Method for establishing pSS prediction model based on multiple clinical examination indexes and pSS prediction system

Abstract

According to the pSS prediction model establishment method based on the multiple clinical examination indexes and the pSS prediction system, multiple machine learning methods and feature screening algorithms are integrated, a new data processing and analysis framework is provided, prediction model establishment of pSS patients is achieved, and the prediction efficiency of the pSS patients is improved. And the contribution degree of important features to model decision is revealed.

Description

Technical Field

[0001] The present invention relates to the field of medical health data analysis, and in particular to the establishment of a prediction model for primary Sjögren's syndrome (pSS), involving feature screening of clinical test data and the application of machine learning methods, and specifically is a method for establishing a pSS prediction model based on multiple clinical test indicators and a pSS prediction system. Background Art

[0002] Primary Sjögren's syndrome (pSS) is a common autoimmune disease. Traditional prediction models usually rely on patients' symptoms and laboratory test results. However, these methods have limited accuracy and sensitivity. In recent years, with the rapid development of data science and machine learning technology, the application of machine learning methods to clinical data analysis can achieve the establishment of early warning models for diseases. In the existing technology, although some studies have explored the application of various machine learning methods in clinical data analysis, due to the high-dimensional characteristics and complexity of clinical data, there is still a lack of effective feature screening and optimization methods for multi-model integration. Summary of the invention

[0003] The purpose of the present invention is to provide a method for establishing a pSS prediction model based on multiple clinical test indicators and a pSS prediction system. By integrating multiple machine learning methods and feature screening algorithms, the present invention proposes a new data processing and analysis framework, realizes the establishment of a prediction model for pSS patients, and reveals the contribution of important features to model decision-making.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for establishing a pSS prediction model based on multiple clinical test indicators, the method comprising the following steps:

[0006] Step (1) Data collection: Collect five test data of all research subjects, including blood routine, urine routine, complete biochemistry, coagulation, and blood lymphocyte subset counts;

[0007] Step (2) Data preprocessing: preprocess the collected data, including deleting rows with more than 80% null values ​​and filling the null values ​​with the average value calculated from the column where the null values ​​are located;

[0008] Step (3) Data standardization: standardizing the clinical test data set of the filtered samples;

[0009] Step (4) Dataset division: divide the standardized dataset into training set and test set in a ratio of 3:1;

[0010] Step (5) Feature importance evaluation and screening: Use RandomForestClassifier, CV2, F-score, XGB, Lasso, and LGBMClassifier to evaluate feature importance, calculate six feature importance values, and screen out 16 important features;

[0011] Step (6) Machine learning model construction: Seven machine learning models are constructed based on the screened features, including RF, SVM, LR, GN, LightGBM, XGBOOST, and KNN models;

[0012] Step (7) Model optimization and evaluation: The model is optimized by five-fold cross validation and hyperparameter adjustment, and the model performance is evaluated using five indicators: ACC, Recall, F1, AUC, and precision based on the five-fold cross validation strategy;

[0013] Step (8) Model interpretation: The SHAP interpretation method is used to calculate the SHAP value of each feature, reflecting the contribution of the feature to the model's prediction ability.

[0014] In the step (2), the average value is used to fill the empty value in the data preprocessing; the feature screening algorithms used in the step (5) feature importance evaluation and screening are RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM.

[0015] Step (6) The models used in the machine learning model construction include RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

[0016] The model optimization and evaluation in step (7) are achieved through five-fold cross validation and grid search parameter adjustment; the model optimization and evaluation in step (7) are based on five indicators, namely ACC, Recall, F1, AUC and Precision.

[0017] In the step (8), the SHAP method is used in the model interpretation to calculate the SHAP value of each feature.

[0018] The present invention also provides a pSS prediction system based on multiple clinical test indicators, the system comprising:

[0019] (1) Data collection module, used to collect clinical test data;

[0020] (2) Data processing module, used for data preprocessing and standardization;

[0021] (3) Feature selection module, which is used to evaluate feature importance and select important features using multiple algorithms;

[0022] (4) a model training module, used to train multiple machine learning models based on the screened features;

[0023] (5) Model optimization module, used to optimize the model through cross-validation and hyperparameter adjustment;

[0024] (6) Model evaluation module, used to evaluate the prediction performance of the model;

[0025] (7) Model explanation module, used to explain the impact of features on model prediction results.

