Anti-gamma interferon autoantibody positive prediction method and system based on machine learning

Abstract

The invention provides an anti-gamma interferon autoantibody positive prediction method and system based on machine learning, and relates to the technical field of machine learning. The method comprises the steps of data acquisition and preprocessing, feature coding and screening, model training and integration, probability output, risk classification and the like. Basic disease information, physical examination data and conventional laboratory indexes are fused to construct a standardized numerical value feature matrix, and a regularization screening and group sparse constraint mechanism is introduced to extract target features. A plurality of supervised machine learning models are adopted to construct a prediction system, and an individual AIGAs positive probability is output in combination with a cross validation and ensemble learning method. Finally, clinical decision is assisted through dynamic threshold setting and visual result display, high precision, high interpretability and good deployability are achieved, and a practical solution is provided for AIGAs intelligent screening in a resource limited scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a method and system for predicting positive anti-interferon gamma autoantibodies based on machine learning. BACKGROUND

[0002] Anti-interferon gamma autoantibody (AIGAs) immunodeficiency syndrome is a kind of acquired immunodeficiency disease in adults, and patients often have multiple sites, repeated and opportunistic infections, but there is no unified diagnostic standard in the world at present. Clinically, it mainly depends on laboratory detection such as ELISA, flow cytometry or Western blot to qualitatively or quantitatively detect AIGAs titer and neutralization activity, the related detection kit is high in price and complicated in process, and is limited to large scientific research institutions or university affiliated hospitals, so that it is difficult for primary hospitals or general hospitals to quickly diagnose, resulting in misdiagnosis or delayed treatment.

[0003] With the digitization of electronic medical records, laboratory examinations and image data, machine learning has been rapidly promoted in medical diagnosis, disease risk prediction and other scenarios. A large number of studies have shown that based on the real-time acquisition of conventional clinical indicators, the prediction model can not only assist in early screening of insufficient manual experience, but also use integrated learning to significantly improve the prediction sensitivity, specificity and generalization ability, and reduce the overall diagnostic cost.

[0004] The current AIGAs detection method highly depends on special reagents and equipment, and cannot be used for epidemiological screening in large sample populations. At the same time, the existing machine learning workflow often lacks feature screening strategies, model integration schemes and external verification mechanisms for AIGAs, making it difficult to ensure the applicability of the model to different hospitals and different ethnic groups. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide an anti-interferon gamma autoantibody positive prediction method and system based on machine learning, which takes into account accuracy, operability and clinical practicability, and provides an efficient and low-cost solution for intelligent screening of AIGAs positive.

[0006] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0007] A method for predicting positive anti-interferon gamma autoantibodies based on machine learning, comprising:

[0008] Collecting the basic disease information, physical examination information of the test object and the laboratory indexes obtained by routine hematology, biochemistry and immunology detection to form an original data sample, and performing missing value filling, outlier rejection and data standardization on the original data sample to obtain a training sample set;

[0009] encoding the training sample set as a numerical feature vector, and screening target features associated with AIGAs positivity by using a regularization algorithm;

[0010] training at least two different types of supervised machine learning models based on the target features and corresponding positive / negative labels;

[0011] performing cross-validation on the supervised machine learning models, selecting a candidate model set according to a preset performance threshold, and implementing ensemble learning on the candidate model set to output an AIGAs positivity probability of the test object;

[0012] comparing the positivity probability with a preset custom threshold, generating a classification result and performing visual display to assist clinical decision-making.

[0013] Preferably, the basic disease information, physical examination information, and laboratory indicators obtained through routine hematology, biochemistry, and immunology detection of the test object are collected to form an original data sample, and the original data sample is subjected to missing value filling, outlier removal, and data standardization to obtain a training sample set, including:

[0014] retrieving the presence or absence of chronic or infectious diseases from an electronic medical record system and encoding them as binary variables to obtain the basic disease information; the chronic or infectious diseases include hepatitis B, chronic kidney disease, diabetes, tuberculosis, and HIV infection;

[0015] measuring and recording the body temperature, systolic and diastolic blood pressure, heart rate, respiratory rate, and body mass index (BMI) by medical personnel to obtain the physical examination information;

[0016] importing quantitative results from a laboratory information system to obtain the laboratory indicators; the quantitative results include white blood cell count, lymphocyte count, CD4 + T cell count, immunoglobulin G, immunoglobulin E, serum globulin, C-reactive protein, and erythrocyte sedimentation rate;

[0017] If the missing rate of any of the basic disease information, physical examination information, and laboratory indicators exceeds 0.33, the indicator is deleted;

[0018] performing Shapiro-Wilk normality test on the retained indicators, and if the test is passed, filling the missing values with the sample mean, otherwise filling the missing values with the sample median;

[0019] calculating the first quartile Q1 and the third quartile Q3 for each numerical indicator, and obtaining the interquartile range IQR = Q3−Q1;

[0020] deleting sample values less than Q1−1.5IQR or greater than Q3+1.5IQR;

[0021] Z-score processing is performed on the remaining numerical indicators to make the mean value 0 and the standard deviation 1, so as to obtain the training sample set.

