue disease (CTD), which can comprehensively analyze multi-dimensional laboratory test indicators, accurately predict the risk of CSU patients with CTD, and provide a scientific basis for clinical screening.

[0006] To achieve the above objectives, embodiments of the present invention provide a machine learning-based method for predicting the association between chronic spontaneous urticaria and connective tissue disease, the method comprising the following steps: Obtain multidimensional laboratory test data of patients with chronic spontaneous urticaria. The laboratory test data includes thyroid peroxidase antibody Anti-TPO, antinuclear antibody ANA, and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4. The laboratory testing index data is imputed and standardized, and derived variables are generated by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables; the deep neural network model includes an input layer, at least two fully connected hidden layers and an output layer connected in sequence, wherein the output of the previous hidden layer is used as the input of the next hidden layer, which is used to pass through layers and abstract feature representation; The multidimensional feature vector is input into the deep neural network model. The high-order interaction relationship between Anti-TPO, ANA and five immune indicators is extracted step by step through nonlinear transformation between each hidden layer. Multi-level feature fusion is completed in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD. Based on the fusion features, the output layer of the deep neural network model calculates and outputs the risk probability value of CSU patients with connective tissue disease CTD; The risk probability values ​​are classified into levels based on preset risk thresholds, generating low-risk, medium-risk, or high-risk prediction results, and outputting corresponding differentiated clinical screening recommendations.

[0007] This invention also provides a machine learning-based predictive system for connective tissue diseases associated with chronic spontaneous urticaria, comprising: The acquisition module is used to acquire multi-dimensional laboratory test indicator data of patients with chronic spontaneous urticaria. The laboratory test indicator data includes thyroid peroxidase antibody Anti-TPO, antinuclear antibody ANA, and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4. The feature vector generation module is used to fill in missing values ​​and standardize the laboratory test index data, and generate derived variables by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables; wherein the deep neural network model includes an input layer, at least two fully connected hidden layers and an output layer connected in sequence, wherein the output of the previous hidden layer serves as the input of the next hidden layer, used to pass through and abstract feature representation layer by layer; The fusion feature processing module is used to input the multidimensional feature vector into the deep neural network model, extract the high-order interaction relationship between Anti-TPO, ANA and the five immune indicators through nonlinear transformation between each hidden layer, and complete the multi-level feature fusion in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD. The probability prediction module is used to calculate and output the risk probability value of CSU patients with connective tissue disease CTD based on the fusion features and the output layer of the deep neural network model. The results processing module is used to classify the risk probability value according to a preset risk threshold, generate low-risk, medium-risk or high-risk prediction results, and output differentiated clinical screening suggestions accordingly.

[0008] In summary, the technical solution of this application, by acquiring multi-dimensional laboratory test data of patients with chronic spontaneous urticaria (CSU), including thyroid peroxidase antibody, antinuclear antibody, and five immune indicators, can comprehensively reflect the patient's immune status. Missing value imputation and standardization of these data, along with the generation of derived variables to form a multi-dimensional feature vector, can eliminate dimensional differences in the data and uncover potential synergistic effects between indicators. Inputting the multi-dimensional feature vector into a deep neural network model can extract high-order interaction relationships between indicators at each level, generating fusion features characterizing the patient's risk of coexisting connective tissue diseases (CTDs). Based on the fusion features, risk probability values ​​are calculated and output, and then classified according to preset risk thresholds to generate prediction results and corresponding differentiated clinical screening recommendations. This improves the accuracy and reliability of predictions, provides clinicians with a more scientific basis for decision-making, helps in the early detection of CSU patients with CTDs, and thus enables timely and appropriate treatment measures to improve patient prognosis. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] To gain a more complete understanding of this application and its beneficial effects, the following description will be provided in conjunction with the accompanying drawings, wherein the same reference numerals in the following description denote the same parts.

[0011] Figure 1 This is a schematic diagram of a scenario for predicting the association between chronic spontaneous urticaria and connective tissue disease in the embodiments of this application; Figure 2 A flowchart is provided for the prediction method of connective tissue disease association in chronic spontaneous urticaria, as described in the embodiments of this application. Figure 3 This is a flowchart illustrating the multidimensional feature vector processing provided in an embodiment of this application. Figure 4 This is a schematic diagram of the process for generating fused feature vectors provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating the process of generating risk probabilities in an embodiment of this application. Figure 6This is a schematic diagram of the structure of the chronic spontaneous urticaria connective tissue disease association prediction system provided in the embodiments of this application; Figure 7 This is another structural schematic diagram of the chronic spontaneous urticaria connective tissue disease association prediction system provided in the embodiments of this application; Figure 8 This is another structural schematic diagram of the chronic spontaneous urticaria connective tissue disease association prediction system provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the chronic spontaneous urticaria connective tissue disease association prediction system provided in the embodiments of this application; Figure 10 A schematic diagram of the structure of a computer device provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of a deep neural network model provided in an embodiment of this application. Detailed Implementation

[0012] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0013] Obviously, the embodiments described below are only a part of the embodiments of this application, and not all of the embodiments; based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this application.

[0014] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0015] It should be noted that the concepts of "first" and "second" mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0016] Furthermore, the terms "first" and "second" used in this application should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more features.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] Figure 1 As shown, a scenario for predicting the association between chronic spontaneous urticaria and connective tissue disease is provided, based on machine learning. This scenario mainly includes a hospital information system, a laboratory information management system, a data processing server, a deep neural network model server, and a clinical diagnostic terminal. The hospital information system, laboratory information management system, data processing server, deep neural network model server, and clinical diagnostic terminal are connected via wired or wireless networks.

[0019] Taking a large tertiary hospital as an example, this hospital receives a large number of patients with chronic spontaneous urticaria (CSU) every day. The hospital's information system stores a massive amount of electronic medical records for patients, including basic information, symptoms, diagnoses, and laboratory test requests. For instance, when a 35-year-old female CSU patient comes to the hospital, the doctor uses the information system to screen electronic medical records that meet the diagnostic criteria for CSU, such as wheals accompanied by itching, occurring at least twice a week for more than six weeks, and excluding other known causes of urticaria. After selecting this patient's electronic medical record, the doctor extracts the corresponding laboratory test requests, which detail the tests requested by the patient, including thyroid peroxidase antibody (Anti-TPO), antinuclear antibody (ANA), and five immune markers (immunoglobulins IgG, IgA, IgM, complement C3, and complement C4).

[0020] The laboratory information management system stores the patient's laboratory test results. Based on the test request information retrieved from the hospital information system, the corresponding test results are retrieved from the laboratory information management system. For example, for the aforementioned 35-year-old female patient, the test results for Anti-TPO, ANA, and five immune markers are retrieved. After obtaining the data, a source consistency verification is required. This is because serum samples collected at different times may be affected by the patient's physical condition, environmental factors, etc., leading to differences in test results. If the indicators do not come from serum sample test reports from the same sampling time point, the sample record will be discarded. After rigorous consistency verification, it was confirmed that all of the patient's indicators came from test reports from the same time point, and the data was valid.

[0021] The data processing server is used to process the acquired data. First, the laboratory test results that have passed consistency checks are structured and categorized according to indicator type, establishing data subsets for the Anti-TPO field, ANA field, and the five immunological markers field, and labeling the corresponding data unit and detection method type for each field. For example, the data unit for Anti-TPO is IU / ml, and the detection method type is chemiluminescence immunoassay; the data unit for ANA is titer, and the detection method type is indirect immunofluorescence immunoassay. Next, preliminary outlier identification is performed on the categorized indicator data, determining whether there are significant outliers based on the upper and lower limits of the medical reference range. For example, the normal reference range for immunoglobulin IgG is 7.0-16.0 g / L. If a patient's IgG test result is 20.0 g / L, significantly exceeding the normal range, the data will be manually reviewed or automatically marked as pending processing. Then, the indicator data that have completed preliminary outlier identification are integrated into a raw data table with a unified format, ensuring that each record contains all ten indicator fields without missing identifiers.

[0022] For missing data items in the original indicators, the data processing server uses an imputation strategy based on the mean of a similar patient group to fill in the missing data. For example, if a patient's immunoglobulin IgA data is missing, other patients of the same sex, age group, and underlying disease status are selected from the database, their average immunoglobulin IgA value is calculated, and this average value is used to replace the missing value. Then, Z-score standardization is performed on all imputed original indicators to convert them into a standard normal distribution with a mean of zero and a variance of one, thus eliminating the impact of dimensional differences on model training. Based on the standardized original indicators, several clinically meaningful ratio-type derived variables are constructed, such as the combined ratios of IgG / IgA, C3 / C4, and Anti-TPO / ANA, and multiple cross-product terms are generated simultaneously to reflect the potential synergistic effects between indicators. All original indicators and newly generated derived variables are merged to form an initial feature set, and a redundancy assessment is performed on this set. Redundant features that are highly linearly correlated with other variables are removed, and the remaining feature combinations are determined as the final multidimensional feature vector.