[0026] The data processing module includes a data preprocessing unit and a data standardization unit, and the feature selection module includes a feature importance evaluation unit and a feature screening unit.

[0027] The present invention also provides a pSS prediction method based on the pSS prediction system based on multiple clinical test indicators, comprising the following steps:

[0028] Step (1) collecting clinical test data through a data collection module;

[0029] Step (2) preprocessing and standardizing the data using a data processing module;

[0030] Step (3) evaluating feature importance and selecting important features through a feature selection module;

[0031] Step (4) using a model training module to train multiple machine learning models based on the selected features;

[0032] Step (5) optimizing the model through a model optimization module;

[0033] Step (6) using a model evaluation module to evaluate the prediction performance of the model;

[0034] Step (7) explains the impact of features on the model prediction results through the model explanation module.

[0035] In the step (3), RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM algorithms are used to evaluate feature importance.

[0036] The machine learning models trained in step (4) are RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention provides a method for establishing a pSS prediction model based on multiple clinical test indicators and a pSS prediction system. Through the integrated optimization of multiple algorithms, accurate prediction of pSS patients is achieved, and the contribution of important features to model decision-making is revealed. The system is particularly suitable for an early prediction model of primary Sjögren's syndrome.

[0039] The results of the present invention show that basic clinical test data based on five results of blood routine, urine routine, complete biochemistry, coagulation, and blood lymphocyte subset counts can achieve effective prediction of pSS patients. The pSS prediction model constructed by using 16 features obtained by the LASSO strategy as the input features of LightGBM achieved excellent performance, with an accuracy of 0.88, an AUC of 0.93, and a precision of 0.89. In addition, the model interpretation was used to explore how the 16 feature indicators affect the model's decision-making. For example, the increase in Glu and UA will improve the model's prediction effect. By comparing the expression of the 16 feature indicators in patients and healthy samples, it was further confirmed that they can be used as key indicators for pSS prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a feature screening result diagram of the present invention, which shows 16 important features screened using six algorithms.

[0041] Figure 2 It is an AUC curve diagram of the machine learning model of the present invention, showing the prediction performance of different models.

[0042] Figure 3 It is the SHAP value diagram of the present invention, which shows the contribution of each feature to the model decision. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0044] Embodiment 1

[0045] A method for establishing a pSS prediction model based on multiple clinical test indicators comprises the following steps:

[0046] 1. Data preprocessing: Collect five test data of all research subjects, delete rows with more than 80% null values ​​and fill the null values ​​with the average value of the column.

[0047] 2. Feature screening: Six algorithms, including RandomForestClassifier, CV2, F-score, XGB, Lasso, and LGBMClassifier, were used to evaluate the importance of features and 16 important features were screened out.

[0048] 3. Model construction: Based on the screened features, seven machine learning models including RF, SVM, LR, GNB, LightGBM, XGBOOST, and KNN are constructed.

[0049] 4. Model optimization and evaluation: The model was optimized through five-fold cross validation and adjustment of hyperparameters, and the model was evaluated using indicators such as ACC, Recall, F1, AUC, and precision.

[0050] 5. Model explanation: Use SHAP value to calculate the contribution of each feature to the model’s predictive ability and explain the model’s decision-making process.

[0051] Embodiment 2

[0052] A method for establishing a pSS prediction model based on multiple clinical test indicators, the method comprising the following steps:

[0053] Step (1) Data collection: Collect five test data including blood routine, urine routine, complete biochemistry, coagulation, and blood lymphocyte subset counts of all research subjects (170,107 samples, a total of 93 indicators);

[0054] Step (2) Data preprocessing: preprocess the collected data, including deleting rows with more than 80% null values ​​and filling the null values ​​with the average value calculated from the column where the null values ​​are located;

[0055] Step (3) Data standardization: standardizing the clinical test data set of the filtered samples;

[0056] Step (4) Dataset division: divide the standardized dataset into training set and test set in a ratio of 3:1;

[0057] Step (5) Feature importance evaluation and screening: Use RandomForestClassifier, CV2, F-score, XGB, Lasso, and LGBMClassifier to evaluate feature importance, calculate six feature importance values, and screen out 16 important features;