[0022] Preferably, the training sample set is encoded as a numerical feature vector, and a regularization algorithm is used to screen target features positively correlated with AIGAs, including:

[0023] For each chronic or infectious disease obtained in the training sample set, a one-hot vector is established according to the ICD-10 three-digit code, and the one-hot vectors are aggregated according to the system to which the disease belongs, to generate a 4-dimensional hierarchical disease vector; the system to which the disease belongs includes respiratory, digestive, endocrine and immune;

[0024] The physical examination information in the training sample set is subjected to Z-score standardization to obtain a physical examination vector;

[0025] Each continuous indicator of the laboratory indicators in the training sample set is subjected to Box-Cox transformation to weaken skewness, and then subjected to Z-score standardization to obtain a laboratory indicator vector;

[0026] The unified numerical feature matrix is formed by splicing in the order of the disease vector, the physical examination vector and the laboratory indicator vector;

[0027] Elastic net logistic regression is performed on the unified numerical feature matrix with AIGAs positive / negative label as the dependent variable, the regularization coefficient is determined using 5-fold cross-validation, and the features with regression coefficient absolute value greater than zero are retained;

[0028] The remaining features are divided into groups according to three types of data modalities, and a Group-LASSO model with group sparse constraint is applied to delete features with all coefficients in the group shrunk to zero, to obtain the target features; the data modalities include disease, physical examination and laboratory.

[0029] Preferably, the remaining features are divided into groups according to three types of data modalities, and a Group-LASSO model with group sparse constraint is applied to delete features with all coefficients in the group shrunk to zero, to obtain the target features, including:

[0030] The remaining features are divided into a basic disease group, a physical examination group and a laboratory indicator group, respectively denoted as index sets and ;

[0031] The objective function is established as ; wherein, is the number of samples in the training sample set; is a sample label vector, 1 for positive and 0 for negative; is a standardized feature matrix arranged in rows; is a regression coefficient vector to be solved; is a sub-vector on the k-th set of feature index ; is a sub-vector on the k-th set of feature index ;is a regularization coefficient; is the number of features in the k-th set;

[0032] The square root-iterative threshold algorithm is adopted to perform group sparse shrinkage update on until the relative change rate of the objective function is less than ;

[0033] If the norm of a group is zero, all features corresponding to the group with zero norm are deleted;

[0034] The features with remaining non-zero coefficients are combined to obtain the target features.

[0035] Preferably, based on the target features and corresponding positive / negative labels, at least two different types of supervised machine learning models are trained, including:

[0036] a) stratified sampling is used on the target feature matrix and labels, and the ratio of 8:2 is used to divide the training subset and the validation subset;

[0037] b) the gradient boosting decision tree model is constructed using the training subset, and the area under the receiver operating characteristic curve (AUC) of the validation subset is used as the monitoring index after each iteration. If the AUC does not improve for ten consecutive rounds, the iteration is stopped;

[0038] c) the linear support vector machine model is constructed using the training subset, and the penalty coefficient C is automatically searched through five-fold cross-validation. The C value with the maximum AUC of the validation subset is used as the final parameter.

[0039] Preferably, cross-validation is performed on the supervised machine learning model, a candidate model set is selected according to a preset performance threshold, and ensemble learning is performed on the candidate model set to output the AIGAs positive probability of the test object, including:

[0040] Each of the supervised machine learning models is subjected to k-fold stratified cross-validation, and the evaluation results of each fold are calculated respectively. The evaluation results include: area under the curve (AUC), sensitivity (SEN), and specificity (SPE);

[0041] The evaluation results of each fold are arithmetically averaged to obtain the overall performance vector of the model (AUC i , SEN i , SPEi );

[0042] candidate model screening: if AUC i ≥ 0.90 and SEN i ≥ 0.85, the ith model is included in the candidate model set, otherwise discarded;

[0043] Pratt scaling is performed on each model in the candidate model set using the validation fold output to map the original confidence to the calibrated probability ;

[0044] According to the results of AUC i , the weight is set in a normalized manner ; where m is the number of candidate models;

[0045] For a single test object, the calibrated probabilities output by each candidate model are weighted and summed to obtain the final positive probability .

[0046] Preferably, the positive probability is compared with a preset custom threshold to generate a classification result and visual display to assist clinical decision-making, including:

[0047] A threshold input interface is provided on the user terminal, and a user-defined threshold T is stored in the system parameter cache; if the user does not input, the default threshold T0=0.5 is read;

[0048] The positive probability P final of a single test object is compared with the threshold T, and if P final ≥ T, it is marked as "positive risk", otherwise as "negative risk";

[0049] The probability interval [0,1] is divided into three segments: [0,T) is low risk, [T,0.75) is medium risk, and [0.75,1] is high risk, and green, yellow, and red visual identifiers are assigned to each segment;

[0050] The visualization module is called on the user terminal, and P final is drawn in the form of a bar progress bar or a circular dial, and "positive risk" or "negative risk" and the visual identifier are superimposed;

[0051] The numerical summary of the corresponding key features of the underlying disease, physical examination and laboratory indicators is displayed synchronously on the visualization interface, and the user is allowed to adjust the threshold T to trigger real-time re-determination, so that the clinician can optimize the decision-making according to the specific situation of the patient.

[0052] A machine learning-based anti-IFN autoantibody positive prediction system, comprising:

[0053] a clinical data acquisition and preprocessing unit configured to acquire basic disease information, physical examination information, and laboratory indexes obtained through routine hematology, biochemistry, and immunology detection of the subject to be tested, form an original data sample, and perform missing value filling, outlier removal, and data standardization on the original data sample to obtain a training sample set;

[0054] a feature encoding and screening unit configured to encode the training sample set into a numerical feature vector and screen target features related to AIGAs positivity using a regularization algorithm;

[0055] a multi-model training unit configured to train at least two different types of supervised machine learning models based on the target features and corresponding positive / negative labels;

[0056] a cross-validation and ensemble learning unit configured to perform cross-validation on the supervised machine learning models, select a candidate model set according to a preset performance threshold, and implement ensemble learning on the candidate model set to output an AIGAs positivity probability of the subject to be tested;

[0057] a classification determination and visualization display unit configured to compare the positivity probability with a preset custom threshold, generate a classification result, and perform visual display to assist clinical decision-making.