[0023] The deep neural network model server receives multidimensional feature vectors output from the data processing server. The deep neural network model consists of an input layer, at least two fully connected hidden layers, and an output layer. The multidimensional feature vectors are input to the input layer, and high-order interactions between Anti-TPO, ANA, and the five immune indicators are extracted level by level through nonlinear transformations between the hidden layers. For example, in the first hidden layer, a primary abstract feature representation is output through weight matrix operations and nonlinear activation function processing; in the second hidden layer, the primary abstract features are used as input, and weighted summation and nonlinear mapping operations are performed again to obtain intermediate feature representations that reflect the second-order interactions between some indicators. As the number of hidden layers increases, the model continuously uncovers more complex interactions. In the last hidden layer, cross-level feature integration is achieved through global pooling or concatenation operations to generate fused features characterizing the risk of CSU patients with connective tissue disease (CTD). Based on the fused features, the output layer calculates and outputs the risk probability value of CSU patients with CTD.

[0024] The clinical diagnostic terminal receives risk probability values ​​output by the deep neural network model server. Based on preset risk thresholds, it categorizes these risk probability values ​​into low-risk, medium-risk, or high-risk predictions, and outputs corresponding differentiated clinical screening recommendations. For example, if a patient's risk probability value falls within the high-risk range, the clinical diagnostic terminal will recommend further specialized examinations, such as tests for anti-double-stranded DNA antibodies and anti-Sm antibodies, as well as skin biopsies and joint imaging examinations, to determine whether a coexisting CTD (concomitant lesioning disease) is present.

[0025] In the above scenario, accurate prediction of the risk of CSU patients with CTD was achieved, providing a scientific basis for clinical diagnosis and treatment and helping to improve the patient's treatment outcome.

[0026] refer to Figure 2 , Figure 2 This is a flowchart illustrating a method for predicting the association between chronic spontaneous urticaria and connective tissue disease, as provided in an embodiment of this application. The method can be executed by computer equipment, such as a server. Specifically, the method for predicting the association between chronic spontaneous urticaria and connective tissue disease provided in this embodiment includes: S10: Obtain multidimensional laboratory test data for patients with chronic spontaneous urticaria (CSU). The laboratory test data includes thyroid peroxidase antibody (Anti-TPO), antinuclear antibody (ANA), and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4.

[0027] In this embodiment, thyroid peroxidase antibody (Anti-TPO) is an autoantibody targeting thyroid peroxidase, and its level changes can reflect the autoimmune status of the thyroid gland. Anti-TPO levels are often elevated in thyroid diseases. For example, in patients with Hashimoto's thyroiditis, Anti-TPO levels are usually significantly higher than the normal range. Antinuclear antibody (ANA) is a group of autoantibodies targeting components within the cell nucleus; ANA detection is of great significance for the diagnosis of autoimmune diseases. ANA positivity is commonly seen in various connective tissue diseases such as systemic lupus erythematosus and Sjögren's syndrome. Immunoglobulins IgG, IgA, and IgM, among the five immune markers, are important components of the human immune system. Complement C3 and complement C4 are important components of the complement system, which participates in immune defense and immune regulation. When the human immune system malfunctions, the levels of these five immune markers will change accordingly.

[0028] In one embodiment, step S11 can be implemented as follows: A1: Screen electronic medical records of patients who meet the diagnostic criteria for chronic spontaneous urticaria from the hospital information system, extract the corresponding laboratory test request information, and retrieve the corresponding laboratory test result data based on the request information.

[0029] In this embodiment, the hospital information system is a database storing patients' electronic medical records and related testing information. Screening patient electronic medical record records that meet the diagnostic criteria for chronic spontaneous urticaria requires specific diagnostic criteria. For example, according to the "Chinese Guidelines for the Diagnosis and Treatment of Chronic Urticaria," a patient presenting with wheals accompanied by itching, occurring at least twice a week for more than six weeks, and excluding other known causes of urticaria, can be diagnosed with chronic spontaneous urticaria. By setting corresponding filtering conditions in the hospital information system, such as diagnostic information and onset time, eligible patient electronic medical record records can be accurately screened. Extracting the corresponding laboratory test request information is to obtain information such as the laboratory tests performed by the patient and the testing time, so as to retrieve the test results data later.

[0030] For example, in a hospital's information system, by setting the filter criteria "diagnosed with chronic spontaneous urticaria and onset within the last three months," 100 patient electronic medical record records were selected. Laboratory test request information was extracted from these records, including the requested tests (such as Anti-TPO, ANA, and the five immune markers) and the test dates. Based on this request information, the corresponding test results were retrieved from the laboratory information management system. This ensures that the data obtained is specific to eligible patients and that the data source is accurate and reliable.

[0031] A2: Verify the consistency of the extracted laboratory test results data to confirm whether Anti-TPO, ANA, and the five immune indicators all come from the serum sample test reports at the same sampling time point. If there is a time discrepancy, the sample record will be removed.

[0032] In this embodiment, source consistency verification is performed to ensure the accuracy and reliability of the data. Serum samples collected at different time points may be affected by the patient's physical condition, environmental factors, etc., leading to differences in test results. Therefore, it is necessary to confirm whether the Anti-TPO, ANA, and five immune indicators all originate from serum sample test reports from the same sampling time point.

[0033] For example, in the obtained test results data, it was found that a patient's Anti-TPO test was conducted on January 1, 2024, while the ANA and five immune markers were tested on January 10, 2024, resulting in a time discrepancy. In such cases, this sample record may affect subsequent analysis results and needs to be removed.

[0034] A3: The laboratory test results that have passed the consistency verification are structured and classified according to the indicator categories. Data subsets of the Anti-TPO field, ANA field and five immune items field are established respectively, and the data units and test method types corresponding to each field are marked.

[0035] In this embodiment, structured classification facilitates data management and analysis. Laboratory test results that have passed consistency checks are categorized according to indicator types, establishing data subsets for the Anti-TPO field, ANA field, and the five immune markers field. Each data subset contains all test results for that indicator. Labeling the data unit and test method type for each field ensures data accuracy and understandability. Different test methods may lead to differences in test results; clearly defining the test method type allows for better data analysis and comparison.

[0036] For example, for the Anti-TPO index, a subset of data for the Anti-TPO field is created, containing Anti-TPO test results for all patients. The data unit for this field is labeled as IU / ml, and the detection method is chemiluminescence immunoassay. Similarly, the ANA index and the five immune markers are categorized and labeled accordingly.

[0037] In one embodiment, a database management system can be used to structure and classify the data. Different tables are created to store subsets of data from the Anti-TPO field, the ANA field, and the five immune markers field. In each table, fields for data units and detection method types are added, and the corresponding information is entered. This facilitates data management and querying, improving the efficiency of data analysis.

[0038] A4: Perform preliminary outlier identification on the categorized indicator data, determine whether there are obvious outliers based on the upper and lower limits of the medical reference range, and manually review or automatically mark suspected erroneously entered data as pending processing.

[0039] In this embodiment, preliminary outlier identification aims to discover anomalies in the data. The upper and lower limits of the medical reference range are normal ranges determined based on extensive clinical data and medical research. By comparing each indicator's data with these limits, it can be determined whether there are significant outliers. For data suspected of being incorrectly entered, manual review or automatic marking as pending processing is required to ensure data accuracy.

[0040] For example, the medical reference range for immunoglobulin IgG is 7.0-16.0 g / L. If a patient's IgG test result is 25.0 g / L, significantly exceeding the reference range, this value may be considered abnormal. In such cases, manual review can be performed to check for data entry errors in the test report, or it can be automatically marked as pending further processing.

[0041] A5: Integrate the Anti-TPO, ANA, and five immune indicators data that have completed the initial outlier identification into a raw data table with a unified format, ensuring that each record contains all ten indicator fields and has no missing identifiers.

[0042] In this embodiment, the data is integrated into a unified format of raw data tables to facilitate subsequent data analysis and processing. Ensuring that each record contains all ten indicator fields (Anti-TPO, ANA, immunoglobulin IgG, IgA, IgM, complement C3, complement C4) without missing markers guarantees data integrity and consistency.

[0043] For example, the data on Anti-TPO, ANA, and five immune indicators, after initial outlier identification, are organized into a standardized format, with each record containing ten indicator fields. Data with missing values ​​are processed accordingly, such as using imputation strategies based on the mean of similar patient groups.

[0044] S20: The laboratory test index data is imputed and standardized, and derived variables are generated by constructing ratios and cross-product terms between indicators to form a multidimensional feature vector containing the original indicators and derived variables; wherein the deep neural network model includes an input layer, at least two fully connected hidden layers and an output layer connected in sequence, wherein the output of the previous hidden layer is used as the input of the next hidden layer, which is used to pass through and abstract feature representation layer by layer.

[0045] In this embodiment, missing value imputation is used to ensure data integrity. During actual data collection, some indicator data may be missing due to various reasons. Imputation is performed using a mean-based strategy for similar patient groups, replacing missing values ​​with the average values ​​of corresponding indicators from other patients of the same sex, age group, and underlying disease state. For example, for a male CSU patient with missing immunoglobulin IgG data, the average immunoglobulin IgG value from other male patients of the same age group and with the same underlying disease state can be calculated and used to replace the missing value. Standardization is used to eliminate dimensional differences in the data and make the indicators comparable. Z-score standardization converts each indicator into a standard normal distribution with a mean of zero and a variance of one. Constructing ratios and cross-product terms between indicators to generate derived variables can uncover potential synergistic effects between indicators. For example, the IgG / IgA ratio can reflect the relative strength of humoral immune responses.