[0058] Step (6) Machine learning model construction: Seven machine learning models are constructed based on the screened features, including random forest RF, support vector machine SVM, logistic regression LR, Gaussian naive Bayes GN, LightGBM, XGBOOST, and K nearest neighbor KNN model;

[0059] Step (7) Model optimization and evaluation: The model is optimized by five-fold cross validation and hyperparameter adjustment, and the model performance is evaluated using five indicators: ACC, Recall, F1, AUC, and precision based on the five-fold cross validation strategy;

[0060] Step (8) Model interpretation: The SHAP interpretation method is used to calculate the SHAP value of each feature, reflecting the contribution of the feature to the model's prediction ability.

[0061] In the step (2), the average value is used to fill the empty value in the data preprocessing; the feature screening algorithms used in the step (5) feature importance evaluation and screening are RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM.

[0062] Step (6) The models used in the machine learning model construction include RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

[0063] The model optimization and evaluation in step (7) are achieved through five-fold cross validation and grid search parameter adjustment; the model optimization and evaluation in step (7) are based on five indicators, namely, accuracy ACC, recall Recall, F1 score F1, area under the receiver operating characteristic curve AUC and precision Precision.

[0064] In the step (8), the SHAP method is used in the model interpretation to calculate the SHAP value of each feature.

[0065] Embodiment 3

[0066] A pSS prediction system based on multiple clinical test indicators, the system comprising:

[0067] (1) Data collection module, used to collect clinical test data;

[0068] (2) Data processing module, used for data preprocessing and standardization;

[0069] (3) Feature selection module, which is used to evaluate feature importance and select important features using multiple algorithms;

[0070] (4) a model training module, used to train multiple machine learning models based on the screened features;

[0071] (5) Model optimization module, used to optimize the model through cross-validation and hyperparameter adjustment;

[0072] (6) Model evaluation module, used to evaluate the prediction performance of the model;

[0073] (7) Model explanation module, used to explain the impact of features on model prediction results.

[0074] The data processing module includes a data preprocessing unit and a data standardization unit, and the feature selection module includes a feature importance evaluation unit and a feature screening unit.

[0075] Embodiment 4

[0076] A pSS prediction method based on the pSS prediction system based on multiple clinical test indicators comprises the following steps:

[0077] Step (1) collecting clinical test data through a data collection module;

[0078] Step (2) preprocessing and standardizing the data using a data processing module;

[0079] Step (3) evaluating feature importance and selecting important features through a feature selection module;

[0080] Step (4) using a model training module to train multiple machine learning models based on the selected features;

[0081] Step (5) optimizing the model through a model optimization module;

[0082] Step (6) using a model evaluation module to evaluate the prediction performance of the model;

[0083] Step (7) explains the impact of features on the model prediction results through the model explanation module.

[0084] In the step (3), RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM algorithms are used to evaluate feature importance.

[0085] The machine learning models trained in step (4) are RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

[0086] The present invention is based on multiple clinical test indicators and combines multiple feature screening algorithms with machine learning methods to establish a pSS early warning model to achieve accurate prediction of pSS.

[0087] Research methods: First, five test data including blood routine, urine routine, biochemistry, coagulation, and blood lymphocyte subset counts were collected from all research subjects (170,107 samples, 93 indicators in total). After that, the data was preprocessed (deleting rows with more than 80% null values ​​and filling null values ​​with the average value calculated from the column where the null value is located), and the clinical test data set of the filtered samples (a total of 94,565 samples) was standardized, and then all samples were divided into training and test sets in a ratio of 3:1. The calculation method and interpretation of high-dimensional data are crucial to understanding model decisions.

[0088] In the present invention, six algorithms, namely RandomForestClassifier, CV2, F-score, XGB, Lasso and LGBMClassifier, are used to evaluate the importance of features, six feature importance values ​​are calculated, and 16 important features are screened out for subsequent analysis.

