Auxiliary diagnostic models for gestational diabetes

Abstract

This invention discloses a method and system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, an auxiliary diagnostic system for gestational diabetes mellitus, an electronic device, and a computer-readable storage medium. The method includes the following steps: selecting a machine learning model that can be used for the auxiliary diagnostic model; acquiring protein biomarker data from multiple samples, wherein the multiple samples are collected from the tears of a subject, and using the protein biomarker data as the raw dataset for the machine learning model; training and evaluating the machine learning model based on the raw dataset, and obtaining the auxiliary diagnostic model based on the evaluation results.

Description

Technical Field

[0001] This invention relates to an auxiliary diagnostic model for gestational diabetes mellitus, specifically to a method and system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, an auxiliary diagnostic system for gestational diabetes mellitus, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Gestational diabetes mellitus (GDM) refers to abnormal glucose metabolism that is first discovered or occurs during pregnancy. It is a type of diabetes that can lead to complications such as fetal malformations, fetal distress, stillbirth, neonatal hypoglycemia, macrosomia, and dystocia or stillbirth. Infants born to mothers with GDM have an increased risk of exposure to chemical imbalances, making early and accurate detection crucial. Clinically, the oral glucose tolerance test (OGTT) at 24-28 weeks of gestation is commonly used to diagnose GDM. However, this mainstream method is invasive, time-consuming, and labor-intensive, requiring multiple blood glucose level checks before and after glucose administration. Furthermore, the sensitivity and specificity of the OGTT are low for late-onset GDM in early pregnancy. Therefore, this method is not suitable for early detection, and there is still an urgent clinical need for more non-invasive and accurate biomarkers.

[0003] Tears are rich in proteins and lipids, which have become important sources of diagnostic biomarkers. Pathophysiological changes in diseases may be reflected in tears, and the diagnostic capability of tear biomarkers has been confirmed in many non-ophthalmic systemic diseases, such as multiple sclerosis and Parkinson's disease. The most common tear sample collection methods include the Schirmer strip method and the capillary method, which allow for non-invasive sample collection. Therefore, tear biomarkers based on comparative proteomics analysis can be used for the diagnosis of GDM.

[0004] Machine learning, a crucial branch of artificial intelligence in modern information science, uses algorithms to analyze massive amounts of data and learn from and summarize problems, thereby making inferences or judgments by simulating the learning methods of the human brain. There are many commonly used machine learning algorithms, which can be categorized by learning style (supervised, unsupervised, semi-supervised, etc.) or by similarity in form or function (classification, regression, decision trees, clustering, deep learning, etc.). While the combination of machine learning and clinical diagnosis has made some progress in recent years, limitations exist due to the characteristics of medical sample data and the applicability of various algorithms, especially in the diagnosis of gestational diabetes mellitus, where no highly accurate AI-assisted diagnostic model has yet been found. Summary of the Invention

[0005] Based on the above situation, the main objective of this invention is to provide a method and system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, an auxiliary diagnostic model for gestational diabetes mellitus, an electronic device, and a computer-readable storage medium. The method preprocesses protein biomarker data from samples and uses it for a machine learning model. Based on the optimized training model, data conversion processing and performance evaluation are combined to finally confirm the optimal auxiliary diagnostic model for gestational diabetes mellitus. The resulting auxiliary diagnostic model can accurately and efficiently diagnose gestational diabetes mellitus based on tear protein biomarkers.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The first aspect of the present invention provides a method for establishing an auxiliary diagnostic model for gestational diabetes mellitus, the method comprising the following steps:

[0008] Select a machine learning model that can be used in the auxiliary diagnostic model;

[0009] Protein biomarker data from multiple samples, wherein the multiple samples are collected from the tears of subjects, and the protein biomarker data is used as the raw dataset for the machine learning model;

[0010] The machine learning model is trained and evaluated based on the original dataset, and the auxiliary diagnostic model is obtained based on the evaluation results.

[0011] Preferably, training and evaluating the machine learning model based on the original dataset includes:

[0012] The original dataset is scaled to obtain preprocessed data;

[0013] The preprocessed data is randomly divided into a training set and a test set. The machine learning model is trained using the training set to obtain a trained model, and the trained model is evaluated using the test set. The optimal trained model is determined based on the evaluation results.

[0014] Based on the scaling process, a data transformation program is generated that can be used for a single sample. The data transformation program is integrated into the optimal training model to obtain an ensemble model. The performance of the ensemble model is evaluated, and the auxiliary diagnostic model is obtained based on the performance evaluation results.

[0015] Preferably,

[0016] The scaling process includes uniform scaling of the original dataset using a Z-value algorithm;

[0017] The data conversion program processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm, so that it is suitable for a single sample as the input data.

[0018] Preferably, the performance evaluation of the integrated model includes:

[0019] Plot the ROC curve and calculate the AUC value of the ensemble model.

[0020] Preferably, the protein biomarker data includes tear protein concentration data for K2C5, PIP, and GLOD4.