[0058] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0059] The present application can make full use of conventional and easily accessible clinical information, and effectively improve the recognition ability of AIGAs positive risk through data standardization, feature selection, and model fusion. This method avoids the dependence on expensive and complex laboratory tests, realizes the rapid preliminary screening and risk grading of patient groups, and is especially suitable for resource-limited primary medical scenarios. In addition, through the custom threshold and visual interactive interface, the clinician is provided with flexible and intuitive auxiliary decision support, which enhances the explainability and applicability of the model results. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0061] Figure 1 a method flowchart provided for the embodiments of the present application;

[0062] Figure 2 a lasso regression cross-validation curve diagram provided for the embodiments of the present application;

[0063] Figure 3 AIGAs-positive decision tree prediction model schematic diagram provided for an embodiment of the present application;

[0064] Figure 4 AIGAs-positive decision tree prediction model ROC curve schematic diagram provided for an embodiment of the present application;

[0065] Figure 5 Random forest model prediction factor importance ranking schematic diagram provided for an embodiment of the present application;

[0066] Figure 6 AIGAs-positive random forest prediction model ROC curve schematic diagram provided for an embodiment of the present application;

[0067] Figure 7 AIGAs-positive support vector machine model ROC curve schematic diagram provided for an embodiment of the present application;

[0068] Figure 8 AIGAs-positive XGBoost prediction model ROC curve schematic diagram provided for an embodiment of the present application;

[0069] Figure 9 XGBoost model prediction factor importance score schematic diagram provided for an embodiment of the present application;

[0070] Figure 10 Artificial neural network model topology diagram provided for an embodiment of the present application;

[0071] Figure 11 AIGAs-positive artificial neural network model ROC curve schematic diagram provided for an embodiment of the present application. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0073] The purpose of the present application is to provide a machine learning-based anti-interferon gamma autoantibody positive prediction method and system, which takes into account accuracy, operability and clinical practicability, and provides an efficient and low-cost solution for intelligent screening of AIGAs-positive.

[0074] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0075] Figure 1 The method flowchart provided by the embodiment of the present application is shown in Figure 1 The present application provides a machine learning-based positive prediction method for anti-interferon gamma autoantibodies, comprising:

[0076] Step 100: Collecting the basic disease information, physical examination information and laboratory indexes obtained through routine hematology, biochemistry and immunology detection of the subject to be tested to form an original data sample, and performing missing value filling, outlier removal and data standardization on the original data sample to obtain a training sample set;

[0077] Step 200: Encoding the training sample set into a numerical feature vector, and screening target features related to AIGAs positivity using a regularization algorithm;

[0078] Step 300: Training at least two different types of supervised machine learning models based on the target features and corresponding positive / negative labels;

[0079] Step 400: Performing cross-validation on the supervised machine learning models, selecting a candidate model set according to a preset performance threshold, and implementing ensemble learning on the candidate model set to output the AIGAs positivity probability of the subject to be tested;

[0080] Step 500: Comparing the positivity probability with a preset custom threshold to generate a classification result and perform visual display to assist clinical decision-making.

[0081] Preferably, the basic disease information, physical examination information and laboratory indexes obtained through routine hematology, biochemistry and immunology detection of the subject to be tested are collected to form an original data sample, and the original data sample is subjected to missing value filling, outlier removal and data standardization to obtain a training sample set, comprising:

[0082] The presence or absence of chronic or infectious diseases is retrieved from an electronic medical record system and encoded as a binary variable to obtain the basic disease information; the chronic or infectious diseases include hepatitis B, chronic kidney disease, diabetes, tuberculosis and HIV infection;

[0083] The body temperature, systolic and diastolic blood pressure, heart rate, respiratory rate and body mass index (BMI) are measured and recorded by medical personnel to obtain the physical examination information;

[0084] The quantitative results are imported from a laboratory information system to obtain the laboratory indexes; the quantitative results include white blood cell count, lymphocyte count, CD4 +T cell count, immunoglobulin G, immunoglobulin E, serum globulin, C-reactive protein, and erythrocyte sedimentation rate;

[0085] If the missing rate of any of the basic disease information, the physical examination information, and the laboratory indicators exceeds 0.33, the indicator is deleted;

[0086] A Shapiro-Wilk normality test is performed on the retained indicators, and if the test is passed, the missing values are filled with the sample mean, otherwise the missing values are filled with the sample median;

[0087] For each numerical indicator, the first quartile Q1 and the third quartile Q3 are calculated, and the interquartile range IQR = Q3 - Q1 is obtained;

[0088] Sample values less than Q1 - 1.5IQR or greater than Q3 + 1.5IQR are deleted;

[0089] Z-score processing is performed on the remaining numerical indicators to obtain a training sample set with a mean of 0 and a standard deviation of 1.

[0090] Illustratively, Q1 is the value at the 25% position after ascending arrangement, referred to as the first quartile. Q3 is the value at the 75% position after ascending arrangement, referred to as the third quartile. IQR is the difference between Q3 and Q1, used to measure the intermediate dispersion of data distribution, and is commonly used for outlier detection.

[0091] Specifically, the embodiment conducts a 36-month prospective and retrospective bidirectional study in a hospital, collects clinical data of AIGAs positive and negative patients, including patient medical history, physical examination results, laboratory examination results, etc. The collected data is cleaned and preprocessed, and indicators with excessive missing values are deleted, and the missing values are filled. For normally distributed indicators, the mean value is filled; for non-normally distributed indicators, the median is filled.