[0089] Afterwards, the present invention constructs seven machine learning models of RF, SVM, LR, GNB, LightGBM, XGBOOST, and KNN. When using the seven machine learning algorithms for prediction, the feature scores of RandomForestClassifier, CV2, F-score, XGB, Lasso, and LGBMClassifier are normalized and then the arithmetic mean is calculated. The selected features are used to train a variety of combined machine learning algorithms, such as 42 combination models such as LASSO+RF, LASSO+KNN, and the corresponding exact feature combination (LASSO+LightGBM) is calculated. Model optimization is achieved by five-fold cross validation and adjustment of hyperparameters. The evaluation of model performance is based on a five-fold cross validation strategy, and is achieved using five indicators: ACC, Recall, F1, ACU, and precision. Finally, the SHAP value of each feature pair is calculated using the SHAP interpretation method to reflect the contribution of the feature to the prediction ability of the entire model.

[0090] 1. Data Preprocessing

[0091] First, we collected data from 170,107 samples of blood routine, urine routine, biochemistry, coagulation, and blood lymphocyte subset counts. We deleted rows with more than 80% null values ​​and filled the null values ​​with the average value of each column. We standardized the data and then divided the data set into a training set and a test set in a ratio of 3:1.

[0092] 2. Feature Screening

[0093] The following six algorithms are used to evaluate feature importance:

[0094] 1. RandomForestClassifier: Uses the random forest algorithm to evaluate the importance of features and calculates the Gini importance of features.

[0095] 2.CV2: Cross-validation evaluates the performance of the model and uses the average cross-validation score to evaluate the importance of features.

[0096] 3. F-score: Use the F-score statistic to evaluate the significance of the feature.

[0097] 4.XGB(XGBoost): Use XGBoost algorithm to calculate feature importance.

[0098] 5.Lasso: Use Lasso regression to filter out features with sparsity constraints.

[0099] 6.LGBMClassifier: Uses the LightGBM algorithm to evaluate the importance of features.

[0100] The 16 important features screened out were used for subsequent analysis.

[0101] 3. Model Construction and Optimization

[0102] Build models using the following seven machine learning algorithms:

[0103] RF (Random Forest): Random forest algorithm uses an integration of multiple decision trees to improve prediction performance.

[0104] SVM (Support Vector Machine): Support vector machine, looking for the hyperplane that maximizes the margin.

[0105] LR (Logistic Regression): Logistic regression, suitable for binary classification problems.

[0106] GNB (Gaussian Naive Bayes): Gaussian Naive Bayes, assuming that features are independent of each other.

[0107] LightGBM: A fast implementation of gradient boosted decision trees suitable for large-scale datasets.

[0108] XGBOOST: A boosted tree algorithm that improves model performance by optimizing the loss function.

[0109] KNN (K-Nearest Neighbors): K nearest neighbor algorithm, classifies unknown samples according to the categories of neighbors.

[0110] For each model, the arithmetic mean is calculated using the normalized feature score results, and various combinations of machine learning models are trained, such as LASSO+RF, LASSO+KNN, and other 42 combination models. The best combination is selected for further optimization. Model optimization is achieved through five-fold cross validation and grid search parameter adjustment.

[0111] IV. Model Evaluation

[0112] In the present invention, in order to comprehensively evaluate the predictive performance of the constructed model, we use a variety of evaluation indicators. These indicators can help us understand the performance of the model in different aspects, such as accuracy, sensitivity, precision and the distinguishing ability of the model. The following is a detailed description of the evaluation indicators used, the calculation formula and its application method.

[0113] 1.ACC (Accuracy)

[0114] Definition: Accuracy is an indicator to measure the overall prediction accuracy of the model, that is, the proportion of samples correctly predicted by the model to the total number of samples. It reflects the overall prediction effect of the model.

[0115] Calculation formula:

[0116]

[0117] in:

[0118] TP (True Positive): The number of samples that are actually positive and predicted to be positive

[0119] TNTNTN (True Negative): The number of samples that are actually negative and predicted to be negative

[0120] FP (False Positive): The number of samples that are actually negative but predicted to be positive

[0121] FN (False Negative): The number of samples that are actually positive but predicted to be negative

[0122] Application method: After the model prediction is completed, the number of correctly predicted samples among all predicted samples is calculated and divided by the total number of samples to obtain the accuracy. A higher accuracy rate indicates that the overall prediction of the model is more accurate, but in the case of class imbalance, the interpretation of the accuracy rate may be biased.