[0021] Preferably, training the machine learning model using the training set to obtain the trained model includes:

[0022] The tear protein concentration data of K2C5, PIP and GLOD4 were used as the optimal feature subset;

[0023] Cross-validation is used to obtain model evaluation results. Based on the model evaluation results, the optimal hyperparameters are determined, and the training model is obtained according to the optimal hyperparameters.

[0024] Preferably, the ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

[0025] Preferably, the subjects include pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

[0026] Preferably, the machine learning model includes deep neural networks, linear support vector machines, radial basis function support vector machines, stochastic gradient descent, and random forests.

[0027] A second aspect of the present invention provides a system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, the system comprising:

[0028] The model unit is used to select a machine learning model that can be used in the auxiliary diagnostic model;

[0029] An input unit is used to acquire protein biomarker data from multiple samples, wherein the multiple samples are collected from the tears of subjects, and the protein biomarker data is used as the raw dataset of the machine learning model.

[0030] The modeling unit is used to train and evaluate the machine learning model based on the original dataset, and obtain the auxiliary diagnostic model based on the evaluation results.

[0031] Preferably, the modeling unit includes:

[0032] The data processing module is used to scale the original dataset and obtain preprocessed data;

[0033] The model training module is used to randomly divide the preprocessed data into a training set and a test set, train the machine learning model using the training set to obtain a trained model, evaluate the trained model using the test set, and determine the optimal training model based on the evaluation results.

[0034] The model building module is used to generate a data transformation program that can be used for a single sample based on the scaling process, integrate the data transformation program into the optimal training model to obtain an ensemble model, evaluate the performance of the ensemble model, and obtain the auxiliary diagnostic model based on the performance evaluation results.

[0035] Preferably,

[0036] The data processing module includes a first sub-module, which can perform uniform-scale processing on the original dataset using the Z-value algorithm.

[0037] The data conversion program includes a second submodule, which processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm to adapt it to a single sample as the input data.

[0038] Preferably, the model building module further includes a performance evaluation module for plotting ROC curves and calculating the AUC value of the ensemble model.

[0039] Preferably, the protein biomarker data includes tear protein concentration data for K2C5, PIP, and GLOD4.

[0040] Preferably, the optimal feature subset of the model training module is the tear protein concentration data of K2C5, PIP, and GLOD4;

[0041] The model training module includes a third sub-module, which can use cross-validation to obtain model evaluation results, determine the optimal hyperparameters based on the model evaluation results, and obtain the training model based on the optimal hyperparameters.

[0042] Preferably, the ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

[0043] Preferably, the subjects include pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

[0044] Preferably, the machine learning model includes deep neural networks, linear support vector machines, radial basis function support vector machines, stochastic gradient descent, and random forests.

[0045] A third aspect of the present invention provides an auxiliary diagnostic system for gestational diabetes mellitus, the auxiliary diagnostic system comprising:

[0046] The input module is used to input diagnostic data, which includes tear protein concentration data of the subject for K2C5, PIP and GLOD4;

[0047] A diagnostic module is used to judge the data to be diagnosed, and the diagnostic module includes an auxiliary diagnostic model obtained by the establishment method described in the first aspect above.

[0048] The output module is used to output the judgment result of the data to be diagnosed, so as to obtain the probability of the subject having gestational diabetes.

[0049] Preferably, the auxiliary diagnostic model is obtained based on the random forest model.

[0050] A fourth aspect of the present invention provides an electronic device, comprising: a processor; and a memory storing a computer program, which, when executed by the processor, enables the establishment method as described in the first aspect above.

[0051] A fifth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to execute the establishment method as described in the first aspect above.

[0052] Compared with the prior art, the present invention has significant advantages and beneficial effects, and has at least the following advantages:

[0053] The present invention provides a method and system for establishing an auxiliary diagnostic model for gestational diabetes mellitus. This method acquires protein biomarker data from multiple subjects' tear fluid samples and uses this data as the raw dataset for a machine learning model. The machine learning model is then trained and evaluated based on this raw dataset, resulting in an auxiliary diagnostic model for gestational diabetes mellitus. This model is suitable for diagnosing gestational diabetes mellitus in large-scale cohorts. Based on the optimally selected machine learning model and applicable protein biomarkers, it can effectively identify gestational diabetes mellitus and exhibits non-invasiveness and good performance. The optimal AUC value for the predictive accuracy of this auxiliary diagnostic model can be above 0.98.

[0054] The present invention provides an auxiliary diagnostic system for gestational diabetes mellitus. The diagnostic module of the system includes an auxiliary diagnostic model obtained according to the above-described method. It is capable of rapid and accurate auxiliary diagnosis of gestational diabetes mellitus based on protein biomarker data (including tear protein concentration data of K2C5, PIP and GLOD4) in multiple samples. It can quickly analyze and judge sample data and output results, thereby providing effective and non-invasive diagnostic basis for clinical applications.

