Method for predicting the level of glutamyl pyruvic transaminase based on machine learning and multi-source clinical data

Abstract

The present application relates to the field of medical information technology, in particular to a method for predicting the level of glutamic oxalacetic transaminase based on machine learning and multi-source clinical data, which first acquires multi-source clinical data including basic signs, blood routine and uric acid; then carries out data preprocessing and positive and negative sample classification; finally, the random forest model with the best AUC is used for training and prediction. The method overcomes the defects of traditional glutamic oxalacetic transaminase detection, which is professional, complex and time-consuming, and can quickly and simply predict the level of glutamic oxalacetic transaminase, providing an efficient auxiliary tool for liver health screening, disease risk warning and patient follow-up management, and has significant clinical practical value.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically a method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data. Background Technology

[0002] Currently, medical research and clinical practice have accumulated massive amounts of multi-source medical data, including patients' basic clinical information and rich quantitative parameters provided by various laboratory test indicators. This data contains potential information closely related to the occurrence, development, diagnosis, and prognosis of diseases. However, this data is often characterized by high dimensionality, heterogeneity, noise, and redundancy. Effectively identifying and filtering key features related to target indicators or diseases and constructing fast and concise predictive models is a significant challenge facing the field of medical information processing. Traditional medical analysis methods often rely on expert experience or limited statistical methods to select features, which may make it difficult to fully explore the complex relationships between multi-source data and may overlook some potentially important predictive factors. With the rapid development of machine learning technology, it has shown great potential in processing complex data and predicting diseases. Aspartate aminotransferase (AST) is a key biomarker for liver function assessment. Elevated AST levels typically indicate hepatocellular damage and are closely associated with various diseases, including viral hepatitis, alcoholic liver disease, and drug-induced liver injury. Currently, clinical methods for detecting AST levels are divided into wet chemistry and dry chemistry methods. Wet chemistry methods allow for the simultaneous analysis of large batches of clinical samples using a single reagent preparation. Dry chemistry methods allow for personalized combination testing of individual samples anytime and anywhere, and are available in test strips and reagent sheets. Test strip technology uses an immersion method, which cannot control the amount of reagent added, resulting in insufficient detection accuracy and failing to meet clinical testing requirements. Reagent sheet technology only has one reaction layer and cannot control the reaction sequence. Therefore, there is an urgent need in this field for a technical solution that can rapidly predict AST levels using existing, readily available clinical data. In the prior art, a relatively rapid clinical test for aspartate aminotransferase (AST) is provided by Chinese Patent Publication No. CN202411063282.X (A Dry Analysis Reagent Sheet for AST Detection, Publication Date: 2024.10.14). This paper provides a dry analysis reagent sheet for AST detection to facilitate clinical use. The reagent sheet allows for uniform diffusion, resulting in high accuracy and suitability for clinical biochemical testing. However, this detection method still requires careful operation by professionals and takes a certain amount of time. In addition, there have been some attempts to screen features from medical data to support disease analysis and prediction. For example, Chinese Patent Publication No. CN202310932812.9 (A method, detection method and application for screening key genes and tag SNPs in diabetic nephropathy complicated with diabetic retinopathy, publication date: 2023.07.27) provides a common key gene affecting the occurrence and development of diabetic nephropathy and diabetic retinopathy, a combination of 5 tag SNP sites on the key gene and its specific screening method, and establishes a machine learning method for predicting diabetic nephropathy and diabetic retinopathy, two diabetic microvascular complications, based on tag SNP information on key genes of diabetic nephropathy and diabetic retinopathy and real-world patient diagnosis and treatment information. That is, there have been some attempts in the prior art to predict diseases based on machine learning, but in terms of aspartate aminotransferase (AST) prediction, there is still a lack of a rapid prediction scheme that only uses routine clinical data and does not require additional testing. In summary, existing technologies still have some limitations in predicting AST levels. Therefore, for existing multi-source heterogeneous medical data, there is a need for a faster, simpler method that can predict AST levels without requiring additional professional intervention. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data, in order to solve the problems of complex operation, low efficiency and high cost in the existing technology for predicting AST. To achieve the above objectives, the present invention provides the following technical solution: a method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data, comprising the following steps: S1. Acquisition and Preprocessing of Multi-Source Clinical Data. Raw clinical data is acquired from hospital information systems, electronic health records, or user input. The clinical data includes, but is not limited to: a) Basic clinical information: age, height, weight, BMI, systolic blood pressure, diastolic blood pressure; b) Complete blood count indicators: white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelet count; c) Uric acid; A total of 17 indicators and a predictive indicator, AST (aspartate aminotransferase) value. In addition, the multi-source clinical data can be divided into anthropometric indicators and laboratory test indicators, which are indirectly related to aspartate aminotransferase (AST). Among them, anthropometry indicators include age, height, weight, BMI, systolic blood pressure, and diastolic blood pressure; Laboratory test indicators include white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelets, and uric acid. In other words, the prediction of the microscopic indicator of aspartate aminotransferase was achieved by combining these anthropometric indicators and laboratory test indicators. S2. Data Preprocessing. The data obtained in step S1 is preprocessed to make it suitable for subsequent model applications. This includes: S21. Missing Value Handling: Check and handle missing data. For example, for features with a missing rate below a preset threshold (e.g., 30%), use statistical imputation (e.g., use the median to imput numerical features, and the mode to imput categorical features). For features with a high missing rate, consider deletion, model prediction imputation, or interpolation. S22. Positive and negative sample classification: The normal range of ALT in clinical practice is 0-40 U / L. Using an ALT value of 40 as the dividing line, cases with an ALT value less than or equal to 40 are classified as positive samples and coded as 1, while those with an ALT value greater than 40 are classified as negative samples and coded as 0. S23. Data standardization: Scaling numerical features to eliminate the effects of differences in feature scales (e.g., using the StandardScaler method provided by sklearn to convert data into a distribution with a mean of 0 and a standard deviation of 1). S24. Feature Correlation Validation: Input all feature matrices and the target variable AST (aspartate aminotransferase), i.e., select all collected clinical indicators. Combined with AST, perform Pearson correlation analysis using IBM SPSS 27.0 (Chicago, IL, USA). Sort the results according to the absolute value of the Pearson correlation coefficient, and select the top 17 with the largest absolute values ​​to filter out the 17 indicators required in step S1. Analyze the grouped data using IBM SPSS 27.0 (Chicago, IL, USA). First, perform binary logistic regression analysis, including AST as the dependent variable and the 17 indicators required in step S1 as covariates. Obtain the predicted probability PRE_1, plot the ROC curve, and obtain the AUC value (an AUC value greater than 0.7 validates the selected indicators as having high correlation significance). S3. Model Training: The collected data is divided into training and testing sets. Based on step S2, preprocessing is performed to train a machine learning prediction model. The input of the machine learning prediction model is the feature described in step S2, and the output is the classification result of whether the aspartate aminotransferase (AST) level is normal. Specifically, for the testing set data, based on the data preprocessed in step S2, a machine learning prediction model is trained. The input of the machine learning prediction model is the 17 indicators preprocessed in step S2, and the output is the predicted probability of whether the AST level is within the normal range (the output range of the model prediction probability is 0 to 1; values ​​greater than 0.5 are considered as a prediction result of 1, and values ​​less than 0.5 are considered as a prediction result of 0). Ten commonly used machine learning models were tested, including: Logistic Regression, Random Forest, Gradient Boosting, SVM, XGBoost, LightGBM, CatBoost, and three voting models (hard voting, average soft voting, and weighted soft voting). The first seven models were used as base models, and their predictions served as input to the voting models. Hard voting followed the majority rule, selecting the model with the most predictions as its prediction. Average soft voting averaged the classification probabilities of the seven models and used this average as the voting model's prediction. Weighted soft voting assigned weights to each model based on their AUC, calculated the predicted classification probabilities, and used this weighted soft voting model's prediction. Model training included hyperparameter tuning, such as using grid search or random search combined with k-fold cross-validation to optimize performance metrics (e.g., AUC). The Random Forest model with the highest AUC was ultimately selected to predict aspartate aminotransferase (AST). Hyperparameter tuning could include the number of estimators and leaf nodes. During training, the problem of data imbalance can be addressed, for example by setting the model classification loss parameter to increase the