Prediction model construction method for traumatic stress liver injury based on machine learning
By constructing a predictive model for traumatic stress-induced liver injury using machine learning, we have addressed the shortcomings of existing assessment methods, enabling early risk assessment and warning, reducing mortality and treatment costs, and improving diagnostic accuracy.
Patent Information
- Application Number
- CN202511070691.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-10-09
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-14
AI Technical Summary
Existing methods for assessing traumatic stress-induced liver injury rely on liver function tests and trauma scoring systems, which are not ideal for assessment and lack effective early warning and intervention mechanisms, thus increasing patients' disease risk and treatment costs.
Machine learning methods were used to construct a predictive model for traumatic stress-induced liver injury. By collecting and processing data from trauma patients, relevant risk factors were screened, and Logistic regression, random forest, decision tree, and XGBoost models were established. Hyperparameters were tuned, and the SHAP algorithm was used for feature visualization. An optimized logistic model and nomogram were established to improve the clinical applicability and accuracy of the model.
It enables early risk assessment and prediction of traumatic stress-induced liver injury, provides an early warning and intervention mechanism, reduces mortality and treatment costs, and improves diagnostic accuracy and clinical efficacy.
Smart Images

Figure CN120954713A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of clinical data analysis and processing technology, and in particular relates to a method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning. Background Technology
[0002] Stress-induced liver injury often occurs after severe multiple traumas. Through various pathological mechanisms, it causes a large number of hepatocytes to apoptosis and be damaged, resulting in varying degrees of liver function impairment in the normal liver. Clinically, it often manifests as abnormal liver enzymes, jaundice, and liver failure. It is a common trauma complication that is easily overlooked in clinical practice. Because the clinical manifestations of stress-related injuries are atypical or masked by other symptoms and signs of trauma, there is a lack of effective clinical warning and intervention mechanisms, which increases the difficulty of clinical diagnosis and treatment. The incidence of stress-induced liver injury varies among different types of trauma. The incidence rate in moderate to severe traumatic brain injury is 23.5%-27.2%, with multiple trauma combined with stress-induced liver injury being the most common, with an incidence rate as high as 44.4%.
[0003] The occurrence of stress-induced liver injury not only increases the patient's disease risk and treatment costs, but severe liver function impairment will lead to an increased mortality rate in trauma patients. Studies have shown that the mortality rate of trauma patients who develop HSI is significantly higher than that of the non-HSI group. Therefore, for patients who develop HSI, drugs that increase the burden on the liver should be discontinued in time and appropriate liver-protective treatment should be given.
[0004] While there is considerable research on the pathogenesis, diagnosis, and treatment of stress-induced liver injury (STLI), a mature model for risk assessment and prediction remains lacking. Existing early assessment methods primarily rely on liver function tests (such as serum biochemical markers, alanine aminotransferase (ALT), and aspartate aminotransferase (AST)) and trauma scoring systems (such as APACHE and MELD scoring systems), but the assessment effectiveness is not yet ideal. In recent years, artificial intelligence (AI) has been widely applied in clinical disease diagnosis and prediction, as well as in assisting clinical treatment decisions. The Los Angeles Children's Hospital, based on... Hospital electronic health records, using recurrent neural networks, predicted the mortality rate of 12,000 PICU patients with an accuracy rate of 93%, significantly higher than hospital rating systems. DeepMind in the US collected medical data from over 700,000 veterans in the US military medical system and trained a deep learning system based on the database, which can accurately predict 55.8% of acute kidney injuries 48 hours in advance, representing a breakthrough in the application of artificial intelligence for disease prediction. Therefore, dimensionality reduction, mining, analysis, and modeling of massive amounts of trauma clinical data, laboratory indicators, and imaging data based on artificial intelligence may be a new approach to assessing stress-induced liver injury.
[0005] To address these issues, a machine learning-based predictive model for traumatic stress-induced liver injury was designed. Summary of the Invention
[0006] The purpose of this application is to provide a method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning. This method can help medical staff monitor patients' conditions, especially the early clinical prediction of traumatic stress-induced liver injury, optimize patient treatment plans, take timely treatment measures, and prevent the condition from deteriorating and the occurrence of poor prognosis.