[0055] The electronic device and computer-readable storage medium of the present invention, by employing the above-described method to establish an auxiliary diagnostic model, can provide rapid and accurate auxiliary diagnosis of gestational diabetes mellitus for large-scale cohorts based on protein biomarkers, and achieve non-invasive auxiliary diagnosis of gestational diabetes mellitus. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the method for establishing an auxiliary diagnostic model for gestational diabetes mellitus according to the present invention.

[0057] Figure 2 This is a schematic diagram of protein concentration data of a protein biomarker validated by ELISA, according to an embodiment of the present invention.

[0058] Figure 3 This is a schematic diagram of protein concentration data of a protein biomarker validated by ELISA, according to an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram of protein concentration data of a protein biomarker validated by ELISA, according to an embodiment of the present invention.

[0060] Figure 5 This is a schematic diagram of AUC values ​​for identifying pregnancy and GDM according to an embodiment of the present invention;

[0061] Figure 6 This is a schematic diagram of the modules of the system for establishing an auxiliary diagnostic model for gestational diabetes mellitus according to the present invention.

[0062] Figure 7 This is a schematic diagram of the modules of the auxiliary diagnostic system for gestational diabetes of the present invention. Detailed Implementation

[0063] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, describes the specific implementation methods, steps, features, and effects of the methods, systems, electronic devices, and computer-readable storage media proposed according to the present invention.

[0064] See Figure 1A method for establishing an auxiliary diagnostic model for gestational diabetes mellitus, the method comprising the following steps:

[0065] Select a machine learning model that can be used in the auxiliary diagnostic model;

[0066] Protein biomarker data from multiple samples, wherein the multiple samples are collected from the tears of subjects, and the protein biomarker data is used as the raw dataset for the machine learning model;

[0067] The machine learning model is trained and evaluated based on the original dataset, and the auxiliary diagnostic model is obtained based on the evaluation results.

[0068] Through the above steps, protein biomarker data that can be used to diagnose gestational diabetes are obtained based on multiple subject samples. The tear protein concentration data in the subject samples is used as the original dataset for model construction. The machine learning model is trained and evaluated based on the original dataset, thereby obtaining an auxiliary diagnostic model for gestational diabetes. This auxiliary diagnostic model is suitable for the diagnosis of gestational diabetes in large-scale cohorts. Based on the optimal machine learning model and applicable protein biomarkers, it can effectively identify gestational diabetes and has non-invasive and good performance. The optimal AUC value of the prediction accuracy of this auxiliary diagnostic model can be above 0.98.

[0069] As an optional embodiment, training and evaluating the machine learning model based on the original dataset includes:

[0070] The original dataset is scaled to obtain preprocessed data;

[0071] The preprocessed data is randomly divided into a training set and a test set. The machine learning model is trained using the training set to obtain a trained model, and the trained model is evaluated using the test set. The optimal trained model is determined based on the evaluation results.

[0072] Based on the scaling process, a data transformation program is generated that can be used for a single sample. The data transformation program is integrated into the optimal training model to obtain an ensemble model. The performance of the ensemble model is evaluated, and the auxiliary diagnostic model is obtained based on the performance evaluation results.

[0073] Through the above steps, the original dataset is scaled to obtain preprocessed data of uniform magnitude. The preprocessed data is used to train the machine learning model to obtain the optimal training model. The corresponding data transformation process of standardization is integrated into the optimal training model to obtain the ensemble model. After performance evaluation, the ensemble model is confirmed as an auxiliary diagnostic model for gestational diabetes mellitus. Thus, the auxiliary diagnostic model is suitable for the diagnosis of gestational diabetes mellitus in large-scale cohorts. The auxiliary diagnostic model obtained based on the optimal machine learning model and applicable protein biomarkers can effectively identify gestational diabetes mellitus and has non-invasiveness and good performance.

[0074] As an optional embodiment,

[0075] The scaling process includes uniform scaling of the original dataset using a Z-value algorithm;

[0076] The data conversion program processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm, so that it is suitable for a single sample as the input data.

[0077] Specifically, the original datasets obtained through the aforementioned steps may be data of different magnitudes. The Z-score algorithm is used to unify the data of different magnitudes into the same magnitude, and the calculated Z-score is used to measure them uniformly to ensure the comparability between the data. When performing data scaling based on the Z-score algorithm, the mean of the population data, the standard deviation of the population data, and individual data (i.e., protein biomarker data of a single sample validated by ELISA) are generally required.

[0078] Furthermore, the original datasets used for machine learning typically include protein biomarker data from multiple samples. This data is usually input into the machine learning model in the form of a data matrix for training. The trained model, after parameter optimization, is evaluated and determined to be a suitable auxiliary diagnostic model. However, in actual use, the input data for this auxiliary diagnostic model is usually a single, unfamiliar sample. This necessitates processing the input data through a data transformation procedure so that the auxiliary diagnostic model can make judgments on individual samples. Therefore, the data processing program processes the raw data of a single sample based on the relevant parameters used in the Z-score algorithm, as detailed in the following formula:

[0079]

[0080] In the formula, Z is the converted protein marker value of a single sample, X is the original protein marker value of a single sample, M is the mean of the protein marker data used in the Z-value algorithm, and s is the standard deviation of the protein marker data used in the Z-value algorithm.