penalty for the prediction of the class with less data. S4. Prediction of Aspartate Aminotransferase (AST) Level: After processing the multi-source medical data of a new subject to be evaluated through the same preprocessing procedure as in step S2, these feature values ​​are input into the random forest prediction model trained in step S3. The model outputs the prediction result of whether the subject's AST level is within the normal range. S5. Model Interpretation (Optional Step): To enhance the model's credibility and clinical applicability, this method may also include a step of interpreting the random forest prediction model trained based on selected features. The SHAP method is preferred. This step helps to understand the model's decision-making rationale and identify key feature drivers and their influence patterns. By calculating the SHAP value of each feature for a single prediction result, a SHAP bar chart can be generated. The SHAP bar chart can show the global importance of each feature (based on the mean absolute SHAP value). Compared with the prior art, the beneficial effects of the present invention are: This invention relates to a method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data. This method improves the relevance of feature selection; specifically, after selecting key features through a feature selection strategy, the effectiveness of the selected features is validated. It can screen specific combinations of key features highly correlated with specific medical events or diseases from multi-source medical data, including clinical and laboratory data. Compared to random simple combinations, this method is more robust and systematic, more effectively identifying features with true predictive value, helping to reduce subsequent model complexity, and potentially improving model stability on different datasets. The method for predicting AST levels based on machine learning and multi-source clinical data involved in this invention can be closely integrated with subsequent models. That is, the feature selection strategy involved (especially the feature validity verification stage) uses the performance of the subsequent prediction model as the basis for evaluating and selecting key features, ensuring that the selected key features have the best joint prediction efficacy for the final prediction task. The method for predicting AST levels based on machine learning and multi-source clinical data disclosed in this invention has high prediction accuracy. Specifically, by comparing multiple popular machine learning models, it can capture the optimal nonlinear relationship between key features and AST, and the prediction accuracy (measured by AUC) is significantly higher than that of linear models. The method for predicting AST levels based on machine learning and multi-source clinical data, as described in this invention, enhances the interpretability of potential models. The key features selected can be used to construct interpretable machine learning models, and by combining them with model interpretation techniques such as SHAP, it helps clinicians understand the prediction logic, increases their confidence in the results, and provides more valuable reference information for individualized clinical decision-making and intervention. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data disclosed in this invention has significant practical potential. The multi-source data types upon which the method relies (routine clinical information, complete blood count information, and uric acid test results) are relatively accessible in many qualified medical institutions. Furthermore, predictions can be completed within seconds, achieving near real-time assessment of liver function and providing physicians with proactive risk warnings, facilitating early intervention. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data, which relates to this invention, helps to improve automation and intelligence. The entire process, from data preprocessing to prediction result output, is fully automated, reducing manual operation costs and subjective errors. Attached Figure Description Figure 1 This is a schematic diagram of the overall process of the method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to the present invention. Figure 2The ROC curve of the random forest model provided in the embodiment of the method for predicting aspartate aminotransferase levels based on machine learning and multi-source clinical data of the present invention; Figure 3 This is a bar chart showing the feature importance of the SHAP random forest model provided in an embodiment of the method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to the present invention. Detailed Implementation The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example, refer to Figure 1 Specifically, it includes the following steps: S1. Acquire multi-source medical data; This study retrospectively collected data from 1169 patients with abnormal liver function who visited a hospital and underwent liver function tests during a specific time period. Based on relevant diagnostic criteria, patients were divided into a normal group (AST ≤ 40 U / L, n=1143) and an abnormal group (AST > 40 U / L, n=26). Data collected included: a) Basic clinical information: age, height, weight, BMI, systolic blood pressure, diastolic blood pressure; b) Complete blood count indicators: white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelet count; c) Uric acid; Specific parameters can be found in Table 1.