[0046] S30: Input the multidimensional feature vector into the deep neural network model, extract the high-order interaction relationship between Anti-TPO, ANA and the five immune indicators through nonlinear transformation between each hidden layer, and complete the multi-level feature fusion in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD.

[0047] In this embodiment, the deep neural network model is a powerful machine learning model consisting of an input layer, hidden layers, and an output layer. Neurons in the hidden layers undergo nonlinear transformations through nonlinear activation functions, enabling them to learn complex patterns and features in the data. The nonlinear transformations between hidden layers can progressively extract higher-order interactions between Anti-TPO, ANA, and the five immune indicators. For example, in the first hidden layer, through weight matrix operations and nonlinear activation function processing, a primary abstract feature representation is output, which contains simple relationships between some indicators. As the number of hidden layers increases, the model can continuously uncover more complex interactions. In the last hidden layer, cross-level feature integration is achieved through global pooling or concatenation operations to generate fused features. These fused features can comprehensively reflect the risk of a patient with coexisting CTD.

[0048] S40: Based on the fusion features, the output layer of the deep neural network model calculates and outputs the risk probability value of CSU patients with connective tissue disease CTD.

[0049] In this embodiment, the fusion feature is a comprehensive feature obtained after processing by a deep neural network model. It contains high-order interaction information between Anti-TPO, ANA, and the five immune indicators. The output layer converts the fusion feature into a risk probability value. The fusion feature vector is mapped to a single real value through a linear transformation unit, and then a sigmoid function is applied for compression transformation to make it fall within the range of 0 to 1, serving as a preliminary risk probability estimate for CSU patients with connective tissue disease CTD.

[0050] In one embodiment, the fused feature vector output from the last hidden layer is input to the linear transformation unit of the output layer, and mapped to a single real value using a learnable parameter matrix. For example, the learnable parameter matrix and the fused feature vector are multiplied to obtain a single real value. This real value is then compressed using the sigmoid function, making it fall within the range of 0 to 1. The expression for the sigmoid function is: y = 1 / (1 + e^(-x)), where x is the input real value and y is the output probability value. The bias term of the output layer is fine-tuned based on the label distribution in historical training samples to correct probability bias caused by class imbalance. During the model inference phase, the pre-trained output layer parameters are fixed, and only new fused features are received as input, generating the corresponding risk probability output in real time. The generated risk probability value is temporarily stored in a temporary buffer for subsequent processing and analysis. This allows for accurate calculation of the risk probability of a patient with CTD, providing a basis for clinical decision-making.

[0051] S50: Classify the risk probability value according to the preset risk threshold, generate low-risk, medium-risk or high-risk prediction results, and output differentiated clinical screening suggestions accordingly.

[0052] In this embodiment, the preset risk threshold is set based on a large amount of clinical data and experience. By comparing the risk probability value with the risk threshold, the patient's risk level can be classified as low risk, medium risk, or high risk. Different risk levels correspond to different clinical screening recommendations. For example, low-risk patients may only need to undergo routine examinations regularly, medium-risk patients may need to undergo some further specialized examinations, and high-risk patients need to undergo more comprehensive and in-depth examinations to determine whether they have coexisting CTD (concurrent lesions).

[0053] In one embodiment, a low-risk threshold of 0.2 and a medium-risk threshold of 0.5 are set. When a patient's risk probability value is less than 0.2, they are considered low-risk, and it is recommended that they undergo routine laboratory tests every six months, including Anti-TPO, ANA, and five immune markers. When the risk probability value is between 0.2 and 0.5, they are considered medium-risk, and it is recommended that they undergo specific tests such as anti-double-stranded DNA antibody and anti-Sm antibody tests within one month. When the risk probability value is greater than 0.5, they are considered high-risk, and it is recommended that they immediately undergo comprehensive connective tissue disease-related examinations, such as skin biopsy and joint imaging. This allows for personalized clinical screening recommendations based on the patient's risk level, improving the accuracy and efficiency of diagnosis.

[0054] In one embodiment, reference Figure 3 Step S20 may include steps S21-S24, which will be described in detail below: S21: For missing data items in the original indicators, a strategy based on the mean of the same patient group is used to fill them in, that is, the missing values ​​are replaced by the average value of the corresponding indicators of other patients of the same gender, age group and underlying disease status.

[0055] For example, for a 40-year-old male CSU patient, immunoglobulin IgM data is missing. Other patients of the same sex (male), age group (35-45 years), and with CSU can be screened from the database, and their average immunoglobulin IgM levels can be calculated. Assuming the average is 1.5 g / L, this average can be used to replace the missing value for this patient.

[0056] S22: Standardize all the original indicators after imputation, so that each indicator is converted into a standard normal distribution with a mean of zero and a variance of one, in order to eliminate the influence of dimensional differences on model training.

[0057] In this embodiment, Z-score standardization can be used. By converting each indicator into a standard normal distribution with a mean of zero and a variance of one, the dimensional differences in the data can be eliminated. Different indicators may have different units and magnitudes, and these dimensional differences can affect the training effect of the model. Standardization makes the indicators comparable, improving the training efficiency and accuracy of the model.

[0058] For example, the raw data for immunoglobulin IgG has a mean of 10.0 g / L and a standard deviation of 2.0 g / L. If a patient's IgG test result is 12.0 g / L, after Z-score standardization, its value is (12.0 - 10.0) / 2.0 = 1.0. This converts data from different magnitudes into a unified standard form.

[0059] S23: Based on the standardized original indicators, construct several clinically significant ratio-type derived variables, including but not limited to the combination ratios of IgG / IgA, C3 / C4, and Anti-TPO / ANA, and simultaneously generate multiple cross-product terms to reflect the potential synergistic effect between indicators.

[0060] In this embodiment, constructing ratio-based derived variables and cross-product terms can uncover potential synergistic effects between indicators. Ratio-based derived variables can reflect the relative relationships between different indicators, while cross-product terms can capture nonlinear synergistic change patterns between indicators. For example, the IgG / IgA ratio can reflect the relative strength of the humoral immune response, and the C3 / C4 ratio can reflect the balance of the complement activation pathway.

[0061] For example, for standardized immunoglobulin IgG and IgA values, the IgG / IgA ratio is calculated. Assuming a patient's standardized IgG value is 1.2 and their standardized IgA value is 0.8, the IgG / IgA ratio is 1.2 / 0.8 = 1.5. Simultaneously, any two standardized raw indicators from different categories are multiplied pairwise on a sample-by-sample basis to generate a cross-product term.

[0062] In one embodiment, step S23 can be implemented as follows: B1: Perform element-wise division on the standardized immunoglobulin IgG and IgA values ​​to generate the IgG / IgA ratio sequence, which is used to characterize the relative strength of the humoral immune response.

[0063] In this embodiment, immunoglobulins IgG and IgA are important components of humoral immunity. The IgG / IgA ratio can reflect the relative strength of the humoral immune response. A high IgG / IgA ratio may indicate that the humoral immune response is dominated by IgG; a low ratio may indicate that IgA plays a more important role in humoral immunity.

[0064] For example, three patients have standardized IgG values ​​of 1.2, 1.5, and 1.8, and standardized IgA values ​​of 0.8, 1.0, and 1.2, respectively. Their IgG / IgA ratios are 1.2 / 0.8 = 1.5, 1.5 / 1.0 = 1.5, and 1.8 / 1.2 = 1.5, respectively. Calculating the IgG / IgA ratio sequence provides a more intuitive understanding of the patient's humoral immune response status.

[0065] B2: The standardized values ​​of complement C3 and complement C4 are calculated on a sample-by-sample basis to form the C3 / C4 ratio sequence, which is used to reflect the balance of the complement activation pathway.

[0066] In this embodiment, complement C3 and complement C4 are important components of the complement system. The balance of the complement activation pathway is crucial for maintaining human immune function. The C3 / C4 ratio can reflect the balance of the complement activation pathway. An abnormal C3 / C4 ratio may indicate an abnormality in the complement activation pathway.

[0067] For example, three patients had standardized complement C3 values ​​of 1.1, 1.3, and 1.5, and standardized complement C4 values ​​of 0.9, 1.1, and 1.3, respectively. Their C3 / C4 ratios are approximately 1.1 / 0.9 ≈ 1.22, 1.3 / 1.1 ≈ 1.18, and 1.5 / 1.3 ≈ 1.15, respectively. Calculating the C3 / C4 ratio sequence allows monitoring of the balance of complement activation pathways.

[0068] B3: Perform a sample-by-sample division operation on the standardized values ​​of thyroid peroxidase antibody Anti-TPO and antinuclear antibody ANA to obtain the Anti-TPO / ANA ratio sequence, which is used to indicate the targeting bias of autoimmune responses.

[0069] In this embodiment, thyroid peroxidase antibody (Anti-TPO) and antinuclear antibody (ANA) are important indicators in autoimmune diseases. The Anti-TPO / ANA ratio sequence can indicate the targeting bias of the autoimmune response. A high Anti-TPO / ANA ratio may indicate that the autoimmune response is more inclined to target thyroid tissue; a low ratio may indicate that the autoimmune response is more inclined to target nuclear components.