[0081] By integrating the data transformation program into the optimal training model obtained through the above steps, the judgment requirements of individual sample data can be met, effectively simulating real-world clinical applications.

[0082] Evaluating the model multiple times on multiple validation sets can better reflect its classification or judgment performance.

[0083] As an optional embodiment, the performance evaluation of the integrated model includes:

[0084] Plot the ROC curve and calculate the AUC value of the ensemble model.

[0085] Specifically, the Receiver Operating Characteristic (ROC) curve is used to analyze binary classification models. The Area Under the Curve (AUC) is used to compare different classification models. The ROC curve of each model is plotted, and the area under the curve is used as an indicator of model performance. It is typically calculated by measuring the area within a 1x1 grid, and the AUC value must be between 0 and 1. A higher AUC value indicates a higher accuracy rate for the classifier (model).

[0086] The above steps can effectively evaluate the performance of the obtained ensemble model to confirm the best ensemble model as an auxiliary diagnostic model for gestational diabetes.

[0087] As an optional embodiment, the subjects include pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

[0088] Specifically, by detecting the dynamic changes in proteomics in the tears of subjects, protein biomarkers associated with pregnancy and GDM can be identified.

[0089] In this embodiment, the recruited subjects included: normal pregnant women in the second trimester (MS group) and pregnant women with gestational diabetes mellitus in the second trimester (GDM group). The second trimester included the second and third trimesters. There were 254 subjects in the MS group and 189 subjects in the GDM group. The inclusion criteria for all subjects included: (1) no history of pregnancy; (2) no chronic diseases; (3) no medication use within 6 months or during pregnancy; (4) no abnormalities during pregnancy (except for the GDM group); (5) no history of eye diseases; and (6) no functional impairment of vital organs found during physical examination. Subsequently, according to the inclusion criteria, 9.8% and 10% of the subjects in the MS group and GDM group, respectively, were excluded. Finally, tear samples were collected from 224 subjects in the MS group and 183 subjects in the GDM group.

[0090] As an optional embodiment, the protein biomarker data includes tear protein concentration data for K2C5, PIP, and GLOD4.

[0091] Specifically, regarding the protein marker K2C5, its full protein name: Type II cytoskeletal 5, encoding gene name: KRT5, Gene ID: 3852, Chinese name: cytokeratin 5, Uniprot accession number: P13647, Uniprot accession name: K2C5_HUMAN.

[0092] Specifically, regarding the protein marker PIP, its full protein name: Prolactin-inducible protein, encoding gene name: PIP, Gene ID: 5304, Chinese name: prolactin-induced protein, Uniprot accession number: P12273, Uniprot accession name: PIP_HUMAN.

[0093] Specifically, regarding the protein marker GLOD4, its full protein name: Glyoxalase Domain-containing4, encoding gene name: GLOD4, Gene ID: 51031, Chinese name: glyoxalase domain-containing protein 4, Uniprot accession number: Q9HC38, Uniprot accession name: GLOD4_HUMAN.

[0094] Thus, the tear protein concentration data of K2C5, PIP, and GLOD4 are used as the original dataset for model construction, and the machine learning model is trained and evaluated accordingly with the applicable protein marker data, providing a data basis for obtaining the best machine learning model.

[0095] As an optional embodiment, in the obtaining of the protein marker data in multiple samples, the proteomic profile of the sample is characterized based on iTRAQ, and the protein marker data is obtained through ELISA verification.

[0096] Specifically, the proteomic profile of the sample is characterized by isobaric Tags for Relative and Absolute Quantification (iTRAQ), and then the protein markers selected based on bioinformatics analysis are verified by enzyme linked immunosorbent assay (ELISA). The protein marker data thus obtained is used for machine learning.

[0097] iTRAQ is a labeled quantitative proteomics method, referring to an analytical method that uses iTRAQ labeling technology and mass spectrometry to quantify the proteome. ELISA is a qualitative or quantitative assay that uses antibodies to bind to and measure target molecules. In this embodiment, iTRAQ is used in the discovery phase to identify optional protein biomarkers, and ELISA is used in the validation phase to validate the protein biomarkers selected in the discovery phase.

[0098] As an optional embodiment, training the machine learning model using the training set to obtain a trained model includes:

[0099] The tear protein concentration data of K2C5, PIP and GLOD4 were used as the optimal feature subset;

[0100] Cross-validation is used to obtain model evaluation results. Based on the model evaluation results, the optimal hyperparameters are determined, and the training model is obtained according to the optimal hyperparameters.