[0092] Preferably, the training sample set is encoded into a numerical feature vector, and a regularization algorithm is used to screen target features related to AIGAs positivity, including:

[0093] For each chronic or infectious disease obtained in the training sample set, a one-hot vector is established according to the ICD-10 three-digit code, and the one-hot vectors are aggregated according to the disease belonging system to generate a 4-dimensional hierarchical disease vector; the disease belonging system includes respiratory, digestive, endocrine, and immune;

[0094] The physical examination information in the training sample set is subjected to Z-score standardization to obtain a physical examination vector;

[0095] performing Box-Cox transformation to weaken skewness and then performing Z-score standardization on each continuous index of the laboratory index in the training sample set to obtain a laboratory index vector;

[0096] splicing the disease vector, the physical examination vector and the laboratory index vector in order to form a unified numerical feature matrix;

[0097] performing elastic net logistic regression on the unified numerical feature matrix with AIGAs positive / negative label as the dependent variable, determining the regularization coefficient using 5-fold cross-validation, and retaining the features with absolute value of regression coefficient greater than zero;

[0098] grouping the retained features according to three types of data modalities, applying Group-LASSO model with group sparsity constraint, deleting the features with all coefficients in the group shrunk to zero, and obtaining the target features; the data modalities include disease, physical examination and laboratory.

[0099] In this embodiment, first, the three types of core information in the training sample set, i.e., basic diseases, physical examinations and laboratory indexes, are structured and uniformly processed. The basic disease information is extracted according to the national standard disease classification code and converted into an aggregated variable representing a specific disease category (such as respiratory, digestive, endocrine and immune system), thereby reducing the dimension while retaining the clinical semantics. The physical examination information is processed by standardization to eliminate the dimensional differences between different indexes, and a combined feature reflecting the relationship between body mass and blood pressure is additionally constructed to improve the model's ability to identify individual risk status. After cleaning, the laboratory index data is adjusted for skewness distribution, and then normalized to ensure that detection items of different sources and different numerical ranges can be efficiently integrated in the same modeling framework. Finally, the above three types of features are spliced in a predetermined order to form a structured numerical feature matrix, which serves as the input basis for model training.

[0100] To improve the generalization ability of the prediction model and ensure the clinical interpretability of the results, this embodiment introduces a two-stage feature screening mechanism. In the first stage, a logistic regression method with a penalty mechanism is used to preliminarily screen all the coded features, and only a feature set with strong association with the AIGAs positive state is retained. In the second stage, the preliminarily screened features are grouped according to the data modalities based on medical semantics, and a feature compression method based on group dimension is further introduced to evaluate the contribution of each group to the prediction task as a whole. If the contribution of a group is very low, all features corresponding to that group are removed. This strategy not only helps to remove redundant information and reduce the risk of overfitting, but also strengthens the correspondence between the model structure and the clinical logic, making the final retained target features more stable and interpretable, especially suitable for building auxiliary diagnosis models deployed in real hospital environments.

[0101] Preferably, the retained features are grouped according to the three data modalities, and the Group-LASSO model with group sparsity constraint is applied to delete the features whose coefficients in the group are shrunk to zero to obtain the target features, including:

[0102] The retained features are divided into basic disease group, physical examination group and laboratory index group, respectively recorded as index sets and ;

[0103] Establishing the objective function ;in, is the number of samples in the training sample set; is the sample label vector, positive is 1 and negative is 0; is the normalized feature matrix of the permutation; is the regression coefficient vector to be determined; for In the kth group of feature index sets subvectors on ; is the regularization coefficient; is the number of features in the kth group;

[0104] The square root-iterative threshold algorithm is used to Perform group sparse shrinkage update until the relative change rate of the objective function is less than ;

[0105] If a group of If the norm is zero, all features corresponding to the group with zero norm are deleted;

[0106] The features containing the remaining non-zero coefficients are merged to obtain the target feature.

[0107] In this embodiment, in order to improve the structured expression ability and clinical interpretability of feature selection, the features after preliminary screening are preferably divided into three functional feature groups according to the data source, namely the basic disease group, the physical examination group, and the laboratory index group, to form a corresponding feature index set. Subsequently, a regression model with group sparse constraints is constructed based on these feature groups, and the contribution of each group of features to the AIGAs positive prediction task is judged holistically. During the model training process, an iterative optimization algorithm is used to update the regression coefficients group by group. After each round of iteration, the relative rate of change of the objective function is calculated to determine whether the convergence condition has been met.

[0108] Specifically, when a certain feature group shows very low predictive contribution in multiple rounds of iterations, the corresponding parameter value will be gradually shrunk to zero. Once the shrinkage result is stable at zero, it is considered that this group of features has no actual value for the classification task, and it will be removed from the model as a whole. Finally, all the features corresponding to the remaining non-zero coefficients form a new feature set, which serves as the core input variable for subsequent model training and inference. This group-by-group compression mechanism not only effectively controls the model complexity caused by high-dimensional features, but also enhances the model's ability to identify the importance between different data modalities, providing a structural foundation for building clinically reliable intelligent auxiliary tools.

[0109] Preferably, based on the target features and corresponding positive / negative labels, at least two different types of supervised machine learning models are trained respectively, including:

[0110] a) Stratified sampling is used on the target feature matrix and labels, and an 8:2 ratio is used to divide them into a training subset and a validation subset;

[0111] b) A gradient boosting decision tree model is constructed using the training subset, and the area under the receiver operating characteristic curve (AUC) of the validation subset is used as a monitoring indicator after each iteration. If there is no improvement in AUC for ten consecutive rounds, the iteration is stopped;

[0112] c) A linear support vector machine model is constructed using the training subset, and the penalty coefficient C is automatically searched through five-fold cross-validation. The C value that maximizes the AUC of the validation subset is used as the final parameter.