[0070] For example, three patients have standardized Anti-TPO values ​​of 1.2, 1.5, and 1.8, and standardized ANA values ​​of 0.8, 1.0, and 1.2, respectively. Their Anti-TPO / ANA ratios are 1.2 / 0.8 = 1.5, 1.5 / 1.0 = 1.5, and 1.8 / 1.2 = 1.5, respectively. By calculating the Anti-TPO / ANA ratio sequence, the target preference of the autoimmune response can be understood.

[0071] B4: Multiply any two standardized raw indices from different categories pairwise on a sample-by-sample basis to generate a set of all possible cross-product terms. This set is used to capture non-linear collaborative change patterns between indices.

[0072] In this embodiment, cross-product terms can capture nonlinear synergistic change patterns between indicators. Complex interactions may exist between different indicators; by generating a set of cross-product terms, these potential synergistic effects can be explored. For example, the cross-product term of immunoglobulin IgG and antinuclear antibody ANA can reflect the synergistic relationship between these two indicators.

[0073] For example, three patients have standardized immunoglobulin IgG values ​​of 1.2, 1.5, and 1.8, and standardized antinuclear antibody (ANA) values ​​of 0.8, 1.0, and 1.2, respectively. Their IgG*ANA cross-product terms are then 1.2*0.8=0.96, 1.5*1.0=1.5, and 1.8*1.2=2.16, respectively. By generating the set of all possible cross-product terms, the nonlinear cooperative variation patterns among the indicators can be comprehensively captured.

[0074] B5: Perform clinical interpretability screening on all generated ratio variables and cross-product terms, eliminating derived variables lacking medical logical support or physiological mechanism association, and retaining feature subsets with clear pathophysiological significance.

[0075] For example, among the generated ratio variables and cross-product terms, if a ratio variable is found to be unrelated to known medical knowledge and physiological mechanisms, such as a randomly combined ratio variable, this variable should be removed, retaining only feature subsets with clear pathophysiological significance, such as IgG / IgA and C3 / C4.

[0076] B6: The retained ratio variables and cross-product terms are uniformly incorporated into the feature engineering process as candidate derived features for subsequent redundancy assessment and model input.

[0077] For example, ratio variables such as IgG / IgA, C3 / C4, and Anti-TPO / ANA, along with related cross-product terms, are incorporated into the feature engineering process. In subsequent redundancy assessments, these candidate derived features are analyzed, and redundant features that are highly linearly correlated with other variables are removed.

[0078] S24: Combine all original indicators with newly generated derived variables to form an initial feature set, and perform a redundancy assessment on the set. Eliminate redundant features that are highly linearly correlated with other variables to reduce dimensionality complexity. Determine the combination of features retained after elimination as the final multidimensional feature vector.

[0079] In this embodiment, all original indicators and newly generated derived variables are merged to form an initial feature set, which allows for the comprehensive utilization of information from both the original indicators and derived variables. Redundancy assessment of this set is performed to reduce dimensionality complexity and avoid model overfitting. Redundant features that are highly linearly correlated with other variables increase model complexity and contribute little to the model's predictive ability. By removing these redundant features, the training efficiency and accuracy of the model can be improved.

[0080] For example, in the initial feature set, a high linear correlation was found between the cross-product term of immunoglobulins IgG and IgA and the IgG / IgA ratio. Through redundancy assessment, the cross-product term was removed, retaining only the IgG / IgA ratio.

[0081] In one embodiment, reference Figure 4 Step S30 may include steps S31-S35, which will be described in detail below: S31: The preprocessed multidimensional feature vector is loaded into the input layer of the deep neural network, and then processed by the weight matrix operation and nonlinear activation function of the first hidden layer in sequence to output the primary abstract feature representation.

[0082] In this embodiment, the preprocessed multidimensional feature vector contains information about the original indicators and derived variables. This vector is then loaded into the input layer of the deep neural network as the model's input. Weight matrix operations in the first hidden layer can linearly combine the input multidimensional feature vector, and nonlinear activation function processing can introduce nonlinear factors, enabling the model to learn complex patterns in the data.

[0083] For example, suppose the multidimensional feature vector is [x1, x2, x3], and the weight matrix of the first hidden layer is [[w11, w12, w13], [w21, w22, w23], [w31, w32, w33]]. After weight matrix operations, a linear combination result [y1, y2, y3] is obtained, where y1 = w11*x1 + w12*x2 + w13*x3, y2 = w21*x1 + w22*x2 + w23*x3, and y3 = w31*x1 + w32*x2 + w33*x3. Then, a non-linear activation function (such as the ReLU function) is applied to [y1, y2, y3] to obtain a primary abstract feature representation.

[0084] S32: The primary abstract features are used as input to the second hidden layer, and weighted summation and nonlinear mapping operations are performed again to obtain intermediate feature representations that can reflect the second-order interactions between some indicators.

[0085] In this embodiment, the primary abstract features contain simple relationships between some indicators. Using these as input to the second hidden layer, and performing weighted summation and nonlinear mapping operations again, allows for the discovery of second-order interactions between indicators. These second-order interactions can reflect more complex relationships between indicators, helping to improve the model's predictive ability.

[0086] For example, suppose the initial abstract features are [z1, z2, z3], and the weight matrix of the second hidden layer is [[u11, u12, u13], [u21, u22, u23], [u31, u32, u33]]. After weighted summation, we get [v1, v2, v3], where v1 = u11*z1 + u12*z2 + u13*z3, v2 = u21*z1 + u22*z2 + u23*z3, and v3 = u31*z1 + u32*z2 + u33*z3. Then, we apply a non-linear activation function to [v1, v2, v3] to obtain the intermediate feature representation.

[0087] In one embodiment, step S32 can be implemented in the following manner: P1: Perform matrix multiplication on the primary abstract feature vector output by the first hidden layer and the learnable weight matrix of the second hidden layer to generate a weighted sum, which represents the linear combination strength of each primary feature at the current level.

[0088] In this embodiment, the primary abstract feature vector output by the first hidden layer contains simple feature information between some indicators. The learnable weight matrix of the second hidden layer is used to linearly combine the primary abstract feature vectors. Through matrix multiplication, a weighted sum is generated, which represents the strength of the linear combination of each primary feature at the current level.

[0089] For example, suppose the primary abstract feature vector output by the first hidden layer is [a1,a2,a3], and the learnable weight matrix of the second hidden layer is [[b11,b12,b13],[b21,b22,b23],[b31,b32,b33]]. After matrix multiplication, we get the weighted sum [c1,c2,c3], where c1=b11*a1+b12*a2+b13*a3, c2=b21*a1+b22*a2+b23*a3, and c3=b31*a1+b32*a2+b33*a3.

[0090] P2: Apply a preset nonlinear activation function to the weighted summation result to perform element-wise transformation, and output the intermediate activation value after nonlinear mapping. This activation value is used to enhance the model's ability to express complex interaction relationships.

[0091] In this embodiment, the preset nonlinear activation function introduces nonlinear factors, enabling the model to learn complex patterns and interactions in the data. Applying a nonlinear activation function to the weighted sum result for element-wise transformation converts the linear combination result into intermediate activation values ​​with nonlinear characteristics. For example, a common nonlinear activation function is ReLU (Rectified Linear Unit), whose expression is given by [expression missing]. When an element in the weighted sum result is positive, it remains unchanged after the ReLU transformation; when the element is negative, it becomes 0 after the transformation. This transformation enhances the model's ability to express complex interactions, preventing the model from only learning linear relationships.

[0092] For example, assuming the weighted summation result is [3,-2,5], after performing an element-wise transformation using the ReLU function, the intermediate activation value obtained is [3,0,5].

[0093] In one embodiment, the activation function module in a deep learning framework can be used to implement a nonlinear transformation of the weighted summation result. Taking TensorFlow as an example, by calling the corresponding activation function interface and taking the weighted summation result as input, intermediate activation values ​​after nonlinear mapping can be obtained. This facilitates nonlinear transformation and improves the model's ability to learn complex interaction relationships.

[0094] P3: The intermediate activation values ​​are grouped into three sub-feature groups based on the original source category of the indicators: thyroid-related, nuclear antibody-related, and immunoglobulin-complement-related. Each sub-feature group corresponds to a pathological mechanism dimension.

[0095] In this embodiment, intermediate activation values ​​are grouped by feature to better analyze feature information under different pathological mechanism dimensions. The thyroid-related sub-feature group includes activation dimensions related to thyroid peroxidase antibody (Anti-TPO) and its derived variables; the nuclear antibody-related sub-feature group includes activation dimensions related to antinuclear antibody (ANA) and its derived variables; and the immunoglobulin-complement-related sub-feature group includes activation dimensions related to the five immune indicators and their derived variables. Each sub-feature group corresponds to a pathological mechanism dimension, which helps to deepen the understanding of the relationships between different pathological mechanisms.