[0101] Specifically, based on the protein biomarker data of the subject samples in the above embodiments, tear samples were collected from 224 MS group subjects and 183 GDM group subjects. For example, the proteomics profiles of the samples were characterized using iTRAQ, and the protein biomarkers in the tear samples were verified by ELISA, including K2C5, CPNS1, PIP, GLOD4, CATD, PPIB, G3P, TPIS, and ENOA (all of which are protein abbreviations). For data observation and comparison, see [link to relevant documentation]. Figure 2 It was found that the levels of certain protein markers (e.g., K2C5, CPNS1, PIP, and GLOD4) gradually increased with pregnancy and were significantly increased in the GDM group. See [link to relevant documentation]. Figure 3 The levels of certain protein markers (e.g., CATD and PPIB) are continuously upregulated during pregnancy and sharply decreased in the GDM group. See [link to relevant documentation]. Figure 4 Certain protein biomarkers (e.g., G3P, TPIS, and ENOA levels) showed significant increases in the MS group and abnormal increases in the GDM group. These protein biomarkers were validated by ELISA, and their ability to identify or classify proteins was assessed using ROC curves and AUC values. See [link to relevant documentation]. Figure 5 For example, certain protein biomarkers (e.g., G3P, PIP, TPIS, GLOD4, and CATD) can serve as the best candidate biomarkers for identifying pregnancy, while certain proteins (e.g., CATD, G3P, K2C5, CPNS1, and ENOA) can serve as the best candidate biomarkers for identifying GDM. Based on this, several suitable protein biomarkers are selected from these protein biomarkers, and the set of their corresponding tear protein concentration data is used as the optimal feature subset to train the machine learning model.

[0102] In machine learning, feature selection is a crucial issue. Its goal is to find the optimal subset of features. Feature selection eliminates irrelevant or redundant features, thereby reducing the number of features, which helps improve model accuracy and reduce runtime. Simultaneously, selecting features or feature subsets that are truly relevant to the problem simplifies the model, aids in understanding the data generation process, and avoids problems such as overfitting. Therefore, determining the optimal feature subset for machine learning models based on several protein biomarkers related to GDM diagnosis helps ensure the performance and accuracy of the diagnostic aid model.

[0103] In this embodiment, tear protein concentration data of K2C5, PIP, and GLOD4 were selected from the protein biomarker data. Specifically, combinations of K2C5 and PIP, K2C5 and GLOD4, PIP and GLOD4, and K2C5, PIP, and GLOD4 were used as feature subsets of the training data for the machine learning model. Deep neural networks, linear support vector machines, radial basis function kernel support vector machines, stochastic gradient descent, and random forests were used as the machine learning models. By plotting ROC curves and evaluating AUC values, the optimal training model and ensemble model based on the random forest model for the K2C5, PIP, and GLOD4 tear protein concentration datasets were found to have the best performance, effectively aiding in the diagnosis of gestational diabetes mellitus, with an AUC value exceeding 0.98.

[0104] Furthermore, overfitting may occur when training machine learning models. Generally, a portion of the training set can be used as a validation set to evaluate the model's training performance and adjust its hyperparameters. The validation set is used to assess the model's general error rate, and hyperparameters are adjusted accordingly to obtain optimal hyperparameters. In the above embodiment, cross-validation is used, such as the commonly used K-fold cross-validation. This involves dividing the training set into K parts, allowing for the reuse of data from the training set. Each part of the data is used both as a training set for model training and as a validation set for model evaluation, maximizing the utilization of all data.

[0105] As an optional embodiment, the ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

[0106] Specifically, the training set is used to build the model, and the test set is used to evaluate the model's metrics, such as accuracy and sensitivity. In the above embodiment, based on the tear protein concentration data of K2C5, PIP, and GLOD4 collected from tear samples of 224 MS group subjects and 183 GDM group subjects, the original dataset obtained was preprocessed and divided into training and test sets at 70% and 30%, respectively. In this embodiment, the random split was repeated 50 times to obtain multiple training and test sets to meet the requirements of model training and evaluation, and to achieve a training model and ensemble model with optimal performance (AUC = 0.981 ± 0.013).

[0107] It should be noted that if the dataset is very large, the test set does not need to be allocated proportionally; it only needs to meet the usage requirements.

[0108] Specifically, in the above embodiments, based on the tear protein concentration data of K2C5, PIP and GLOD4 collected from tear samples from 224 MS group subjects and 183 GDM group subjects, ten-fold cross-validation was used. After multiple evaluations on multiple validation sets, it can well reflect the classification or judgment performance of the auxiliary diagnostic model.

[0109] As an optional embodiment, the machine learning model includes deep neural networks, linear support vector machines, radial basis function support vector machines, stochastic gradient descent, and random forests.

[0110] In the above embodiments, different machine learning models are selected and trained accordingly based on selected protein biomarker data to determine the best-performing training model and ensemble model, which can be used as an auxiliary diagnostic model for gestational diabetes.

[0111] See Figure 6 A system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, the system comprising:

[0112] The model unit is used to select a machine learning model that can be used in the auxiliary diagnostic model;

[0113] An input unit is used to acquire protein biomarker data from multiple samples, wherein the multiple samples are collected from the tears of subjects, and the protein biomarker data is used as the raw dataset of the machine learning model.