[0007] To achieve the above objectives, this application provides a method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury, comprising:
[0008] We collected basic information and clinical test data of trauma patients from multiple medical centers, identified patients with traumatic stress-induced liver injury who met the criteria, and established a database of traumatic stress-induced liver injury.
[0009] By reviewing relevant literature and consulting with clinical experts, general information such as personal history in medical records was extracted to screen out important research indicators.
[0010] Missing values in the original dataset of the database were imputed using the median, and the data was processed using standardization.
[0011] Count data were expressed as rates and percentages. In the training set, analysis of variance was used for continuous variables, and correlation analysis was performed using the chi-square distribution for discrete values to screen for relevant risk factors in patients with traumatic stress-induced liver injury. Factors with a significance (P) value less than 0.05 were ranked from highest to lowest and used as input variables for the model.
[0012] We used a search-based combination approach to perform regression search to reduce variables and obtain the optimal variable combination. We then built Logistic regression model, random forest model, decision tree model, and XGBoost model on the training set.
[0013] By using functions to visualize the parameters of complex models, and based on the first four models, we build an optimized logistic model and nomogram to improve the clinical applicability of the models;
[0014] The evaluation method for the models was to use 5-fold hierarchical cross-validation to assess the predictive performance of the four models for post-traumatic stress liver injury. ROC curves were constructed using the true positive rate (TPR) and false positive rate (FPR) of the models. The model performance was evaluated based on the area under the average ROC curve (AUC), sensitivity, precision, specificity, and accuracy. The model with the highest accuracy was selected for evaluation. The optimized logistic model was further evaluated using calibration curves and DCA decision curves to improve its clinical utility.
[0015] Hyperparameter tuning of the model can further improve its performance.
[0016] Furthermore, the method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury provided in this application includes, in part, obtaining patients with traumatic stress-induced liver injury who meet the criteria:
[0017] Obtain clinical indicators, laboratory indicators, underlying liver disease, and changes in relevant indicators after trauma for all trauma patients aged 18 years or older, and meet one of the following criteria:
[0018] Mild cases: Abnormal levels of alanine aminotransferase (ALT) and aspartate aminotransferase (AST), with or without fatigue, gastrointestinal symptoms, jaundice, etc.
[0019] Severe cases: Acute or subacute liver failure, clinically manifested as extreme fatigue, severe loss of appetite, vomiting and other gastrointestinal symptoms, elevated transaminases, elevated bilirubin >171umol / L (or daily increase >17.1umol / L), prothrombin activity PTA <40% or INR >1.5, and may develop grade II or higher hepatic encephalopathy.
[0020] Furthermore, the method for constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this application, wherein the criteria for obtaining traumatic stress-induced liver injury patients should exclude those who have died, have significant damage to the liver and kidneys, have acquired immunodeficiency syndrome, have liver tumors, or have a history of spontaneous liver injury.
[0021] Furthermore, the method for constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this application includes the following preliminary patient screening indicators: traumatic brain injury, degree of burns, whether there are multiple injuries, injury classification, hospital admission condition, abdominal rebound tenderness, hemoperitoneum, whether surgery has been performed, smoking history, body temperature-categorical variable, injury site classification score, burn area, age, pulse pressure, systolic blood pressure, injury time, and heart rate.
[0022] The clinical laboratory data includes: the patient's personal history, physical measurements, injury indicators upon admission, and laboratory indicators within 12 hours of admission.
[0023] Furthermore, the method for constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this application includes the following test indicators that have been preliminarily screened through statistical analysis within the 12 hours: hepatitis B surface antigen, hepatitis B e antigen, mean corpuscular hemoglobin level, white blood cell count, creatine kinase, platelet distribution width, neutrophil percentage, red blood cell distribution width (SD), eosinophil percentage, hemoglobin, neutrophil count, lactate dehydrogenase, and C-reactive protein.