[0096] For example, among the intermediate activation values, the activation dimensions corresponding to Anti-TPO and its derived variables are classified into the thyroid-related sub-feature group, the activation dimensions corresponding to ANA and its derived variables are classified into the nuclear antibody-related sub-feature group, and the activation dimensions corresponding to the five immune indicators and their derived variables are classified into the immunoglobulin-complement-related sub-feature group.

[0097] In one embodiment, step P3 may specifically include the following: P31: Analyze the original index source information of the input features of the second hidden layer and establish a mapping index table from hidden layer neurons to original laboratory indicators.

[0098] In this embodiment, parsing the original indicator source information of the input features of the second hidden layer is to clarify the original laboratory indicator corresponding to each hidden layer neuron. By establishing a mapping index table from hidden layer neurons to original laboratory indicators, the information source of each neuron can be clearly understood, providing a basis for subsequent feature grouping and analysis.

[0099] For example, assuming the second hidden layer has 10 neurons, by parsing the original indicator source information of the input features, it is found that neuron 1 corresponds to the immunoglobulin IgG indicator, neuron 2 corresponds to the antinuclear antibody ANA indicator, and so on. These correspondences are recorded in a mapping index table.

[0100] P32: Assign each dimension in the intermediate activation values ​​to the corresponding pathological mechanism category according to the mapping index table to complete the feature attribution labeling.

[0101] For example, according to the mapping index table, the intermediate activation value dimension corresponding to thyroid peroxidase antibody Anti-TPO and its derived variables is labeled as thyroid-related, the dimension corresponding to antinuclear antibody ANA and its derived variables is labeled as nuclear antibody-related, and the dimension corresponding to the five immune indicators and their derived variables is labeled as immunoglobulin-complement-related.

[0102] P33: The activation dimensions of the thyroid peroxidase antibody Anti-TPO and its derived variables are classified into the thyroid-related sub-feature group.

[0103] In this embodiment, the activation dimension of the thyroid peroxidase antibody Anti-TPO and its derived variables is included in the thyroid-related sub-feature group to integrate thyroid-related feature information. The thyroid-related sub-feature group can reflect the immune status of the thyroid and related pathological mechanisms.

[0104] For example, in the intermediate activation values, after feature attribution labeling, the activation dimensions corresponding to Anti-TPO and its derived variables are extracted to form thyroid-related sub-feature groups.

[0105] P34: The activation dimensions of the antinuclear antibody (ANA) and its derived variables are classified into the ANA-related sub-feature group.

[0106] For example, in the intermediate activation values, after feature attribution labeling, the activation dimensions corresponding to ANA and its derived variables are extracted to form a nuclear antibody-related sub-feature group.

[0107] P35: The activation dimensions of the five immune indicators and their derived variables are classified into the immunoglobulin-complement related sub-feature group.

[0108] For example, in the intermediate activation values, after feature attribution labeling, the activation dimensions corresponding to the five immune indicators (immunoglobulins IgG, IgA, IgM, complement C3, and complement C4) and their derived variables are extracted to form an immunoglobulin-complement related sub-feature group.

[0109] P4: Calculate the pairwise product of features within each sub-feature group to generate a second-order interaction term within the group. This interaction term is used to explicitly model the synergistic or antagonistic effects between similar pathological factors.

[0110] In this embodiment, the pairwise multiplication of features within each sub-feature group is calculated to generate second-order interaction terms within the group. This is to explicitly model the synergistic or antagonistic effects between similar pathological factors. Complex relationships may exist between similar pathological factors, and generating second-order interaction terms can capture these relationships.

[0111] For example, in the thyroid-related sub-feature group, there are features A, B, and C. Second-order interaction terms such as A*B, A*C, and B*C can be calculated. These second-order interaction terms can reflect the synergistic or antagonistic effects between thyroid-related pathological factors.

[0112] P5: The second-order interaction terms within the group are concatenated and fused with the original intermediate activation values ​​to form an expanded intermediate feature representation vector, which contains both first-order abstract information and local second-order interaction information.

[0113] In this embodiment, the intra-group second-order interaction terms are concatenated and fused with the original intermediate activation values ​​to form an expanded intermediate feature representation vector. This is to comprehensively utilize first-order abstract information and local second-order interaction information. First-order abstract information reflects the basic characteristics of the features, while local second-order interaction information reflects the complex relationships between features. Combining the two can improve the expressive power of the features.

[0114] For example, the intra-group second-order interaction terms of the thyroid-related sub-feature group are concatenated with the original intermediate activation values ​​to form an expanded intermediate feature expression vector. Assuming the original intermediate activation values ​​have 5 dimensions, the intra-group second-order interaction terms have 3 dimensions, and the resulting expanded intermediate feature expression vector has 8 dimensions.

[0115] S33: If the model contains more hidden layers, the output of the previous layer is passed to the next layer for similar processing, deepening the learning ability of the complex relationship patterns between Anti-TPO, ANA and the five immune parameters layer by layer.

[0116] In this embodiment, if the model contains more hidden layers, the output of the previous layer is passed to the next layer for similar processing, which can continuously deepen the learning ability of the complex correlation patterns between Anti-TPO, ANA, and the five immune markers. Each hidden layer can further abstract and transform the input features to uncover more complex correlation patterns.

[0117] For example, if the model has a third hidden layer, the intermediate feature representation vector output from the second hidden layer is used as the input to the third hidden layer, and weighted summation and nonlinear mapping operations are performed again. By continuously adding hidden layers, the model can learn higher-level feature representations and improve its ability to learn complex association patterns.

[0118] In one embodiment, a multi-layered network structure of a deep learning framework can be used, where the output of the previous layer is passed to the next layer for similar processing. By setting appropriate weight matrices and activation functions, the learning ability to understand the complex correlation patterns between Anti-TPO, ANA, and the five immune markers can be deepened layer by layer. This can improve the model's performance and prediction accuracy.

[0119] S34: Based on the output of the penultimate hidden layer, an attention mechanism or feature weighting module is introduced to dynamically adjust the importance weights of features from different sources, thereby strengthening the attention to key pathological correlation signals.

[0120] In this embodiment, an attention mechanism or feature weighting module is introduced based on the output of the penultimate hidden layer to dynamically adjust the importance weights of features from different sources. Different features may contribute differently to the model's prediction results; by introducing an attention mechanism or feature weighting module, the focus on key pathological correlation signals can be strengthened.

[0121] For example, among the features output from the penultimate hidden layer, some features are closely related to the risk of CSU patients with CTD, while others are less correlated. An attention mechanism or feature weighting module can be used to increase the weight of features closely related to risk and decrease the weight of features with weaker correlation.

[0122] S35: The weighted features are fed into the last hidden layer. Cross-level feature integration is achieved through global pooling or splicing operations, and finally the fused feature vector representing the risk of CSU patients with connective tissue disease CTD is output.

[0123] In this embodiment, the weighted features are fed into the last hidden layer, and cross-level feature integration is achieved through global pooling or splicing operations, which can fuse feature information from different levels. Cross-level feature integration can comprehensively utilize feature information from each level to form a more comprehensive and accurate fused feature vector that characterizes the risk of CSU patients with connective tissue disease CTD.

[0124] For example, the weighted features are fed into the last hidden layer, and global pooling is used to aggregate features from different dimensions, or concatenation is used to join features from different levels together. The final output fused feature vector contains all the important information from the input layer to the last hidden layer.

[0125] In one embodiment, global pooling layers or concatenation layers in a deep learning framework can be used to achieve cross-level feature integration. Weighted features are taken as input, and a fused feature vector is obtained through global pooling or concatenation operations. This improves the model's ability to represent the risk of CSU patients with concurrent CTD.

[0126] In one embodiment, reference Figure 5 Step S40 may include steps S41-S45, which will be described in detail below: S41: Input the fused feature vector output from the last hidden layer into the linear transformation unit of the output layer, and map it to a single real value through the learnable parameter matrix.

[0127] In this embodiment, the fused feature vector output from the last hidden layer contains high-order interaction information between Anti-TPO, ANA, and the five immune indicators. Inputting this feature vector into the linear transformation unit of the output layer and mapping it to a single real value using a learnable parameter matrix is ​​to convert the fused feature vector into a numerical value that can be used to calculate the risk probability.

[0128] For example, suppose the fused feature vector output by the last hidden layer is [d1,d2,d3], and the learnable parameter matrix of the output layer is [e1,e2,e3]. Through linear transformation, a single real value f = e1*d1 + e2*d2 + e3*d3 is obtained.

[0129] S42: Apply the sigmoid function to compress and transform this real value so that it falls within the range of 0 to 1, and use it as the risk probability value of CSU patients with connective tissue disease CTD.

[0130] In this embodiment, the sigmoid function is applied to the real values ​​for compression transformation in order to convert them into probability values. The output of the sigmoid function ranges from 0 to 1, which conforms to the definition of probability. Through this transformation, the real values ​​obtained by linear transformation can be converted into a preliminary risk probability estimate for CSU patients with connective tissue disease CTD.

[0131] For example, assuming the real value obtained by linear transformation is 2.5, applying the sigmoid function for compression transformation yields the value, which can be used as a preliminary risk probability estimate for CSU patients with connective tissue disease CTD.

[0132] In one embodiment, step S42 can be implemented as follows: K1: Receives a single real value generated by the linear transformation unit of the output layer, which is the unnormalized raw prediction score.