[0113] Specifically, the program packages of machine learning algorithms such as random forest, decision tree, support vector machine, XGBoost, and artificial neural network are installed and called in the R language environment, the selected predictive factors are used as input, and AIGAs positive or negative is used as output to construct a prediction model. The parameters of the model are set, such as the number of decision trees in random forest and the maximum depth in XGBoost, and the model is trained and optimized through cross-validation and other methods.

[0114] Preferably, cross-validation is performed on the supervised machine learning model, a candidate model set is selected according to a pre-set performance threshold, and ensemble learning is performed on the candidate model set to output the AIGAs positive probability of the test object, including:

[0115] Each of the supervised machine learning models is subjected to k-fold stratified cross-validation, and the evaluation results of each fold are calculated respectively; the evaluation results include: area under the curve (AUC), sensitivity (SEN), and specificity (SPE);

[0116] The evaluation results of each fold are arithmetically averaged to obtain a model overall performance vector (AUC i , SEN i , SPE i);

[0117] Candidate model screening: if AUC i ≥ 0.90 and SEN i ≥ 0.85, the ith model is included in the candidate model set, otherwise discarded;

[0118] Pratt scaling is performed on each model in the candidate model set using the validation fold output to map the original confidence to the calibrated probability .

[0119] According to the results of AUC i , the weights are set in a normalized manner ; where m is the number of candidate models;

[0120] For a single test object, the calibrated probabilities output by each candidate model are weighted and summed to obtain the final positive probability .

[0121] Further, the present embodiment first screens the general conditions, clinical symptoms and signs, laboratory test results, etc. that are different between the AIGAs-positive and AIGAs-negative groups, selects 29 indicators that can be obtained at the time of admission or within a short time after admission, and includes these prediction indicators in the lasso regression. As shown in Figure 2 , when lambda is 0.076, the selected variables include 11 prediction factors such as the number of underlying diseases (No. of underlying diseases), superficial lymphadenopathy (Superficiallymphadenopathy), disseminated infection (Disseminated infection), white blood cell count (WBC), lymphocyte count (L), CD4+ T lymphocyte count (CD4+ T), IgG, IgE, GLB, CRP, and ESR.

[0122] The present embodiment uses different machine learning algorithms, uses the 11 prediction factors selected by lasso regression as model prediction indicators, and uses AIGAs positivity or negativity as outcome indicators to construct an IGAs-positive machine learning model. A random sequence is set (random seed is 2024), and the data is divided into a training set (166 cases, 91 AIGAs-positive patients and 75 AIGAs-negative patients) and a test set (72 cases, 42 AIGAs-positive patients and 30 AIGAs-negative patients) in a ratio of 7:3. All subsequent machine learning methods use these two data subsets as training samples and test samples. As shown in Figure 3The AIGAs-positive decision tree prediction model was constructed by training samples. The test samples were substituted into the model, and the prediction probability was calculated. The ROC curve was drawn with the prediction probability value of the optimal model and the actual classification result of the test sample, as shown in Figure 4 The area under the curve was 0.923, and the 95% confidence interval was [0.866-0.980]. The judgment result of the decision tree model on the test sample was: 34 cases of AIGAs-positive patients were correctly judged, and 27 cases of AIGAs-negative patients were correctly judged; the accuracy (consistency rate) was 84.72%, the sensitivity was 80.95%, the specificity was 90.00%, the Youden index was 0.71, the positive predictive value was 91.89%, and the negative predictive value was 77.14%.

[0123] The AIGAs-positive random forest prediction model was constructed by training samples, and the importance of the prediction factors was sorted as shown in Figure 5 The test samples were substituted into the model, and the prediction classification result was calculated. The ROC curve was drawn with the prediction classification result of the optimal model and the actual classification of the test sample, as shown in Figure 6 The area under the curve was 0.948, and the 95% confidence interval was [0.899-0.996]. The judgment result of the random forest model on the test sample was: 40 cases of AIGAs-positive patients were correctly judged, and 25 cases of AIGAs-negative patients were correctly judged; the accuracy was 90.28%, the sensitivity was 95.24%, the specificity was 83.33%, the Youden index was 0.77, the positive predictive value was 88.89%, and the negative predictive value was 92.59%.

[0124] The AIGAs-positive support vector machine prediction model was constructed by training samples, and the test samples were substituted into the model. The prediction probability was calculated, and the ROC curve was drawn with the prediction probability value of the optimal model and the actual classification result of the test sample, as shown in Figure 7 The area under the curve was 0.935, and the 95% confidence interval was [0.881-0.989]. The judgment result of the support vector machine model on the test sample was: 40 cases of AIGAs-positive patients were correctly judged, and 24 cases of AIGAs-negative patients were correctly judged; the accuracy was 88.89%, the sensitivity was 95.24%, the specificity was 80.00%, the Youden index was 0.75, the positive predictive value was 86.96%, and the negative predictive value was 92.31%.

[0125] The AIGAs-positive XGBoost prediction model was constructed by training samples, and the test samples were substituted into the model. The prediction probability was calculated, and the ROC curve was drawn with the prediction probability value of the optimal model and the actual classification result of the test sample, as shown in Figure 8The area under the curve is 0.965, and the 95% confidence interval is [0.925-1.000]. The judgment results of the XGBoost model on the test samples are: 38 cases of AIGAs positive patients are correctly judged, and 28 cases of AIGAs negative patients are correctly judged; the accuracy rate is 91.67%, the sensitivity is 90.48%, the specificity is 93.33%, the Youden index is 0.85, the positive predictive value is 95.00%, and the negative predictive value is 87.50%. The XGBoost model also gives a score of 11 prediction factors on the importance of the model, and the higher the score, the more important it is, such as Figure 9 As shown in the figure.