[0123] 2.Recall (Recall rate, sensitivity)

[0124] Definition: Recall is an indicator that measures the model's ability to predict actual positive samples, indicating the proportion of actual positive samples that are correctly predicted as positive. A high recall means that most of the actual positive samples are correctly identified by the model.

[0125] Calculation formula:

[0126]

[0127] Application method: In the prediction results, calculate the ratio of the number of samples correctly predicted as positive (TP) to the number of actual positive samples (TP+FN). High recall is particularly important for application scenarios that need to avoid missed detection (such as disease detection).

[0128] 3.F1 Score

[0129] Definition: The F1 score is the harmonic mean of precision and recall, which is used to comprehensively consider the precision and recall of the model. It is particularly suitable for measuring the performance of the model in the case of class imbalance because it focuses on both false positives and missed positives.

[0130] Calculation formula:

[0131]

[0132] in:

[0133]

[0134] Application method: The F1 score is calculated by calculating the model's precision and recall. The F1 score is a balance between precision and recall. The higher the value, the better the overall performance of the model, especially in certain class-skewed datasets.

[0135] 4.AUC (Area Under Curve)

[0136] Definition: AUC is the abbreviation of the area under the ROC curve (receiver operating characteristic curve), which is used to measure the model's ability to distinguish between positive and negative samples. The closer the AUC value is to 1, the stronger the model's ability to distinguish is.

[0137] Calculation method:

[0138] First, draw the ROC curve, with the vertical axis as True Positive Rate (TPR) and the horizontal axis as False Positive Rate (FPR).

[0139] The calculation formulas for TPR (True Positive Rate) and FPR (False Positive Rate) are:

[0140]

[0141] Calculate the area under the curve, which is the AUC value.

[0142] Application method: By gradually changing the prediction probability threshold from 0 to 1, draw the ROC curve of TPR to FPR. The higher the AUC value, the better the model performs in distinguishing positive and negative samples. AUC is an overall evaluation of the performance of the model at different thresholds.

[0143] 5. Precision

[0144] Definition: Precision (also called accuracy) refers to the proportion of samples predicted to be positive that are actually positive. High precision means that the model has high confidence in the positive prediction.

[0145] Calculation formula:

[0146]

[0147] Application method: In the prediction results, calculate the ratio of the number of samples correctly predicted as positive (TP) to the number of samples predicted as positive (TP+FP). High accuracy means that the model makes fewer mistakes when predicting positive, which is particularly suitable for use in scenarios where misjudgment (false detection) needs to be avoided.

[0148] Summary of Model Evaluation Methods

[0149] The present invention adopts the above five evaluation indicators to comprehensively analyze the predictive ability of different machine learning models and their combinations. In practical applications, the performance of each model on the training set and the test set is evaluated through a 5-fold cross-validation strategy, and the average value of each evaluation indicator is calculated to ensure the robustness and generalization ability of the model.

[0150] By calculating and analyzing these indicators, we can determine the best feature combination and model parameters, and finally select the optimal model for the prediction of primary Sjögren's syndrome (pSS) patients. In addition, the SHAP value method is used to interpret the feature contribution in the model prediction results, which further verifies the scientificity and reliability of the selected model.

[0151] 5. Model Interpretation

[0152] In this paper, in order to better understand the decision-making mechanism of the model and the impact of each feature on the prediction results, we use the SHAP (SHapley Additive exPlanations) value method. This method is based on cooperative game theory and explains the prediction behavior of the model by calculating the marginal contribution of each feature to the model prediction, providing a global and local explanation of the importance of the feature and its impact direction.

[0153] SHAP value is a method to explain the prediction results of machine learning models. Its core idea comes from Shapley value in game theory. Shapley value is used to measure the contribution of each participant to the total benefit in a cooperative game. In machine learning models, features are regarded as participants, and the output of the model is regarded as the total benefit. SHAP value measures the contribution of each feature to the model prediction.

[0154] Global interpretation: understanding the importance of a feature across the entire dataset.

[0155] Local explanations: Understanding the contribution of a feature in a single prediction instance.

[0156] SHAP value calculation method:

[0157] The calculation of the SHAP value is based on the average of the marginal contributions of all possible feature combinations, and the formula is as follows:

[0158]

[0159] in:

[0160] φ i (f, x) is the SHAP value of feature i, which indicates the contribution of feature i to the model prediction results.