[0024] Furthermore, the method for constructing a machine learning-based prediction model for traumatic stress-induced liver injury provided in this application includes, in which the regression search algorithm comprises:
[0025] Input: Feature matrix X, target variable y, model evaluation criteria (model performance metrics (accuracy, AUC, etc.), stopping condition (stop when the maximum number of iterations is reached or the performance improvement is not significant);
[0026] Initialization: Select a full field as the seed (parent), and initialize the optimal model as the full field model;
[0027] while the parent generation has not met the cessation conditions;
[0028] For each field, use `IN parent`;
[0029] Generate offspring: Delete the current field and build a new model;
[0030] Calculate the performance of the offspring (using criterion);
[0031] Select the optimal offspring (the offspring with the best performance);
[0032] If the performance of the best offspring is better than that of the parent;
[0033] Update the parent generation to the best offspring;
[0034] else: Stop the loop (the parent is a local optimum);
[0035] Output: Locally optimal feature subset, optimal model.
[0036] Furthermore, the method for constructing a machine learning-based prediction model for traumatic stress-induced liver injury provided in this application includes visualizing complex model parameters using functions and employing the SHAP algorithm to evaluate the importance of each indicator to the model. SHAP (SHapley Additive exPlanations) is a method for interpreting the prediction results of machine learning models, and the SHAP value is calculated based on the following formula:
[0037]
[0038] Where N is the set of all features, S is the subset of features excluding feature i, f(S) is the SHAP value of feature i, and φ i (f) is the model's predicted output given a subset of features S.
[0039] Furthermore, the method for constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this application includes a 5-fold hierarchical cross-validation evaluation method. This method involves dividing the dataset into five equal subsets, using four subsets as training data each time, and using the remaining subset as validation data. This process is repeated five times, with a different subset used as validation data each time. Finally, the average of the five validation results is used as the performance evaluation metric of the model.
[0040] Furthermore, the method for constructing a machine learning-based prediction model for traumatic stress-induced liver injury provided in this application involves constructing an ROC curve using the model's true positive rate (TPR) and false positive rate (FPR), evaluating model performance based on the area under the average ROC curve (AUC), sensitivity, precision, specificity, and accuracy, and selecting the model with the highest accuracy for evaluation.
[0041] The expressions for sensitivity, specificity, and accuracy are:
[0042]
[0043] TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.
[0044] Furthermore, the method for constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this application includes hyperparameter tuning of the model based on Bayesian optimization. Bayesian optimization assumes that the search model space is Gaussian distributed and uses a Gaussian process to iteratively calculate new optimal parameters that are expected to be improved compared to the current optimal parameters. The Bayesian optimization algorithm is as follows:
[0045]
[0046] Input: f is the model, M is the Gaussian fitting function, X is the parameters, and S is the parameter selection algorithm (AcquisitionFunction).
[0047] Initialize the dataset D for fitting the Gaussian distribution as (x, y), where x is the hyperparameter and y is the result of the hyperparameter x (such as accuracy).
[0048] Iterate T times;
[0049] In each iteration, a Gaussian distribution function is fitted using the D dataset;
[0050] Based on the fitted function, and according to the Acquisition Function (such as the Expected Improvement algorithm), select a parameter xi in the parameter space that is better than the current optimal solution;
[0051] Substitute the parameters xi into the model f (train a model) to obtain the corresponding yi (the accuracy of the new model, etc.);
[0052] (xi,yi) are re-added to the fitted dataset D, and the iteration continues.
[0053] Compared with existing technologies, this application has the following advantages: it innovatively uses machine learning to construct a stress-induced liver injury assessment model, and establishes a SHAP function to visualize model features. The model features used for feature importance have a guiding role in clinical diagnosis. Furthermore, it establishes a simple and intuitive nomogram model to accurately assess and predict the risk of traumatic stress-induced liver injury, which helps to establish an early warning and clinical intervention mechanism for traumatic stress-induced liver injury. Attached Figure Description
[0054] Figure 1 This is a simplified flowchart of the method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning, provided in an embodiment of the present invention.