[0004] Table 1. List of parameters collected in step S1 of the embodiment S2, Data Preprocessing; S21. Missing value handling: The missing data ratio in the example is less than 30%, and there are only 1169 patients, which is a small sample. Therefore, simple filling is performed. The median is used to fill numerical features, and the mode is used to fill categorical features. S22. Positive and negative sample classification: The normal range of ALT in clinical practice is 0-40 U / L. Using an ALT value of 40 as the dividing line, cases with an ALT value less than or equal to 40 are classified as positive samples and coded as 1, while those with an ALT value greater than 40 are classified as negative samples and coded as 0. S23. Data Standardization: Scale all numerical features using the StandardScaler method provided by sklearn to convert the data into a distribution with a mean of 0 and a standard deviation of 1. S24. Feature Correlation Validation: Input all feature matrices and the target variable AST (aspartate aminotransferase), i.e., select all collected clinical indicators. Combined with AST, perform Pearson correlation analysis using IBM SPSS 27.0 (Chicago, IL, USA). Sort the results according to the absolute value of the Pearson correlation coefficient, and select the top 17 with the largest absolute values ​​to filter out the 17 indicators required in S1. Analyze the grouped data using IBM SPSS 27.0 (Chicago, IL, USA). First, perform binary logistic regression analysis, including AST as the dependent variable, and age, height, weight, BMI, systolic blood pressure, diastolic blood pressure, white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelet count, and uric acid as covariates. Select "Forward: Conditional" for the method and "Probability" for saving. Obtain the predicted probability PRE_1. The second step is to plot the ROC curve, including the predicted probability PRE_1 as the test variable and AST as the state variable, with a state variable value defined as 2. In the display, check the ROC curve, diagonal reference line, standard error, and confidence interval. The obtained AUC value is 0.754, which is greater than 0.7, validating that the selected 17 indicators can serve as key features for predicting AST levels. To build and evaluate the predictive model, a dataset containing a total of 1169 cases was used. To ensure the scientific rigor and independence of the development and testing process, stratified random sampling was employed to divide the dataset into two subsets in a 4:1 ratio: a training set (935 cases) for feature selection and model training, and an independent test set (234 cases) for final model performance validation. Stratified sampling aimed to ensure that the class distribution of the target variable (abnormal AST levels) remained consistent across the training and test sets, thus avoiding sampling bias. Specifically, the data distribution was as follows: the training set included 20 patients with abnormal AST levels and 915 patients with normal AST levels; the test set included 6 patients with abnormal AST levels and 228 patients with normal AST levels. In subsequent model development processes, including model selection and training, only the training set data was used. The final trained model will be evaluated on a test set independent of the development set to objectively reflect its generalization ability. It is important to emphasize that before applying the model, the data in both the training and testing sets will be transformed according to a unified and complete preprocessing workflow. S3, Model Training; Using a training dataset (Train Set) containing data from control patients with normal AST levels and patients with abnormal AST levels, data preprocessing was performed based on step S2. Ten machine learning models were trained, including logistic regression, random forest, gradient boosting, SVM, XGBoost, LightGBM, CatBoost, and three voting models (hard voting, average soft voting, and weighted soft voting). The model selection objective was to maximize the average AUC and accuracy. The AUC of the logistic regression model was 0.694, and the accuracy was 0.726; the AUC of the random forest model was 0.884, and the accuracy was 0.974; the AUC of the gradient boosting model was 0.637, and the accuracy was 0.970; the AUC of the SVM model was 0.178, and the accuracy was 0.944; the AUC of the XGBoost model was 0.734, and the accuracy was 0.974; the AUC of the LightGBM model was 0.682, and the accuracy was 0.974; the AUC of the CatBoost model was 0.705, and the accuracy was 0.974; the AUC of the hard voting model was 0.5, and the accuracy was 0.974; the AUC of the average soft voting model was 0.774, and the accuracy was 0.974; the AUC of the weighted soft voting model was 0.774, and the accuracy was 0.974. The random forest model, which had the highest AUC and accuracy, was ultimately selected as the prediction model used in this invention. The parameter space for optimizing the random forest search includes: n_estimators ([100, 200]) and num_leaves ([15, 31]). In this example, the two sets of samples differ significantly, so class_weight='balanced' is set to handle potential class imbalance. Finally, the optimal parameter combination is determined to be n_estimators=200 and num_leaves=15. S4, prediction of aspartate aminotransferase (AST) levels; On the test set data, the aspartate aminotransferase (AST) level of patients with abnormal liver function was predicted. First, the same preprocessing procedure as in step S2 was applied to the test set data. Then, the extracted feature values ​​were input into the random forest model with the optimal parameter combination obtained in step S3. The model outputs the predicted AST level for each patient. Based on a preset probability threshold of 0.5, patients with a predicted probability greater than or equal to 0.5 were classified as "abnormal" (corresponding data is represented as 1), and patients with a predicted probability less than 0.5 were classified as "normal" (corresponding data is represented as 0). like Figure 2As shown, the ROC curve of the random forest model on the test set is displayed in detail. This curve reflects the relationship between the true positive rate and false positive rate of the random forest model under different classification thresholds. The AUC is 0.89 (rounded to two decimal places), which is a quantitative indicator of the overall classification performance of the random forest model. S5. Model Explanation; The random forest model trained in step S4 is interpreted using the SHAP (SHapley Additive exPlanations) library. The SHAP value for each sample in the test set is calculated and averaged. Generate a SHAP feature importance bar chart as follows: Figure 3 As shown, this graph sorts 17 features based on their mean absolute SHAP value, quantifying their global importance. Figure 3 The BMI average SHAP value has the highest absolute value, indicating that it has the strongest global impact on the model's prediction results; age, uric acid, and systolic blood pressure decrease in that order, while weight has the weakest impact. It should be noted that in this article, relational terms such as first and second are only used to refer to... Distinguishing one entity or operation from another does not necessarily require or imply any such actual relationship or order between those entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data, characterized in that, Includes the following steps: S1. Obtain multi-source clinical data; S2. Preprocess and verify the feature correlation of the multi-source clinical data; S3. Train the prediction model based on the preprocessed data; S4. Input the clinical data of the subjects to be evaluated into the trained prediction model, and the prediction model will then... Output the prediction results of whether AST (aspartate aminotransferase) is within the normal range.

2. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that, The multi-source clinical data includes: a) Basic clinical information: age, height, weight, BMI, systolic blood pressure, diastolic blood pressure; b) Complete blood count indicators: white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelet count; c) Uric acid.

3. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that, Step S2 includes: S21. Handling missing values; S22. Based on the clinical normal range of AST (aspartate aminotransferase) of 0-40 U / L, positive and negative samples are divided with an ALT value of 40 as the dividing line. S23. Standardize the numerical features; S24. Use Pearson correlation analysis and binary logistic regression to verify the effectiveness of the features.

4. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that: The prediction model in step S3 is a random forest prediction model.

5. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 4, characterized in that: In step S3, the prediction model training process includes hyperparameter tuning, whereby the hyperparameters include the number of decision trees, the number of leaf nodes, and the class weights.

6. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that, It also includes: step S5, using the SHAP method to interpret the prediction model and generate a feature importance bar chart.

7. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that: In step S3, when training the prediction model, the collected multi-source clinical data is divided into a training set and a test set. Based on step S2, preprocessing is performed to train the machine learning prediction model. The machine learning prediction model takes as input the features that are verified for correlation in step S2 and outputs the classification result of whether the aspartate aminotransferase level is normal.

8. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that: In step S4, the clinical data of the subject to be evaluated needs to be preprocessed according to step S2 before being input into the trained prediction model.

9. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 3, characterized in that: Step S24 includes inputting the complete feature matrix and the target variable aspartate aminotransferase (AST), performing Pearson correlation analysis using IBM SPSS 27.0, selecting the top 17 with the largest absolute values, and finally performing binary logistic regression analysis using IBM SPSS 27.0 to verify the effectiveness of the key features.

10. The method for predicting aspartate aminotransferase (AST) levels based on machine learning and multi-source clinical data according to claim 1, characterized in that: The multi-source clinical data includes anthropometric indicators and laboratory test indicators; Among them, anthropometry indicators include age, height, weight, BMI, systolic blood pressure, and diastolic blood pressure; Laboratory test indicators include white blood cell count, neutrophil count, NLR, monocyte count, lymphocyte count, red blood cell count, hemoglobin, neutrophil percentage, lymphocyte percentage, platelets, and uric acid.

Images