[0133] In this embodiment, a single real value generated by the linear transformation unit of the output layer is received. This real value is the unnormalized raw prediction score. This score reflects the model's preliminary judgment on the risk of CSU patients with connective tissue disease CTD, but its range may be large and it does not have probabilistic significance.

[0134] For example, the single real value generated by the linear transformation unit of the output layer is 3.2, which is the unnormalized raw prediction score.

[0135] K2: Substitute the original predicted score into the predefined sigmoid function expression for element-wise function mapping, and output a continuous value between zero and one.

[0136] In this embodiment of the application, the original predicted score is substituted into the predefined sigmoid function expression for element-wise function mapping in order to convert the original predicted score into a continuous value between zero and one, so that it has a probabilistic meaning.

[0137] The expression for the sigmoid function is: ; where x is the original predicted score.

[0138] For example, substituting the original predicted score of 3.2 into the sigmoid function expression, we get ≈0.96, which is a continuous value between zero and one.

[0139] K3: Perform floating-point precision normalization on the continuous numerical value, converting it into a standard probability representation that conforms to probabilistic semantics.

[0140] In this embodiment, floating-point precision normalization is performed on continuous numerical values ​​to ensure that they conform to the standard probabilistic representation form of probabilistic semantics. Because computers may have precision issues when processing floating-point numbers, it is necessary to normalize continuous numerical values ​​to more accurately represent probabilities.

[0141] For example, the continuous value obtained by the sigmoid function is 0.956789. After performing floating-point precision normalization to retain a certain number of decimal places, such as four decimal places, we get 0.9568.

[0142] K4: Verify the consistency between the standard probability representation and the baseline risk distribution recorded during the training phase to determine if there is a systematic bias.

[0143] In this embodiment, the consistency check between the standard probability representation and the baseline risk distribution recorded during the training phase is performed to determine whether there is a systematic shift in the probability output by the model. The baseline risk distribution is statistically obtained during the training phase and reflects the risk probability distribution under normal circumstances. If the probability output by the model differs significantly from the baseline risk distribution, it may indicate a problem with the model.

[0144] For example, the baseline risk distribution recorded during the training phase showed that the risk probability of CSU patients with CTD was commonly between 0.1 and 0.3. However, the standard probability representation of the model output was 0.8, which differed significantly from the baseline risk distribution and may indicate a systematic bias.

[0145] K5: If a probability distribution shift is detected, the probability calibration module is triggered, and a monotonic order-preserving adjustment strategy is used to fine-tune the output probability to ensure that it remains logically consistent with the true positive rate.

[0146] In this embodiment, if a probability distribution shift is detected, the probability calibration module is triggered. A monotonic order-preserving adjustment strategy is used to fine-tune the output probability to ensure that the model's output probability remains logically consistent with the true positive rate. The monotonic order-preserving adjustment strategy guarantees that the adjusted probabilities maintain a monotonically increasing relationship and do not change their relative magnitude.

[0147] For example, if a shift in the probability distribution output by the model is detected, the probability calibration module is triggered. A monotonic order-preserving adjustment strategy is used to fine-tune the output probability to make it closer to the true positive rate. For instance, if the model output probability is 0.8, after adjustment it becomes 0.3, which better reflects the true positive rate.

[0148] In this embodiment, the parameters of the output layer have been pre-trained and fine-tuned based on the label distribution in historical training samples, and remain fixed during the model inference stage to correct for probability bias caused by class imbalance. Specifically: By combining the label distribution in historical training samples, the bias term of the output layer is fine-tuned to correct the probability bias caused by class imbalance.

[0149] In this embodiment, the bias term of the output layer is fine-tuned based on the label distribution in historical training samples to correct the probability bias caused by class imbalance. In real-world data, the number of CSU patients with combined CTDs may differ significantly from the number of samples without combined CTDs, leading to class imbalance. This imbalance affects the model's training and prediction results; fine-tuning the bias term of the output layer can alleviate this problem.

[0150] For example, in the historical training samples, the number of CSU patients with combined CTDs was relatively small, while the number of samples without combined CTDs was relatively large. Analysis of the label distribution revealed that this class imbalance caused the model to predict a lower probability of combined CTDs. Fine-tuning the bias term of the output layer increased the probability weight of the combined CTD category to correct the probability bias.

[0151] During the model inference phase, the pre-trained output layer parameters are fixed, and only new fused features are received as input to generate the corresponding risk probability output in real time.

[0152] In this embodiment, during the model inference phase, the pre-trained output layer parameters are fixed, and only new fusion features are received as input to generate the corresponding risk probability output in real time. This is to ensure the stability and efficiency of the model. Since the output layer parameters are determined during the training phase, there is no need to adjust these parameters during the inference phase. Simply inputting the new fusion features into the model allows for the rapid generation of the corresponding risk probability output.

[0153] For example, in practical applications, when the fusion features of a new CSU patient are input into the model, the parameters of the output layer are fixed, the fusion features are input into the model, and the model will generate the risk probability output of the patient having CTD in real time.

[0154] S43: Temporarily store the generated risk probability value in a temporary cache.

[0155] For example, if a risk probability value of 0.7 is generated, it is temporarily stored in a temporary cache. This risk probability value can then be used to classify risk levels and generate clinical screening recommendations.

[0156] Accordingly, to better implement the above methods, this application also provides a predictive system for connective tissue disease association in chronic spontaneous urticaria. For example... Figure 6 As shown, the chronic spontaneous urticaria connective tissue disease association prediction system 60 includes: The acquisition module 61 is used to acquire multi-dimensional laboratory test index data of patients with chronic spontaneous urticaria. The laboratory test index data includes thyroid peroxidase antibody Anti-TPO, antinuclear antibody ANA, and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4. The feature vector generation module 62 is used to fill in missing values ​​and standardize the laboratory test index data, and generate derived variables by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables; wherein the deep neural network model includes an input layer, at least two fully connected hidden layers and an output layer connected in sequence, wherein the output of the previous hidden layer is used as the input of the next hidden layer, which is used to pass through and abstract feature representation layer by layer; The fusion feature processing module 63 is used to input the multidimensional feature vector into the deep neural network model, extract the high-order interaction relationship between Anti-TPO, ANA and the five immune indicators through nonlinear transformation between each hidden layer, and complete the multi-level feature fusion in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD. The probability prediction module 64 is used to calculate and output the risk probability value of CSU patients with connective tissue disease CTD based on the fusion features and the output layer of the deep neural network model. The results processing module 65 is used to classify the risk probability value according to a preset risk threshold, generate low-risk, medium-risk or high-risk prediction results, and output differentiated clinical screening suggestions accordingly.

[0157] In one embodiment, the acquisition module 61 is specifically used for: The hospital information system screens electronic medical records of patients who meet the diagnostic criteria for chronic spontaneous urticaria and extracts the corresponding laboratory test request information. Based on the request information, the corresponding laboratory test results are retrieved. The extracted laboratory test results data were checked for consistency of source to confirm whether Anti-TPO, ANA and the five immune indicators all came from the serum sample test reports at the same sampling time point. If there was a time deviation, the sample record was removed. Laboratory test results that have passed the consistency verification are structured and categorized according to indicator categories. Data subsets for the Anti-TPO field, ANA field, and five immune parameters field are established, and the data unit and test method type corresponding to each field are marked. Preliminary outlier identification is performed on the categorized indicator data. Based on the upper and lower limits of the medical reference range, it is determined whether there are obvious outlier values. Suspected erroneous data is manually reviewed or automatically marked as pending processing. The data of Anti-TPO, ANA and five immune indicators that have completed the initial outlier identification are integrated into a raw data table with a unified format, ensuring that each record contains all ten indicator fields and has no missing markers.

[0158] In one embodiment, reference Figure 7 The feature vector generation module 62 includes: The imputation processing submodule 621 is used to impute missing data items in the original indicators using an imputation strategy based on the mean of the same patient group. This includes replacing missing values ​​with the average value of the corresponding indicators of other patients of the same gender, age group and underlying disease status. The standard processing submodule 622 is used to perform standardization processing on all the original indicators after filling, so that each indicator is converted into a standard normal distribution with a mean of zero and a variance of one, in order to eliminate the influence of dimensional differences on model training. The variable construction submodule 623 is used to construct several clinically significant ratio-type derived variables based on the standardized original indicators, including the combination ratios of IgG / IgA, C3 / C4, and Anti-TPO / ANA, and simultaneously generate multiple cross-product terms to reflect the potential synergistic effect between indicators. The vector processing submodule 624 is used to merge all the original indicators and newly generated derived variables to form an initial feature set, and to perform a redundancy evaluation on the set, remove redundant features that are highly linearly correlated with other variables to reduce dimensional complexity, and determine the feature combination retained after removal as the final multidimensional feature vector.