[0114] The modeling unit is used to train and evaluate the machine learning model based on the original dataset, and obtain the auxiliary diagnostic model based on the evaluation results.

[0115] Therefore, the obtained auxiliary diagnostic model is suitable for the diagnosis of gestational diabetes mellitus in large-scale cohorts. Based on the optimized machine learning model and applicable protein biomarkers, it can effectively assist in the diagnosis of gestational diabetes mellitus, and has non-invasiveness and good performance. The optimal AUC value of the auxiliary diagnostic model can reach above 0.98.

[0116] As an optional embodiment, the modeling unit includes:

[0117] The data processing module is used to scale the original dataset and obtain preprocessed data;

[0118] The model training module is used to randomly divide the preprocessed data into a training set and a test set, train the machine learning model using the training set to obtain a trained model, evaluate the trained model using the test set, and determine the optimal training model based on the evaluation results.

[0119] The model building module is used to generate a data transformation program that can be used for a single sample based on the scaling process, integrate the data transformation program into the optimal training model to obtain an ensemble model, evaluate the performance of the ensemble model, and obtain the auxiliary diagnostic model based on the performance evaluation results.

[0120] Therefore, the data processing module scales the original dataset to obtain preprocessed data of a uniform scale. The model training module uses the preprocessed data to train the machine learning model to obtain the optimal training model. The model building module integrates the corresponding data transformation process with the optimal training model to obtain the ensemble model. After performance evaluation, the ensemble model is confirmed as an auxiliary diagnostic model for gestational diabetes mellitus. Thus, this auxiliary diagnostic model is suitable for the diagnosis of gestational diabetes mellitus in large-scale cohorts. Based on the optimal machine learning model and applicable protein biomarkers, it can effectively assist in the diagnosis of gestational diabetes mellitus and has non-invasiveness and good performance.

[0121] As an optional embodiment,

[0122] The data processing module includes a first sub-module, which can perform uniform-scale processing on the original dataset using the Z-value algorithm.

[0123] The data conversion program includes a second submodule, which processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm to adapt it to a single sample as the input data.

[0124] Therefore, the first submodule uses the Z-score algorithm to convert data of different magnitudes into the same magnitude and uses the calculated Z-score to measure them, so as to ensure the comparability between data. The second submodule processes the input data of the auxiliary diagnostic model to meet the diagnostic needs of a single sample data and can effectively simulate real clinical applications.

[0125] As an optional embodiment, the model building module further includes a performance evaluation module for plotting ROC curves and calculating the AUC value of the ensemble model.

[0126] Therefore, by plotting ROC curves and calculating AUC values ​​through the performance evaluation module, the performance of the obtained ensemble model can be effectively evaluated to confirm that the best ensemble model is the auxiliary diagnostic model for gestational diabetes.

[0127] As an optional embodiment, the protein biomarker data includes tear protein concentration data for K2C5, PIP, and GLOD4.

[0128] For a detailed description of the protein biomarkers K2C5, PIP, and GLOD4 (including full protein name, encoding gene, Gene ID, Chinese name, Uniprot receiver number, and Uniprot receiver name), please refer to the details above.

[0129] Therefore, the obtained auxiliary diagnostic model is suitable for the diagnosis of gestational diabetes mellitus in large-scale cohorts. Based on the optimized machine learning model and applicable protein biomarkers, it can effectively assist in the diagnosis of gestational diabetes mellitus, and has non-invasiveness and good performance. The optimal AUC value of the auxiliary diagnostic model can reach above 0.98.

[0130] As an optional embodiment, the optimal feature subset of the model training module is the tear protein concentration data of K2C5, PIP and GLOD4;

[0131] The model training module includes a third sub-module, which can use cross-validation to obtain model evaluation results, determine the optimal hyperparameters based on the model evaluation results, and obtain the training model based on the optimal hyperparameters.

[0132] Therefore, cross-validation through the third sub-module allows for the reuse of data from the training set. Each piece of data is used as both a training set and a validation set for model evaluation, maximizing the use of all data. Furthermore, multiple validation sets are used for multiple evaluations, which better reflects the model's classification or judgment performance.

[0133] As an optional embodiment, the ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

[0134] Specifically, based on the tear protein concentration data of K2C5, PIP, and GLOD4 collected from tear samples of 224 MS group subjects and 183 GDM group subjects, the original dataset obtained was preprocessed and divided into training and test sets at 70% and 30% respectively. In this embodiment, the random split was repeated 50 times to obtain multiple training and test sets to meet the requirements of model training and evaluation. Ten-fold cross-validation was adopted, and multiple evaluations were performed on multiple validation sets, which can well reflect the classification or judgment performance of the auxiliary diagnostic model, and the training model and ensemble model with the best performance (AUC = 0.981 ± 0.013) were achieved.

[0135] As an optional embodiment, the subjects include pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

[0136] Therefore, by detecting the dynamic changes in proteomics in the tears of normal pregnant women and pregnant women with gestational diabetes mellitus in the second trimester, multiple protein biomarkers related to pregnancy and GDM can be effectively identified.