[0055] Figure 2 This is a SHAP feature variable importance diagram of four prediction model construction methods for traumatic stress liver injury based on machine learning provided in the embodiments of the present invention;
[0056] Figure 3 This is the ROC curve of the method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning provided in this embodiment of the invention.
[0057] Figure 4 This is a nomogram of predicted probability grouping statistics for the method of constructing a prediction model for traumatic stress-induced liver injury based on machine learning provided in this embodiment of the invention.
[0058] Figure 5 This is an optimized logistic calibration curve of the method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning provided in this embodiment of the invention.
[0059] Figure 6 This is the optimized logistic model DCA decision curve of the machine learning-based predictive model construction method for traumatic stress-induced liver injury provided in this embodiment of the invention.
[0060] Figure 7 This is a decision curve analysis diagram of the method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning, provided in an embodiment of the present invention. Detailed Implementation
[0061] 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.
[0062] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings, and the technical effects that the present invention can achieve will be further explained.
[0063] This study analyzed the clinical data of trauma patients using mature machine learning models such as Logistic Regression, Decision Tree, XGBoost, and Random Forest, analyzed the influencing factors of post-traumatic stress liver injury, and constructed an early HSI prediction model to provide a reference for timely and accurate intervention plans for patients.
[0064] Step 1: Collect basic information and clinical laboratory data from trauma patients from multiple medical centers, removing indicators with a missing rate greater than 30%. Finally, data from 677 patients were collected and compiled according to the established criteria for inclusion in subsequent studies. A database of traumatic stress-induced liver injury was established to identify eligible patients.
[0065] Step 2: By reviewing relevant literature and consulting with clinical experts, general information was extracted from the patient's personal history, admission injury details, and laboratory indicators in the medical records. Important research indicators were then screened out, with a total of 63 characteristic variables selected, including 39 binary variables and 24 continuous variables.
[0066] Step 3: Use the median to impute missing values in the original dataset from Step 1, and then perform data processing using standardization.
[0067] Step 4: Count data are expressed as rates and percentages. In the training set, analysis of variance is used for continuous variables, and correlation analysis is performed on discrete values using chi-square distribution. Those with stress-induced liver injury are the case group, and those without stress-induced liver injury are the control group. Multiple research indicators are included. Factors with a significance level (P) less than 0.05 were ranked from highest to lowest and used as input variables for the model. Among them, the following 30 indicators showed statistically significant differences (P<0.05): traumatic brain injury, burn severity, multiple injuries, injury classification, hospital admission condition, abdominal rebound tenderness, hemoperitoneum, surgical history, smoking history, body temperature-categorical variable, injury site classification score, burn area, age, pulse pressure, systolic blood pressure, injury time, heart rate, hepatitis B surface antigen, hepatitis B e antigen, mean corpuscular hemoglobin, white blood cell count, creatine kinase, platelet distribution width, neutrophil percentage, red blood cell distribution width (SD), eosinophil percentage, hemoglobin, neutrophil count, lactate dehydrogenase, and C-reactive protein. (See Tables 1 and 2).
[0068] Table 1. Comparison of binary classification variables included in the model for HSI and non-HSI groups [Example (%)]
[0069]
[0070] Table 2 Comparison of quantitative variables between the HSI group and the non-HSI group
[0071]
[0072] Step 5: Model Building and Research. SPSS 24.0 and Python 3.9 software were used for data processing. Count data were expressed as rates and percentages. In the training set, analysis of variance was used for continuous variables, and correlation analysis was performed using chi-square distribution for discrete values to screen relevant risk factors for HSI. Factors with p-values less than 0.05 were ranked from highest to lowest and used as input variables. The occurrence of traumatic stress-induced liver injury was used as the outcome variable. Logistic regression, random forest, decision tree, and XGBoost models were built in the training set. Then, based on the search and combination algorithm, regression search was performed to reduce variables and obtain the optimal variable combination. The regression search algorithm can be described in detail using the following framework:
[0073] 1. Input: Feature matrix X, target variable y, model evaluation criteria (model performance metrics (accuracy, AUC, etc.), stopping condition (stop when the maximum number of iterations is reached or the performance improvement is not significant).