[0161] S is a subset of the feature set and does not contain feature i.

[0162] N is the set of all features.

[0163] f(S) is the model output for making predictions using only the feature set S.

[0164] |S| represents the number of elements in set S, and |N| represents the total number of features.

[0165] f(S|U{i})-f(S) represents the marginal contribution of feature i to the model prediction results based on the feature set S.

[0166] The main steps to calculate the SHAP value are as follows:

[0167] 1. Model training: First, train a machine learning model based on all features.

[0168] 2. Combination calculation: Calculate the changes in the model prediction output under all feature combinations.

[0169] 3. SHAP value summation: Based on all possible feature combinations, calculate the marginal contribution of the features to the model output and finally take the average value.

[0170] How to use SHAP value:

[0171] 1. Feature contribution analysis: The SHAP value can be used to understand the average contribution of each feature in the prediction model. A larger SHAP value (whether positive or negative) indicates that the feature has a greater impact on the model prediction results.

[0172] 2. Feature influence direction: The sign of the SHAP value indicates the direction in which the feature affects the prediction result. A positive SHAP value means that the increase of the feature will cause the prediction result to change in the positive direction, and vice versa.

[0173] 3. Model interpretation visualization: Using SHAP values, a series of interpretable visualization charts can be generated, including:

[0174] SHAP Summary Plot: Shows a global overview of feature importance and influence direction, with each point representing a data sample.

[0175] SHAP Dependence Plot: shows how the SHAP value of a single feature changes with its feature value, reflecting the relationship between the feature and the prediction result.

[0176] Application of SHAP value in the present invention:

[0177] In the present invention, the SHAP value method is used to interpret and analyze the pSS prediction model, which is mainly reflected in the following aspects:

[0178] 1. Feature Importance Ranking: By calculating the SHAP value of each feature, the importance of all features in the model is ranked. For example, features such as Glu (glucose), GLO (glutathione oxidoreductase), and Hb (hemoglobin) have large SHAP values, indicating that these features contribute significantly to the model prediction results.

[0179] 2. Feature impact analysis:

[0180] Effect of Glu: The analysis showed that an increase in Glu significantly increased the probability of the model predicting pSS, which was related to the abnormalities of glucose metabolism in pSS patients.

[0181] Effect of GLO: Changes in GLO levels are related to immune response, and elevated GLO plays a certain auxiliary diagnostic role in the prediction of pSS models.

[0182] Effect of Hb: Changes in Hb levels may be related to the anemia symptoms of pSS patients. The model prediction results show that decreased Hb is correlated with a high risk of pSS.

[0183] 3. Explanation of individual predictions: For each specific patient sample, calculate the SHAP value corresponding to its features to explain why the model predicts that the sample is pSS or healthy. For example, in the prediction of some pSS patients, it was found that the SHAP values ​​of Glu and GLO were positive and large, while the SHAP value of Hb was negative. Combining this information can explain the prediction results of the model.

[0184] The SHAP value method provides a clear explanation of the model decision mechanism by calculating the marginal contribution of features to model prediction. In the present invention, SHAP value analysis revealed how important features such as Glu, GLO and Hb affect the decision of the pSS prediction model. These findings not only help improve the transparency and interpretability of the model, but also provide clinicians with additional diagnostic clues and potential pathophysiological insights.

[0185] 6. Algorithm formula

[0186] 1 Feature Importance Formula

[0187] Feature importance is a measure of the contribution of each feature to the model's predictions. It helps identify which features are most useful in the model and helps optimize and interpret the model. In this paper, we used a variety of algorithms to evaluate feature importance. The following are the methods and formulas for calculating feature importance for each algorithm.

[0188] 1. Random Forest Classifier Feature Importance

[0189] Definition: The feature importance of random forest is based on how well the feature splits the dataset in multiple decision trees. The importance of a feature is determined by the sum of the information gain it brings when splitting nodes in all trees.

[0190] Calculation formula:

[0191]

[0192] in:

[0193] I j is the importance of feature j.

[0194] T is the set of all trees.

[0195] t represents a node in a single tree.

[0196] n(t) is the number of samples in node t.

[0197] n is the total number of samples.