[0137] As an optional embodiment, the machine learning model includes deep neural networks, linear support vector machines, radial basis function support vector machines, stochastic gradient descent, and random forests.

[0138] Therefore, by selecting different machine learning models and training them based on selected protein biomarker data, the best-performing training model and ensemble model can be determined for use as an auxiliary diagnostic model for gestational diabetes.

[0139] See Figure 7 An auxiliary diagnostic system for gestational diabetes mellitus, the auxiliary diagnostic system comprising:

[0140] The input module is used to input diagnostic data, which includes tear protein concentration data of the subject for K2C5, PIP and GLOD4;

[0141] A diagnostic module is used to judge the data to be diagnosed, and the diagnostic module includes an auxiliary diagnostic model obtained according to the above-described establishment method;

[0142] The output module is used to output the judgment result of the data to be diagnosed, so as to obtain the probability of the subject having gestational diabetes.

[0143] For a detailed description of the protein biomarkers K2C5, PIP, and GLOD4 (including full protein name, encoding gene, Gene ID, Chinese name, Uniprot receiver number, and Uniprot receiver name), please refer to the details above.

[0144] Therefore, the diagnostic module of this auxiliary diagnostic system includes an auxiliary diagnostic model obtained according to the above-mentioned method, which can perform rapid and accurate diagnosis of gestational diabetes based on selected protein biomarkers, realize rapid analysis and judgment of sample data and output results, thereby providing effective and non-invasive diagnosis of gestational diabetes for clinical applications.

[0145] As an optional embodiment, the auxiliary diagnostic model is obtained based on the random forest model.

[0146] In this embodiment, tear samples were collected from 224 subjects in the MS group and 183 subjects in the GDM group. The proteomic profiles of the samples were characterized using iTRAQ, and protein biomarker data in the tear samples were obtained through ELISA validation. After analysis and comparison, the tear protein concentration data of K2C5, PIP, and GLOD4 were selected from the protein biomarker data. The combinations of K2C5 and PIP, K2C5 and GLOD4, PIP and GLOD4, and K2C5, PIP, and GLOD4 were used as feature subsets of the training data for the machine learning model. Deep neural networks, linear support vector machines, radial basis function kernel support vector machines, stochastic gradient descent, and random forests were used as machine learning models. By plotting ROC curves and evaluating AUC values, the optimal training model and ensemble model based on the K2C5, PIP, and GLOD4 tear protein concentration datasets based on the random forest model were found to have the best performance, effectively assisting in the diagnosis of gestational diabetes mellitus, with an AUC value above 0.98.

[0147] When identifying pregnancy, the optimal training models of all machine learning models with K2C5, PIP and GLOD4 as the optimal feature subsets all showed high performance. However, when identifying GDM, the optimal training model of the random forest model with K2C5, PIP and GLOD4 as the optimal feature subsets achieved the best performance with an AUC of 0.981±0.013.

[0148] When identifying pregnancy, the ensemble model of all machine learning models with K2C5, PIP and GLOD4 as the optimal feature subsets showed high performance. However, when identifying GDM, the ensemble model of random forest model with K2C5, PIP and GLOD4 as the optimal feature subsets achieved the best performance with an AUC of 0.981±0.013.

[0149] Therefore, the auxiliary diagnostic model based on the random forest model and selected tear protein biomarkers demonstrated high performance in identifying GDM. Although the high performance stemmed in part from the stringent subject inclusion criteria, this study produced accurate, non-invasive tear protein biomarker signatures and improved the application of tear proteomics in GDM research.

[0150] The present invention also provides an electronic device, including: a processor; and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it is capable of implementing the establishment method as described in the above embodiments.

[0151] The present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to execute the establishment method as described in the above embodiments.

[0152] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for establishing an auxiliary diagnostic model for gestational diabetes mellitus, characterized in that, The method includes the following steps: Select a machine learning model that can be used in the auxiliary diagnostic model; Protein biomarker data were obtained from multiple samples, including tear samples from subjects. This protein biomarker data was used as the raw dataset for the machine learning model. The protein biomarker data consisted of tear protein concentration data, specifically the concentrations of selected K2C5, PIP, and GLOD4 proteins. The protein biomarker K2C5 (Type II cytoskeletal 5, gene ID: 3852, Uniploid receiver number: P13647, Uniploid receiver name: K2C5_HUMAN) was used. The protein biomarker PIP (Prolactin-inducible protein, gene ID: 5304, Uniploid receiver number: P12273, Uniploid receiver name: PIP_HUMAN) was also used. The protein biomarker GLOD4 (Glyoxalase Domain-containing protein) was used.