[0074] 2. Initialization: Select a full field as the seed (parent), and initialize the optimal model as the full field model.
[0075] 3. While the parent generation has not met the stopping condition:
[0076] 4. For each field, use `IN parent`:
[0077] 5. Generate offspring: Delete the current field and build a new model.
[0078] 6. Calculate the performance of the offspring (using criterion)
[0079] 7. Select the optimal offspring (the offspring with the best performance).
[0080] 8. If the performance of the best offspring is better than that of the parent:
[0081] 9. Update the parent generation to the best offspring generation.
[0082] 10. else: Stop the loop (if the parent is a local optimum).
[0083] 11. Output: Locally optimal feature subset, optimal model
[0084] Step Six: Visualize the parameters of complex models using functions, and based on the first four models, build an optimized logistic model and nomogram to improve the clinical applicability of the model;
[0085] The parameters of complex models are visualized using functions. Finally, the SHAP algorithm is used to evaluate the importance of each metric to the model. SHAP (SHapley Additive exPlanations) is a method for interpreting the prediction results of machine learning models. The SHAP value is calculated based on the following formula:
[0086]
[0087] Where N is the set of all features, S is the subset of features excluding feature i, f(S) is the SHAP value of feature i, and φ i (f) is the model's predicted output given a subset of features S.
[0088] The visualization results of the SHAP model are shown in the figure. The results show the importance percentage of the top 20 features of each model, which clearly shows the risk factors affecting patients with traumatic stress-induced liver injury.
[0089] After comparing the performance of the four models, although the models performed well, they had too many feature parameters. To further reduce the number of parameters and improve clinical applicability, the top 10 features of the four models were extracted to build a logistic model, which included traumatic brain injury, lactate dehydrogenase, hemoglobin, neutrophil percentage, mean corpuscular hemoglobin content, creatine kinase, red blood cell distribution width (SD), injury site classification score, age, and hemoperitoneum. A nomogram model was also built for visual visualization.
[0090] Step 7: Model Evaluation Metrics
[0091] This paper evaluates the predictive performance of the four models for post-traumatic stress-induced liver injury using 5-fold hierarchical cross-validation. 5-fold cross-validation is a commonly used model evaluation method. It divides the dataset into five equal subsets, using four subsets as training data and the remaining subset as validation data in each iteration. This process is repeated five times, using a different subset each time. Finally, the average of the five validation results is used as the model's performance metric. Its advantage lies in reducing the risk of overfitting and obtaining a more stable and reliable performance evaluation.
[0092] ROC curves were constructed using the model's true positive rate (TPR) and false positive rate (FPR). Model performance was evaluated based on the area under the average ROC curve (AUC), sensitivity, precision, specificity, and accuracy. The model with the highest accuracy was selected for evaluation.
[0093] The expressions for sensitivity, specificity, and accuracy are:
[0094]
[0095] TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.
[0096] The results are shown in Table 3. The area under the ROC curve (AUC) of each model reached a maximum of 0.92, indicating excellent predictive performance. Specifically, the nomogram model's predictive performance was evaluated using the calibration curve, revealing that the fitted curve closely overlapped with the reference line, indicating strong diagnostic capability. Furthermore, the DCA decision curve showed that the diagnostic model had significant clinical value in predicting the probability of HSI formation in the range of 0.2-0.9.
[0097] The performance metrics of each model are shown in Table 3 below:
[0098] Table 3 Comparison of model performance based on search combinatorial algorithm
[0099]
[0100]
[0101] Step 8: Perform hyperparameter tuning on the model to further improve its performance.
[0102] The parameter tuning method used in this invention is Bayesian optimization, which is a solution in this case. Bayesian optimization assumes that the search model space is Gaussian distributed, and uses a Gaussian process to iteratively calculate new optimal parameters that are expected to be improved compared to the current optimal parameters.
[0103] The Bayesian optimization algorithm is as follows:
[0104]
[0105] Input: f is the model, M is the Gaussian fitting function, X is the parameters, and S is the parameter selection algorithm (AcquisitionFunction).