[0198] ΔGini(t, j) is the change in the Gini index of node t under the condition of feature j.

[0199] Method: Calculate the weighted sum of the changes in the Gini index of features at all nodes in each tree, and then average the results across all trees in the entire random forest.

[0201] 2. Cross-validation score (CV2)

[0202] Definition: Cross-validation scoring determines the importance of a feature by evaluating its impact on model performance on different training and validation set splits.

[0203] Calculation formula:

[0204]

[0205] in:

[0206] I j is the importance of feature j.

[0207] K is the number of cross-validation folds.

[0208] n k is the number of samples in the kth fold.

[0209] Score k (f i , x j ) is the prediction score of sample i using feature j in the kth fold.

[0210] Method: During the cross-validation process, the model is trained and the contribution of features to the performance of the validation set model is calculated, and the feature importance is averaged.

[0211] 3.F-score

[0212] Definition: F-score is used to measure the significance of each feature. It calculates the mean difference and variance ratio of positive and negative samples on the feature.

[0213] Calculation formula:

[0214]

[0215] in:

[0216] F j is the F-score of feature j.

[0217] and are the means of feature j in positive and negative samples, respectively.

[0218] and are the variances of feature j in positive and negative samples, respectively.

[0219] Method: Calculate the mean and variance of each feature on positive and negative samples to obtain the F-score. The larger the value, the more significant the feature.

[0220] 4. XGBoost Feature Importance

[0221] Definition: XGBoost's feature importance is based on the frequency of feature usage and information gain at each split node. It measures the contribution of a feature in a decision tree model.

[0222] Calculation formula:

[0223]

[0224] in:

[0225] I j is the importance of feature j.

[0226] T is the set of all nodes.

[0227] Gain(t, j) is the information gain of node t conditioned on feature j.

[0228] Method: Calculate the split contribution of each feature in all trees and take the average to get the feature importance.

[0229] 5. Lasso Regression Feature Importance

[0230] Definition: Lasso regression achieves feature selection by adding an L1 regularization term to the regression model. The importance of a feature is measured by the size of its corresponding regression coefficient.

[0231] Calculation formula:

[0232]

[0233] in:

[0234] w is the regression coefficient vector.

[0235] x i is the eigenvector of sample i.

[0236] y i is the actual value of sample i.

[0237] λ is the regularization parameter.

[0238] ||w||1 is the L1 norm of the coefficient vector.

[0239] Method: Introduce L1 regularization in the regression model to shrink unimportant feature coefficients to zero, thereby screening out important features. The importance of non-zero coefficient features is measured based on their absolute value.

[0240] 6. LightGBM Feature Importance

[0241] Definition: LightGBM's feature importance is based on the contribution of the feature to the objective function loss when splitting a node. It is similar to XGBoost's feature importance calculation.

[0242] Calculation formula:

[0243]

[0244] in:

[0245] I j is the importance of feature j.

[0246] T is the set of all nodes.

[0247] Split_Gain(t,j) is the split gain of node t conditioned on feature j.

[0248] Method: LightGBM calculates feature importance by the gain contribution of each feature in splitting across all trees.

[0249] By calculating the feature importance through the above algorithms, we can identify the features that contribute most to the model prediction. These methods evaluate the importance of features from different perspectives, making the model interpretation more comprehensive and accurate. In the present invention, by integrating these algorithms, 16 important features were selected for subsequent analysis and model construction, ensuring the efficiency and interpretability of the model.

[0250] Research results: The results of the present invention show that basic clinical test data based on five results of blood routine, urine routine, complete biochemistry, coagulation, and blood lymphocyte subset counts can achieve effective prediction of pSS patients. The pSS prediction model constructed by using 16 features obtained by the LASSO strategy as the input features of LightGBM achieved excellent performance, with an accuracy of 0.88, an AUC of 0.93, and a precision of 0.89. In addition, model interpretation was used to explore how the 16 feature indicators affect the model's decision-making. For example, an increase in Glu and UA will improve the model's prediction effect. By comparing the expression of the 16 feature indicators in patients and healthy samples, it was further confirmed that they can be used as key indicators for pSS prediction.