4. Encoding gene name: GLOD4, Gene ID: 51031, Chinese name: glyoxal functional domain covered protein 4, Uniprot receiver number: Q9HC38, Uniprot receiver name: GLOD4_HUMAN; The machine learning model is trained and evaluated based on the original dataset, and the auxiliary diagnostic model is obtained based on the evaluation results. The combination of K2C5 and PIP, K2C5 and GLOD4, PIP and GLOD4, and K2C5, PIP and GLOD4 are respectively used as feature subsets of the training data. The training and evaluation of the machine learning model based on the original dataset includes: The original dataset is scaled to obtain preprocessed data; The preprocessed data is randomly divided into a training set and a test set. The machine learning model is trained using the training set to obtain a trained model, and the trained model is evaluated using the test set. The optimal trained model is determined based on the evaluation results. Based on the scaling process, a data transformation program that can be used for a single sample is generated. The data transformation program is integrated into the optimal training model to obtain an ensemble model. The performance of the ensemble model is evaluated, and the auxiliary diagnostic model is obtained based on the performance evaluation results. The performance evaluation of the ensemble model includes: plotting the ROC curve, calculating the AUC value of the ensemble model, and obtaining the ensemble model based on the random forest model with K2C5, PIP, and GLOD4 tear protein concentration data as the optimal feature subset.

2. The method for establishing as described in claim 1, characterized in that, The scaling process includes uniform scaling of the original dataset using a Z-value algorithm; The data conversion program processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm, so that it is suitable for a single sample as the input data.

3. The method for establishing as described in claim 1, characterized in that, The step of training the machine learning model using the training set to obtain the trained model includes: The tear protein concentration data of K2C5, PIP and GLOD4 were used as the optimal feature subset; Cross-validation is used to obtain model evaluation results. Based on the model evaluation results, the optimal hyperparameters are determined, and the training model is obtained according to the optimal hyperparameters.

4. The method for establishing as described in claim 3, characterized in that, The ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

5. The method for establishing as described in claim 1, characterized in that, The subjects included pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

6. A system for establishing an auxiliary diagnostic model for gestational diabetes mellitus, characterized in that, The system establishment includes: The model unit is used to select a machine learning model that can be used in the auxiliary diagnostic model; The input unit is used to acquire protein biomarker data from multiple samples, wherein the samples are tear samples from the subjects, and the protein biomarker data is used as the raw dataset of the machine learning model. The tear protein concentration data of K2C5, PIP and GLOD4 are selected from the protein biomarker data. A modeling unit is used to train and evaluate the machine learning model based on the original dataset, and obtain the auxiliary diagnostic model based on the evaluation results, wherein the combination of K2C5 and PIP, the combination of K2C5 and GLOD4, the combination of PIP and GLOD4, and the combination of K2C5, PIP and GLOD4 are respectively used as feature subsets of the training data. The modeling unit includes: The data processing module is used to scale the original dataset and obtain preprocessed data; The model training module is used to randomly divide the preprocessed data into a training set and a test set, train the machine learning model using the training set to obtain a trained model, evaluate the trained model using the test set, and determine the optimal training model based on the evaluation results. The model building module is used to generate a data transformation program that can be used for a single sample based on the scaling process, integrate the data transformation program into the optimal training model to obtain an ensemble model, evaluate the performance of the ensemble model, and obtain the auxiliary diagnostic model based on the performance evaluation results. The model building module also includes a performance evaluation module, which is used to plot the ROC curve, calculate the AUC value of the ensemble model, and obtain the ensemble model based on the random forest model with K2C5, PIP and GLOD4 tear protein concentration data as the optimal feature subset.

7. The system as described in claim 6, characterized in that, The data processing module includes a first sub-module, which can perform uniform-scale processing on the original dataset using the Z-value algorithm. The data conversion program includes a second submodule, which processes the input data of the auxiliary diagnostic model according to the parameters used in the Z-value algorithm to adapt it to a single sample as the input data.

8. The system as described in claim 6, characterized in that, The optimal feature subset of the model training module is the tear protein concentration data of K2C5, PIP, and GLOD4. The model training module includes a third sub-module, which can use cross-validation to obtain model evaluation results, determine the optimal hyperparameters based on the model evaluation results, and obtain the training model based on the optimal hyperparameters.

9. The system as described in claim 8, characterized in that, The ratio of the training set to the test set is 7:3, and the cross-validation is 10-fold cross-validation.

10. The system as described in claim 6, characterized in that, The subjects included pregnant women with normal pregnancy in the second trimester and pregnant women with gestational diabetes in the second trimester.

11. An auxiliary diagnostic system for gestational diabetes mellitus, characterized in that, The auxiliary diagnostic system includes: The input module is used to input diagnostic data, which includes tear protein concentration data of the subject for K2C5, PIP and GLOD4; A diagnostic module is used to judge the data to be diagnosed, and the diagnostic module includes an auxiliary diagnostic model obtained by the establishment method according to any one of claims 1 to 5; The output module is used to output the judgment result of the data to be diagnosed, so as to obtain the probability of the subject having gestational diabetes.

12. An electronic device, characterized in that, include: processor; as well as A memory storing a computer program, which, when executed by the processor, enables the creation method as described in any one of claims 1 to 5.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is used to run in order to implement the establishment method as described in any one of claims 1 to 5.

Images