[0126] An artificial neural network prediction model for AIGAs positive patients is constructed by training samples, as shown in Figure 10 The test sample is substituted into the model, the prediction probability is calculated, and the ROC curve is drawn with the prediction probability value of the optimal model and the actual classification result of the test sample, as shown in Figure 11 The area under the curve is 0.935, and the 95% confidence interval is [0.884-0.986]. The judgment results of the artificial neural network model on the test samples are: 38 cases of AIGAs positive patients are correctly judged, and 24 cases of AIGAs negative patients are correctly judged; the accuracy rate is 86.11%, the sensitivity is 90.48%, the specificity is 80.00%, the Youden index is 0.70, the positive predictive value is 86.36%, and the negative predictive value is 95.71%.

[0127] Specifically, 10 AIGAs positive patients and 10 AIGAs negative patients in the external validation data are substituted into 5 different machine learning prediction models, and the results are shown in Table 1. The decision tree model correctly judges 8 positive patients and 9 negative patients; the random forest model correctly judges 10 positive patients and 8 negative patients; the support vector machine model correctly judges 10 positive patients and 9 negative patients; the XGBoost model correctly judges 8 positive patients and 9 negative patients; and the artificial neural network model correctly judges 8 positive patients and 9 negative patients.

[0128] As shown in Table 1, the samples incorrectly judged by different machine learning models in the external validation data are not the same. By integrating 5 algorithms, more than 3 algorithms are finally judged as positive, i.e. judged as positive, and vice versa. The correct judgment rate of 10 AIGAs positive patients is 100.00%, and the correct judgment rate of 10 AIGAs negative patients is 90.00%.

[0129] Table 1. External data validation and integrated algorithm prediction of AIGAs-positive results of different machine learning models

[0130] External verification sample number Decision tree model Random forest model Support vector machine model XGBoost model Artificial neural network model Integration algorithm Final judgment result Positive 1 + + + + + 4 / 5 Positive Positive 2 + + + + + 5 / 5 Positive Positive 3 - + + - + 3 / 5 Positive Positive 4 + + + + - 4 / 5 Positive Positive 5 + + + + + 5 / 5 Positive Positive 6 + + + + + 5 / 5 Positive Positive 7 + + + + + 5 / 5 Positive Positive 8 - + + - + 3 / 5 Positive Positive 9 + + + + + 5 / 5 Positive Positive 10 + + + + + 5 / 5 Positive Positive sample accuracy % 80 100 100 80 90 100 Negative 1 - - - - - 0 / 5 Negative Negative 2 - - - - - 0 / 5 Negative Negative 3 - - - - - 0 / 5 Negative Negative 4 + - - - - 1 / 5 Negative Negative 5 - + + + + 4 / 5 Positive Negative 6 - - - - - 0 / 5 Negative Negative 7 - - - - - 0 / 5 Negative Negative 8 - - - - - 0 / 5 Negative Negative 9 - - - - - 0 / 5 Negative Negative 10 - + - - + 2 / 5 Negative Negative sample accuracy % 90 80 90 90 80 90

[0131] Preferably, the positive probability is compared with a preset custom threshold value, a classification result is generated and visualized to assist clinical decision-making, including:

[0132] A threshold value input interface is provided on the user terminal, and a user-defined threshold value T is stored in the system parameter cache; if the user does not input, the default threshold value T0=0.5 is read;

[0133] The positive probability P of a single test object final Compared with the threshold value T, if P final ≥T, it is marked as “positive risk”, otherwise it is marked as “negative risk”;

[0134] The probability interval [0, 1] is divided into three segments: [0, T) is low risk, [T, 0.75) is medium risk, and [0.75, 1] is high risk, and green, yellow and red visual identifiers are assigned to each segment;

[0135] The user terminal calls the visualization module, and P final is drawn in the form of a bar progress bar or a circular instrument panel, and is superimposed with “positive risk” or “negative risk” and the visual identifier;

[0136] The numerical summary of the corresponding key features of the underlying disease, physical examination and laboratory indicators is synchronously displayed on the visualization interface, and the user is allowed to adjust the threshold value T to trigger real-time re-determination, so that the clinician can optimize the decision according to the specific situation of the patient.

[0137] In this embodiment, a user-defined threshold mechanism is preferably introduced to enhance the flexibility and clinical adaptability of the model prediction results. The system provides a threshold value input interface on the user terminal, allowing physicians to input personalized judgment thresholds according to different clinical backgrounds or population characteristics. If the user does not set it, the default threshold value 0.5 is automatically used as the judgment standard. The system compares the positive probability of the test object with the threshold value in real time, and if the result is higher than or equal to the threshold value, it is marked as “positive risk”, otherwise it is marked as “negative risk”. In addition, in order to facilitate intuitive understanding and risk classification, the system divides the positive probability into three intervals, corresponding to low, medium and high risk, and uses green, yellow and red for visual coding, thereby realizing the quantitative and perceptible expression of the prediction results.

[0138] Further, to improve the interpretability and practicality of the prediction result, the system calls a graphics module at the visualization terminal, renders the positive probability output by the model in the form of a bar progress bar or a circular instrument panel, and automatically superimposes a risk level label and a color identifier. At the same time, the interface dynamically presents the key input features most relevant to the prediction, including the basic disease condition, physical examination values and laboratory index overview, to assist doctors in tracing the source of the model's determination. The interface also supports manual threshold adjustment function, and the system immediately re-evaluates the current individual risk level after the user modifies the threshold each time, updates the display result, ensures that the doctor can make a flexible judgment according to the individual differences of the patient in actual use, and realizes the deep integration of intelligent prediction result and clinical experience.