[0106] Initialize the dataset D, (x, y), to fit the Gaussian distribution, where x is a hyperparameter and y is the result of the hyperparameter x (such as accuracy).
[0107] Iterate T times
[0108] In each iteration, a Gaussian distribution function is fitted using the D dataset.
[0109] Based on the fitted function, and according to the Acquisition Function (such as the Expected Improvement algorithm), a parameter xi that is better than the current optimal solution is selected in the parameter space.
[0110] Substitute the parameters xi into the model f (train a model) to obtain the corresponding yi (the accuracy of the new model, etc.).
[0111] (xi,yi) are re-added to the fitted dataset D, and the iteration continues.
[0112] The key parameters involved in the four models of this invention are: Random Forest: n_estimators = 100, min-n = 40, tress = 1991, mtry = 10; Decision Tree Model: criterion = "entropy".
[0113] splitter="best", random_state="42, class_weight="balanced"; catboost: iterations="Logloss", learning_rate="Logloss", learning_rate="0.1, depth="6; logistic regression model: max_iter="auto", multi_class="auto", intercept_scaling="l2".
[0114] In summary, this invention uses various machine learning techniques to predict traumatic stress-induced liver injury and screen out risk factors that influence this condition, providing significant guidance for the detection and treatment of patients with traumatic stress-induced liver injury.
[0115] Any aspects of this invention not described in detail are well-known to those skilled in the art.
[0116] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning or limited experimentation on the basis of the prior art should be within the scope of protection defined by the claims.
Claims
1. A method for constructing a predictive model for traumatic stress-induced liver injury based on machine learning, characterized in that, include: We collected basic information and clinical test data of trauma patients from multiple medical centers, identified patients with traumatic stress-induced liver injury who met the criteria, and established a database of traumatic stress-induced liver injury. By reviewing relevant literature and consulting with clinical experts, general information such as personal history in medical records was extracted to screen out important research indicators. Missing values in the original dataset of the database were imputed using the median, and the data was processed using standardization. Count data were expressed as rates and percentages. In the training set, analysis of variance was used on continuous variables, and correlation analysis was performed on discrete values using chi-square distribution to screen for relevant risk factors in patients with traumatic stress-induced liver injury. For relevant factors with a significance (P) value less than 0.05, they were ranked from high to low and used as input variables for the model; We used a search-based combination approach to perform regression search to reduce variables and obtain the optimal variable combination. We then built Logistic regression model, random forest model, decision tree model, and XGBoost model on the training set. By using functions to visualize the parameters of complex models, and based on the first four models, we build an optimized logistic model and nomogram to improve the clinical applicability of the models; The evaluation method for the models was to use 5-fold hierarchical cross-validation to assess the predictive performance of the four models for post-traumatic stress liver injury. ROC curves were constructed using the true positive rate (TPR) and false positive rate (FPR) of the models. The model performance was evaluated based on the area under the average ROC curve (AUC), sensitivity, precision, specificity, and accuracy. The model with the highest accuracy was selected for evaluation. The optimized logistic model was further evaluated using calibration curves and DCA decision curves to improve its clinical utility. Hyperparameter tuning of the model can further improve its performance.
2. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, The criteria for obtaining patients with traumatic stress-induced liver injury include: Obtain clinical indicators, laboratory indicators, underlying liver disease, and changes in relevant indicators after trauma for all trauma patients aged 18 years or older, and meet one of the following criteria: Mild cases: Abnormal levels of alanine aminotransferase (ALT) and aspartate aminotransferase (AST), with or without fatigue, gastrointestinal symptoms, jaundice, etc. Severe cases: Acute or subacute liver failure, clinically manifested as extreme fatigue, severe loss of appetite, vomiting and other gastrointestinal symptoms, elevated transaminases, elevated bilirubin >171umol / L (or daily increase >17.1umol / L), prothrombin activity PTA <40% or INR >1.5, and may develop grade II or higher hepatic encephalopathy.
3. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, Patients meeting the criteria for traumatic stress-induced liver injury should exclude those who are deceased, have significant liver and kidney damage, have acquired immunodeficiency syndrome, have liver tumors, or have a history of spontaneous liver injury.
4. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, The initial screening criteria for patients include: traumatic brain injury, degree of burns, whether there are multiple injuries, injury classification, admission condition, abdominal rebound tenderness, hemoperitoneum, whether surgery has been performed, smoking history, body temperature-categorical variable, injury site classification score, burn area, age, pulse pressure, systolic blood pressure, time of injury, and heart rate. The clinical laboratory data includes: the patient's personal history, physical measurements, injury indicators upon admission, and laboratory indicators within 12 hours of admission.
5. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 4, characterized in that, The test indicators selected through preliminary statistical analysis within the 12 hours include: hepatitis B surface antigen, hepatitis B e antigen, mean corpuscular hemoglobin, white blood cell count, creatine kinase, platelet distribution width, neutrophils %, red blood cell distribution width (SD), eosinophils %, hemoglobin, neutrophil count, lactate dehydrogenase, and C-reactive protein.
6. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, The regression search algorithm includes: Input: Feature matrix X, target variable y, model evaluation criteria (model performance metrics (accuracy, AUC, etc.), stopping condition (stop when the maximum number of iterations is reached or the performance improvement is not significant); Initialization: Select a full field as the seed (parent), and initialize the optimal model as the full field model; while the parent generation has not met the cessation conditions; For each field, use `IN parent`; Generate offspring: Delete the current field and build a new model; Calculate the performance of the offspring (using criterion); Select the optimal offspring (the offspring with the best performance); If the performance of the best offspring is better than that of the parent; Update the parent generation to the best offspring; else: Stop the loop (the parent is a local optimum); Output: Locally optimal feature subset, optimal model.
7. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, The method involves visualizing complex model parameters using functions and employing the SHAP algorithm to evaluate the importance of each metric to the model. SHAP (SHapley Additive exPlanations) is a method for interpreting the prediction results of machine learning models. The SHAP value is calculated based on the following formula: Where N is the set of all features, S is the subset of features excluding feature i, f(S) is the SHAP value of feature i, and φ i (f) is the model's predicted output given a subset of features S.
8. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, Five-fold hierarchical cross-validation evaluation involves dividing the dataset into five equal subsets. Each time, four subsets are used as training data, and the remaining subset is used as validation data. This process is repeated five times, using a different subset as validation data each time. Finally, the average of the five validation results is used as the performance evaluation metric for the model.
9. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 8, characterized in that, ROC curves were constructed using the model's true positive rate (TPR) and false positive rate (FPR). Model performance was evaluated based on the area under the average ROC curve (AUC), sensitivity, precision, specificity, and accuracy. The model with the highest accuracy was selected for evaluation. The expressions for sensitivity, specificity, and accuracy are: TP represents a true positive, TN represents a true negative, FP represents a false positive, and FN represents a false negative.
10. The method for constructing a machine learning-based predictive model for traumatic stress-induced liver injury as described in claim 1, characterized in that, The hyperparameter tuning of the model described above is based on Bayesian optimization. Bayesian optimization assumes that the search model space follows a Gaussian distribution and uses a Gaussian process to iteratively calculate new optimal parameters that are expected to improve upon the current optimal parameters. The Bayesian optimization algorithm is as follows: Input: f is the model, M is the Gaussian fitting function, X is the parameter, S is the parameter selection algorithm. The dataset D is initialized with Gaussian distribution fitting as (x, y), where x is the hyperparameter and y is the result of the hyperparameter x (such as accuracy). Iterate T times; In each iteration, a Gaussian distribution function is fitted using the D dataset; Based on the fitted function, and according to the Acquisition Function (such as the Expected Improvement algorithm), select a parameter xi in the parameter space that is better than the current optimal solution; Substitute the parameters xi into the model f (train a model) to obtain the corresponding yi (the accuracy of the new model, etc.); (xi,yi) are re-added to the fitted dataset D, and the iteration continues.