[0159] In one embodiment, reference Figure 8 The fusion feature processing module 63 includes: The activation processing submodule 631 is used to load the preprocessed multidimensional feature vector into the input layer of the deep neural network, and then process it through the weight matrix operation and nonlinear activation function of the first hidden layer in sequence to output the primary abstract feature representation. The mapping operation submodule 632 is used to take the primary abstract features as the input of the second hidden layer, and perform weighted summation and nonlinear mapping operations again to obtain intermediate feature expressions that can reflect the second-order interaction between some indicators. The feature processing submodule 633 is used to pass the output of the previous layer to the next layer for similar processing if the model contains more hidden layers, thereby deepening the learning ability of the complex correlation patterns between Anti-TPO, ANA and the five immune markers layer by layer. Based on the output of the penultimate hidden layer, an attention mechanism or feature weighting module is introduced to dynamically adjust the importance weights of features from different sources, and strengthen the attention to key pathological correlation signals. The weighted features are sent to the last hidden layer, and cross-level feature integration is achieved through global pooling or splicing operations, and finally outputting a fused feature vector that represents the risk of CSU patients with connective tissue disease CTD.

[0160] In one embodiment, reference Figure 9 The probability prediction module 64 may include: The transformation processing submodule 641 is used to input the fused feature vector output from the last hidden layer to the linear transformation unit of the output layer and map it to a single real value through a learnable parameter matrix. The compression processing submodule 642 is used to apply the sigmoid function to compress the real value so that it falls within the range of 0 to 1, which is used as the risk probability value of CSU patients with connective tissue disease CTD, and as the risk probability estimate of CSU patients with connective tissue disease CTD; wherein, the parameters of the output layer have been pre-trained and fine-tuned in combination with the label distribution in historical training samples, and remain fixed during the model inference stage to correct the probability bias problem caused by class imbalance; Storage submodule 643 is used to temporarily store the generated risk probability values ​​in a temporary cache area.

[0161] refer to Figure 11 In this embodiment of the application, the structure of the deep neural network model may include: 1. Input layer The input layer receives a preprocessed multidimensional feature vector. This vector contains multidimensional laboratory test data from patients with chronic spontaneous urticaria (CSU), such as raw indicators like thyroid peroxidase antibody (Anti-TPO), antinuclear antibody (ANA), immunoglobulin IgG, IgA, IgM, complement C3, and complement C4, as well as derived variables generated by constructing ratios and cross-product terms between the indicators. The number of neurons in the input layer is consistent with the dimension of the multidimensional feature vector, with each neuron corresponding to one feature, responsible for feeding the data into the model for subsequent processing.

[0162] 2. Hidden layer There are at least two hidden layers, which are fully connected. The output of the previous hidden layer serves as the input of the next hidden layer, and is used to pass through layers and abstract feature representations.

[0163] First hidden layer: The multidimensional feature vector of the input layer is multiplied by the learnable weight matrix of the first hidden layer to obtain a weighted sum. Then, a non-linear activation function (such as the ReLU function, expressed as f(x)=max(0,x)) is applied to the result for element-wise transformation, and a primary abstract feature representation is output.

[0164] Subsequent hidden layers: Subsequent hidden layers continue to process and abstract the features output by the previous layer. Each layer uses weighted summation and nonlinear mapping operations to continuously uncover higher-order interactions between Anti-TPO, ANA, and the five immune indicators. For example, the second hidden layer takes the primary abstract features output by the first hidden layer as input, performs weighted summation and nonlinear mapping again, and obtains intermediate feature representations that can reflect the second-order interactions between some indicators. As the number of hidden layers increases, the model can learn more complex feature patterns and interaction relationships.

[0165] The penultimate hidden layer: Based on the output of the penultimate hidden layer, an attention mechanism or feature weighting module is introduced. The attention mechanism first calculates the attention score for each feature. Assuming the feature vector output by the penultimate hidden layer is h=[h1,h2,…,hn], the attention score ei = transpose of w multiplied by hi is obtained by performing a dot product operation between the learnable weight vector w and the feature vector. Then, the attention scores are normalized using the softmax function, resulting in the attention weight αi = exp(ei) divided by the sum of all exp(ej) (j from 1 to n). Finally, the feature vector is weighted and summed according to the attention weights to obtain the feature vector h' = sum of all αi multiplied by hi (i from 1 to n). The feature weighting module then evaluates the importance of each feature, assigns corresponding weights, and processes the feature vector h using a learnable weight matrix W to obtain the weighted feature vector h'' = W multiplied by h.

[0166] 3. Output layer The output layer receives the fused feature vector from the last hidden layer. First, the fused feature vector is input into the linear transformation unit of the output layer, and mapped to a single real value using a learnable parameter matrix. Then, the sigmoid function (expressed as y = 1 divided by (1 + e to the power of -x)) is applied to this real value for compression, ensuring it falls within the 0-1 interval. This serves as a preliminary risk probability estimate for CSU patients with connective tissue disease (CTD).

[0167] The model training process is as follows: The hospital information system screens electronic medical records of patients who meet the diagnostic criteria for chronic spontaneous urticaria, extracts the corresponding laboratory test request information, and retrieves the corresponding laboratory test results data to form the original dataset.

[0168] The original dataset underwent missing value imputation and standardization. Ratios and cross-product terms between indicators were constructed to generate derived variables, forming a multidimensional feature vector containing both original indicators and derived variables. Specifically, missing data items were imputed using a mean-based imputation strategy for similar patient groups. Z-score standardization was performed on all imputed original indicators. Clinically meaningful ratio-type derived variables and cross-product terms were constructed. Redundancy was assessed on the generated feature set, and redundant features were removed.

[0169] The preprocessed dataset is divided into training, validation, and test sets. The training set accounts for approximately 70%-80% of the dataset and is used for model training; the validation set accounts for approximately 10%-15% of the dataset and is used to evaluate model performance and tune hyperparameters during training; the test set accounts for approximately 10%-15% of the dataset and is used to evaluate the model's generalization ability.

[0170] Initialize the weight matrix and bias terms of the deep neural network model. The weight matrix can be initialized using random initialization methods, such as Xavier or He initialization, to ensure the model effectively learns data features in the early stages of training. The bias terms are typically initialized as a zero vector. Simultaneously, determine the learning rate α, which controls the step size for updating the model parameters. If the learning rate is too large, the model may skip optimal solutions, leading to non-convergence; if the learning rate is too small, the model's training speed will slow down.

[0171] Define the loss function: Use the binary cross-entropy loss function as the model's loss function. For binary classification problems, the binary cross-entropy loss function can effectively measure the difference between the model's prediction and the true label.

[0172] Its expression is: .

[0173] Where N is the sample size. Let be the true label (0 or 1) of the i-th sample. Let be the predicted probability of the model for the i-th sample.

[0174] During training, data from the training set is input into the model in batches. Forward propagation is used to calculate predictions, and then the loss value is calculated based on the loss function. Next, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the model parameters are updated according to the optimization algorithm. This process is repeated until the model's performance reaches a satisfactory level or the preset number of training epochs is reached.

[0175] The implementation details of each module are provided in the preceding method embodiments and will not be repeated here. The technical effects achieved by each module and device are described in the foregoing method embodiments.

[0176] like Figure 10 As shown, this application embodiment also provides a computer device 90 (which may be an intelligent body fat scale), which includes a processor 901 and a memory 902, wherein the memory 902 stores a computer program, and when the computer program is executed by the processor 901, the processor 901 performs the steps of any of the methods described above.

[0177] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0178] The embodiments, implementation methods, and related technical features of this application can be combined and substituted for each other without conflict.

[0179] The above are merely preferred embodiments of this application and do not constitute any limitation on this application. Any simple modifications, equivalent changes, or alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application shall still fall within the scope of the technical solution of this application.

Claims

1. A machine learning-based method for predicting the association between chronic spontaneous urticaria and connective tissue disease, characterized in that, Includes the following steps: Obtain multidimensional laboratory test data of patients with chronic spontaneous urticaria. The laboratory test data includes thyroid peroxidase antibody Anti-TPO, antinuclear antibody ANA, and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4. The laboratory test index data are imputed and standardized, and derived variables are generated by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables. The deep neural network model includes an input layer, at least two fully connected hidden layers, and an output layer connected in sequence. The output of the previous hidden layer serves as the input of the next hidden layer, which is used to pass through layers and abstract feature representations. The multidimensional feature vector is input into the deep neural network model. The high-order interaction relationship between Anti-TPO, ANA and five immune indicators is extracted step by step through nonlinear transformation between each hidden layer. Multi-level feature fusion is completed in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD. Based on the fusion features, the output layer of the deep neural network model calculates and outputs the risk probability value of CSU patients with connective tissue disease CTD; The risk probability values ​​are classified into levels based on preset risk thresholds, generating low-risk, medium-risk, or high-risk prediction results, and outputting corresponding differentiated clinical screening recommendations.