[0139] Corresponding to the above method, the embodiment also provides an anti-interferon gamma autoantibody positive prediction system based on machine learning, comprising:

[0140] A clinical data acquisition and preprocessing unit is configured to acquire basic disease information, physical examination information and laboratory indexes obtained through routine hematology, biochemistry and immunology detection of a to-be-tested object, form an original data sample, and perform missing value filling, outlier removal and data standardization on the original data sample to obtain a training sample set.

[0141] A feature encoding and screening unit is configured to encode the training sample set into a numerical feature vector, and screen target features related to AIGAs positivity by using a regularization algorithm.

[0142] A multi-model training unit is configured to train at least two different types of supervised machine learning models based on the target features and corresponding positive / negative labels.

[0143] A cross-validation and ensemble learning unit is configured to perform cross-validation on the supervised machine learning models, select a candidate model set according to a preset performance threshold, and implement ensemble learning on the candidate model set to output the AIGAs positivity probability of the to-be-tested object.

[0144] A classification determination and visualization display unit is configured to compare the positivity probability with a preset custom threshold, generate a classification result and perform visualization display to assist clinical decision-making.

[0145] The beneficial effects of the present application are as follows:

[0146] The present invention provides a machine learning-based method for predicting the positive risk of anti-γ interferon autoantibodies (AIGAs), which significantly reduces the reliance on expensive laboratory tests. It uses easily accessible underlying diseases, physical examinations, and routine laboratory indicators to achieve intelligent prediction of the risk of AIGAs positivity. It has good scalability and deployability, and is particularly suitable for large-scale primary screening in resource-constrained primary medical institutions.

[0147] This invention innovatively integrates the structured compression idea of ​​Group-LASSO through the feature selection mechanism of "group modeling + regularized screening", which not only significantly reduces the model complexity and overfitting risk, but also improves the interpretability of the model output, making the final retained features highly consistent with the clinical knowledge system, and enhancing doctors' trust and acceptance of the model.

[0148] During the model training and prediction process, the present invention adopts an integrated learning strategy to fuse multiple supervised machine learning models, and integrates the prediction results of different models through adaptive weight combination, which further improves the accuracy, stability and generalization ability of the prediction, and adapts to usage scenarios in different data centers or population backgrounds.

[0149] The present invention also constructs an interactive visual analysis interface, allowing clinical users to dynamically set prediction thresholds based on actual conditions and view the model's risk level and key feature sources, thereby enhancing the model's practicality and flexibility in clinical decision-making assistance and providing technical support for achieving accurate screening and personalized medicine.

[0150] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0151] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for predicting the positive anti-interferon-γ autoantibody based on machine learning, characterized in that: include: Collect basic disease information, physical examination information, and laboratory indicators obtained through routine hematology, biochemistry, and immunology tests of the subjects to be tested to form an original data sample, and perform missing value filling, outlier removal, and data standardization on the original data sample to obtain a training sample set; Encoding the training sample set into a numerical feature vector, and using a regularization algorithm to screen target features associated with AIGAs positivity; Based on the target features and the corresponding positive / negative labels, respectively, training at least two different types of supervised machine learning models; Performing cross-validation on the supervised machine learning model, selecting a set of candidate models based on a preset performance threshold, and performing ensemble learning on the set of candidate models to output the probability of AIGAs positivity of the subject to be tested; The positive probability is compared with a preset custom threshold to generate a classification result and visualize it to assist clinical decision-making.

2. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 1, characterized in that: The basic disease information, physical examination information and laboratory indicators obtained through routine hematology, biochemistry and immunology tests of the subjects to be tested are collected to form raw data samples. The raw data samples are then filled with missing values, outliers are removed and data normalized to obtain a training sample set, including: The presence or absence of chronic or infectious diseases is retrieved from the electronic medical record system and coded into a binary variable to obtain the underlying disease information; the chronic or infectious diseases include: hepatitis B, chronic kidney disease, diabetes, tuberculosis and HIV infection; The medical staff measures and records body temperature, systolic and diastolic blood pressure, heart rate, respiratory rate, and body mass index (BMI) to obtain the physical examination information; Import quantitative results from the test information system to obtain the laboratory indicators; the quantitative results include: white blood cell count, lymphocyte count, CD4 + T-cell count, immunoglobulin G, immunoglobulin E, serum globulins, C-reactive protein, and erythrocyte sedimentation rate; If the missing rate of any of the basic disease information, physical examination information, and laboratory indicators exceeds 0.33, the indicator will be deleted; Perform the Shapiro-Wilk normality test on the retained indicators. If the test is passed, the missing values ​​are filled with the sample mean, otherwise they are filled with the sample median; For each numerical indicator, calculate the first quartile Q1 and the third quartile Q3, and obtain the interquartile range IQR = Q3 − Q1; Delete sample values ​​smaller than Q1−1.5IQR or larger than Q3+1.5IQR; The remaining numerical indicators are subjected to Z-score processing to make the mean value 0 and the standard deviation 1, so as to obtain the training sample set.

3. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 1, characterized in that: The training sample set is encoded into a numerical feature vector, and a regularization algorithm is used to screen target features associated with positive AIGAs, including: For each chronic or infectious disease obtained in the training sample set, a one-hot vector is created according to the ICD-10 three-digit code, and the one-hot vector is aggregated according to the system to which the disease belongs to generate a four-dimensional hierarchical disease vector; the disease belongs to the respiratory, digestive, endocrine and immune systems; performing Z-score normalization on the physical examination information in the training sample set to obtain a physical examination vector; For each continuous indicator of the laboratory index in the training sample set, a Box-Cox transformation is first performed to weaken the skewness, and then a Z-score standardization is performed to obtain a laboratory index vector; Concatenate the disease vector, the physical examination vector, and the laboratory index vector in order to form a unified numerical feature matrix; Using AIGAs positive / negative labels as dependent variables, elastic net logistic regression was performed on the unified numerical feature matrix. 5-fold cross validation was used to determine the regularization coefficient, and features with absolute values ​​of regression coefficients greater than zero were retained. The retained features are grouped according to three data modalities, and a Group-LASSO model with group sparsity constraints is applied to delete features whose coefficients within the group are shrunk to zero to obtain the target features; the data modalities include: disease, physique, and laboratory.

4. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 3, characterized in that: The retained features are grouped according to the three data modalities. The Group-LASSO model with group sparsity constraints is applied to delete the features whose coefficients in the group are shrunk to zero, and the target features are obtained, including: The retained features are divided into basic disease group, physical examination group and laboratory index group, respectively recorded as index sets and ; Establishing the objective function ;in, is the number of samples in the training sample set; is the sample label vector, positive is 1 and negative is 0; is the normalized feature matrix of the permutation; is the regression coefficient vector to be determined; for In the kth group of feature index sets subvectors on ; is the regularization coefficient; is the number of features in the kth group; The square root-iterative threshold algorithm is used to Perform group sparse shrinkage update until the relative change rate of the objective function is less than ; If a group of If the norm is zero, all features corresponding to the group with zero norm are deleted; The features containing the remaining non-zero coefficients are merged to obtain the target feature.

5. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 1, characterized in that: Based on the target features and corresponding positive / negative labels, at least two different types of supervised machine learning models are trained, including: a) using stratified sampling to divide the target feature matrix and labels into a training subset and a validation subset in a ratio of 8:2; b) Use the training subset to build a gradient boosting decision tree model. After each iteration, the area under the receiver operating characteristic curve (AUC) of the validation subset is used as the monitoring indicator. If there is no improvement in the AUC for ten consecutive rounds, the iteration is stopped. c) A linear support vector machine model was constructed using the training subset, and the penalty coefficient C was automatically searched through five-fold cross-validation, with the C value that maximized the AUC of the validation subset being used as the final parameter.

6. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 5, characterized in that: Performing cross-validation on the supervised machine learning model, selecting a set of candidate models based on a preset performance threshold, and performing ensemble learning on the set of candidate models to output the AIGAs-positive probability of the subject to be tested, including: A k-fold stratified cross-validation is performed on each of the supervised machine learning models to calculate the evaluation results of each fold; the evaluation results include: area under the curve (AUC), sensitivity, and specificity; The arithmetic mean of the evaluation results of each fold is taken to obtain the overall performance vector of the model (AUC i , SEN i , SPE i ); Candidate model screening: If AUC i ≥0.90 and SEN i If ≥0.85, the i-th model is included in the candidate model set, otherwise it is discarded; Perform Pratt scaling on the validation fold output for each model in the candidate model set, mapping the raw confidence to a calibrated probability. ; Based on AUC i The result is to set the weight in a normalized way. ; Where m is the number of candidate models; For a single object to be tested, the calibration probability of each candidate model output is Perform weighted summation to obtain the final positive probability .

7. The method for predicting positive anti-interferon-γ autoantibodies based on machine learning according to claim 5, characterized in that: The positive probability is compared with a preset custom threshold to generate a classification result and visualize it to assist clinical decision-making, including: A threshold input interface is provided on the user terminal, and the user-defined threshold T is stored in the system parameter cache; if the user does not input, the default threshold T0=0.5 is read; The positive probability P for a single test object final Compare with the threshold T, if P final ≥T, marked as "positive risk", otherwise marked as "negative risk"; The probability interval [0,1] is divided into three segments: [0,T) is low risk, [T,0.75) is medium risk, and [0.75,1] is high risk, and each segment is assigned a green, yellow, or red visual identifier. Call the visualization module on the user terminal and final Draw in the form of a bar progress bar or a circular dashboard, with "positive risk" or "negative risk" and the visual identifier superimposed; The visual interface simultaneously displays a numerical summary of key features corresponding to underlying diseases, physical examination and laboratory indicators, and allows users to adjust the threshold T to trigger real-time re-judgment, allowing clinicians to optimize decisions based on the patient's specific situation.

8. A machine learning-based anti-interferon-γ autoantibody positive prediction system, characterized in that: include: The clinical data collection and preprocessing unit is used to collect the basic disease information, physical examination information and laboratory indicators obtained through routine hematology, biochemistry and immunology tests of the subjects to be tested to form raw data samples, and perform missing value filling, outlier removal and data standardization on the raw data samples to obtain a training sample set; a feature encoding and screening unit, configured to encode the training sample set into a numerical feature vector and screen target features associated with positive AIGAs using a regularization algorithm; a multi-model training unit, configured to train at least two different types of supervised machine learning models based on the target features and corresponding positive / negative labels; a cross-validation and ensemble learning unit, configured to perform cross-validation on the supervised machine learning model, select a set of candidate models based on a preset performance threshold, perform ensemble learning on the set of candidate models, and output a positive probability of AIGAs for the subject to be tested; The classification determination and visualization display unit is used to compare the positive probability with a preset custom threshold, generate a classification result and display it visually to assist clinical decision-making.

Images