[0251] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only preferred examples of the invention and are not intended to limit the present invention. Without departing from the novel spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for establishing a pSS prediction model based on multiple clinical test indicators, characterized in that: The method comprises the following steps: Step (1) Data collection: Collect five test data of all research subjects, including blood routine, urine routine, complete biochemistry, coagulation, and blood lymphocyte subset counts; Step (2) Data preprocessing: preprocess the collected data, including deleting rows with more than 80% null values ​​and filling the null values ​​with the average value calculated from the column where the null values ​​are located; Step (3) Data standardization: standardizing the clinical test data set of the filtered samples; Step (4) Dataset division: divide the standardized dataset into training set and test set in a ratio of 3:1; Step (5) Feature importance evaluation and screening: Use RandomForestClassifier, CV2, F-score, XGB, Lasso, and LGBMClassifier to evaluate feature importance, calculate six feature importance values, and screen out 16 important features; Step (6) Machine learning model construction: Seven machine learning models are constructed based on the screened features, including random forest (RF), support vector machine (SVM), logistic regression (LR), Gaussian naive Bayes (GN), LightGBM, XGBOOST, and K nearest neighbor (KNN) models; Step (7) Model optimization and evaluation: The model was optimized by five-fold cross validation and hyperparameter adjustment, and the model performance was evaluated using five indicators: accuracy (ACC), recall (Recall), F1 score, area under the receiver operating characteristic curve (AUC), and precision (Precision) based on the five-fold cross validation strategy; Step (8) Model interpretation: The SHAP interpretation method is used to calculate the SHAP value of each feature, reflecting the contribution of the feature to the model's prediction ability.

2. The method for establishing a pSS prediction model based on multiple clinical test indicators according to claim 1, characterized in that: In the step (2), the average value is used to fill the empty value in the data preprocessing; the feature screening algorithms used in the step (5) feature importance evaluation and screening are RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM.

3. The method for establishing a pSS prediction model based on multiple clinical test indicators according to claim 1, characterized in that: Step (6) The models used in the machine learning model construction include RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

4. The method for establishing a pSS prediction model based on multiple clinical test indicators according to claim 1, characterized in that: The model optimization and evaluation in step (7) are achieved through five-fold cross validation and grid search parameter adjustment; the model optimization and evaluation in step (7) are based on five indicators, namely ACC, Recall, F1, AUC and Precision.

5. The method for establishing a pSS prediction model based on multiple clinical test indicators according to claim 1, characterized in that: In the step (8), the SHAP method is used in the model interpretation to calculate the SHAP value of each feature.

6. A pSS prediction system based on multiple clinical test indicators, characterized in that: The system includes: (1) Data collection module, used to collect clinical test data; (2) Data processing module, used for data preprocessing and standardization; (3) Feature selection module, which is used to evaluate feature importance and select important features using multiple algorithms; (4) a model training module, used to train multiple machine learning models based on the screened features; (5) Model optimization module, used to optimize the model through cross-validation and hyperparameter adjustment; (6) Model evaluation module, used to evaluate the prediction performance of the model; (7) Model explanation module, used to explain the impact of features on model prediction results.

7. The pSS prediction system based on multiple clinical test indicators according to claim 6, characterized in that: The data processing module includes a data preprocessing unit and a data standardization unit, and the feature selection module includes a feature importance evaluation unit and a feature screening unit.

8. A pSS prediction method based on the pSS prediction system based on multiple clinical test indicators according to claim 6 or 7, characterized in that: The following steps are involved: Step (1) collecting clinical test data through a data collection module; Step (2) preprocessing and standardizing the data using a data processing module; Step (3) evaluating feature importance and selecting important features through a feature selection module; Step (4) using a model training module to train multiple machine learning models based on the selected features; Step (5) optimizing the model through a model optimization module; Step (6) using a model evaluation module to evaluate the prediction performance of the model; Step (7) explains the impact of features on the model prediction results through the model explanation module.

9. The pSS prediction method according to claim 8, characterized in that: In the step (3), RandomForestClassifier, CV2, F-score, XGBoost, Lasso and LightGBM algorithms are used to evaluate feature importance.

10. The pSS prediction method according to claim 8 or 9, characterized in that: The machine learning models trained in step (4) are RF, SVM, LR, GNB, LightGBM, XGBOOST and KNN models.

Images