2. The method according to claim 1, characterized in that, The acquisition of multidimensional laboratory test data for patients with chronic spontaneous urticaria (CSU) includes thyroid peroxidase antibody (Anti-TPO), antinuclear antibody (ANA), and five immune indicators, including: The hospital information system screens electronic medical records of patients who meet the diagnostic criteria for chronic spontaneous urticaria and extracts the corresponding laboratory test request information. Based on the request information, the corresponding laboratory test results are retrieved. The extracted laboratory test results data were checked for consistency of source to confirm whether Anti-TPO, ANA and the five immune indicators all came from the serum sample test reports at the same sampling time point. If there was a time deviation, the sample record was removed. Laboratory test results that have passed the consistency verification are structured and categorized according to indicator categories. Data subsets for the Anti-TPO field, ANA field, and five immune parameters field are established, and the data unit and test method type corresponding to each field are marked. Preliminary outlier identification is performed on the categorized indicator data. Based on the upper and lower limits of the medical reference range, it is determined whether there are obvious outlier values. Suspected erroneous data is manually reviewed or automatically marked as pending processing. The data of Anti-TPO, ANA and five immune indicators that have completed the initial outlier identification are integrated into a raw data table with a unified format, ensuring that each record contains all ten indicator fields and has no missing markers.

3. The method according to claim 2, characterized in that, The process involves imputing missing values ​​and standardizing the laboratory test indicator data, and generating derived variables by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables, including: For missing data items in the original indicators, an imputation strategy based on the mean of the same patient group is used to fill them in. This includes replacing missing values ​​with the average of the corresponding indicators of other patients of the same gender, age group and underlying disease status. Standardization is performed on all the original indicators after filling, so that each indicator is converted into a standard normal distribution with a mean of zero and a variance of one, in order to eliminate the influence of dimensional differences on model training. Based on the standardized original indicators, several clinically significant ratio-based derived variables were constructed, including the combined ratios of IgG / IgA, C3 / C4, and Anti-TPO / ANA, and multiple cross-product terms were generated simultaneously to reflect the potential synergistic effect between the indicators. All original indicators and newly generated derived variables are combined to form an initial feature set. Redundancy assessment is performed on this set, and redundant features that are highly linearly correlated with other variables are removed to reduce dimensionality complexity. The combination of features retained after removal is determined as the final multidimensional feature vector.

4. The method according to claim 3, characterized in that, The process involves inputting the multidimensional feature vector into the deep neural network model, extracting high-order interaction relationships between Anti-TPO, ANA, and five immune indicators through nonlinear transformations between hidden layers, and performing multi-level feature fusion in the last hidden layer to generate fused features characterizing the risk of CSU patients with connective tissue disease (CTD), including: The preprocessed multidimensional feature vector is loaded into the input layer of the deep neural network, and then processed by the weight matrix operation and nonlinear activation function of the first hidden layer to output a primary abstract feature representation. The primary abstract features are used as input to the second hidden layer, and weighted summation and nonlinear mapping operations are performed again to obtain intermediate feature representations that can reflect the second-order interaction between some indicators. If the model contains more hidden layers, the output of the previous layer is passed to the next layer for similar processing, deepening the learning ability of the complex relationship patterns between Anti-TPO, ANA and the five immune markers layer by layer. Based on the output of the penultimate hidden layer, an attention mechanism or feature weighting module is introduced to dynamically adjust the importance weights of features from different sources, thereby strengthening the attention to key pathological correlation signals. The weighted features are fed into the last hidden layer, and cross-level feature integration is achieved through global pooling or splicing operations. Finally, a fused feature vector representing the risk of CSU patients with connective tissue disease CTD is output.

5. The method according to claim 4, characterized in that, The calculation and output of the risk probability value of CSU patients with connective tissue disease CTD based on the fusion features by the output layer of the deep neural network model includes: The fused feature vector output from the last hidden layer is input into the linear transformation unit of the output layer, and mapped to a single real value through a learnable parameter matrix. The real value is compressed using the sigmoid function to fall within the range of 0 to 1, and is used as the risk probability value of CSU patients with connective tissue disease CTD, as an estimate of the risk probability of CSU patients with connective tissue disease CTD; wherein, the parameters of the output layer have been pre-trained and fine-tuned in combination with the label distribution in historical training samples, and remain fixed during the model inference stage to correct the probability bias problem caused by class imbalance. The generated risk probability values ​​are temporarily stored in a temporary cache.

6. A machine learning-based predictive system for chronic spontaneous urticaria as a connective tissue disease, characterized in that, The system includes: The acquisition module is used to acquire multi-dimensional laboratory test indicator data of patients with chronic spontaneous urticaria. The laboratory test indicator data includes thyroid peroxidase antibody Anti-TPO, antinuclear antibody ANA, and five immune indicators, including immunoglobulin IgG, IgA, IgM, complement C3, and complement C4. The feature vector generation module is used to fill in missing values ​​and standardize the laboratory test index data, and generate derived variables by constructing ratios and cross-product terms between indicators, forming a multidimensional feature vector containing the original indicators and derived variables; wherein the deep neural network model includes an input layer, at least two fully connected hidden layers and an output layer connected in sequence, wherein the output of the previous hidden layer serves as the input of the next hidden layer, used to pass through and abstract feature representation layer by layer; The fusion feature processing module is used to input the multidimensional feature vector into the deep neural network model, extract the high-order interaction relationship between Anti-TPO, ANA and the five immune indicators through nonlinear transformation between each hidden layer, and complete the multi-level feature fusion in the last hidden layer to generate fusion features that characterize the risk of CSU patients with connective tissue disease CTD. The probability prediction module is used to calculate and output the risk probability value of CSU patients with connective tissue disease CTD based on the fusion features and the output layer of the deep neural network model. The results processing module is used to classify the risk probability value according to a preset risk threshold, generate low-risk, medium-risk or high-risk prediction results, and output differentiated clinical screening suggestions accordingly.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: The hospital information system screens electronic medical records of patients who meet the diagnostic criteria for chronic spontaneous urticaria and extracts the corresponding laboratory test request information. Based on the request information, the corresponding laboratory test results are retrieved. The extracted laboratory test results data were checked for consistency of source to confirm whether Anti-TPO, ANA and the five immune indicators all came from the serum sample test reports at the same sampling time point. If there was a time deviation, the sample record was removed. Laboratory test results that have passed the consistency verification are structured and categorized according to indicator categories. Data subsets for the Anti-TPO field, ANA field, and five immune parameters field are established, and the data unit and test method type corresponding to each field are marked. Preliminary outlier identification is performed on the categorized indicator data. Based on the upper and lower limits of the medical reference range, it is determined whether there are obvious outlier values. Suspected erroneous data is manually reviewed or automatically marked as pending processing. The data of Anti-TPO, ANA and five immune indicators that have completed the initial outlier identification are integrated into a raw data table with a unified format, ensuring that each record contains all ten indicator fields and has no missing markers.

8. The system according to claim 6, characterized in that, The feature vector generation module includes: The imputation processing submodule is used to impute missing data items in the original indicators using an imputation strategy based on the mean of the same patient group. This includes replacing missing values ​​with the average value of the corresponding indicators of other patients of the same gender, age group and underlying disease status. The standard processing submodule is used to perform standardization processing on all the original indicators after filling, so that each indicator is converted into a standard normal distribution with a mean of zero and a variance of one, in order to eliminate the impact of dimensional differences on model training. The variable construction sub-model is used to construct several clinically meaningful ratio-type derived variables based on the standardized original indicators, including the combination ratios of IgG / IgA, C3 / C4, and Anti-TPO / ANA, and simultaneously generate multiple cross-product terms to reflect the potential synergistic effect between indicators. The vector processing submodule is used to merge all the original indicators and newly generated derived variables to form an initial feature set, and to perform a redundancy evaluation on the set, removing redundant features that are highly linearly correlated with other variables to reduce dimensionality complexity, and determining the combination of features retained after removal as the final multidimensional feature vector.

9. The system according to claim 8, characterized in that, The fusion feature processing module includes: The activation processing submodule is used to load the preprocessed multidimensional feature vector into the input layer of the deep neural network, and then process it through the weight matrix operation and nonlinear activation function of the first hidden layer to output the primary abstract feature representation. The mapping operation submodule is used to take the primary abstract features as input to the second hidden layer, and then perform weighted summation and nonlinear mapping operations to obtain intermediate feature representations that can reflect the second-order interaction between some indicators. The feature processing submodule is used to pass the output of the previous layer to the next layer for similar processing if the model contains more hidden layers, thereby deepening the learning ability of the complex correlation patterns between Anti-TPO, ANA and the five immune markers layer by layer. Based on the output of the penultimate hidden layer, an attention mechanism or feature weighting module is introduced to dynamically adjust the importance weights of features from different sources, strengthening the attention to key pathological correlation signals. The weighted features are then sent to the last hidden layer, and cross-level feature integration is achieved through global pooling or splicing operations, finally outputting a fused feature vector that represents the risk of CSU patients with connective tissue disease CTD.

10. The system according to claim 9, characterized in that, The probability prediction module includes: The transformation processing submodule is used to input the fused feature vector output from the last hidden layer into the linear transformation unit of the output layer, and map it into a single real value through a learnable parameter matrix. The compression processing submodule is used to apply the sigmoid function to compress the real value so that it falls within the range of 0 to 1, which is used as the risk probability value of CSU patients with connective tissue disease CTD, and as the risk probability estimate of CSU patients with connective tissue disease CTD; wherein, the parameters of the output layer have been pre-trained and fine-tuned in combination with the label distribution in historical training samples, and remain fixed during the model inference stage to correct the probability bias problem caused by class imbalance; The storage submodule is used to temporarily store the generated risk probability values ​​in a temporary cache.

Images