Risk assessment method and system for vancomycin-related acute kidney injury

Abstract

The invention relates to the technical field of medical monitoring, in particular to a risk assessment method and system for vancomycin-related acute kidney injury, and the method comprises the steps: collecting clinical data of a patient receiving vancomycin treatment, carrying out grading labeling, and dividing the clinical data into a training set and a test set; performing data expansion on the training set sample by adopting a conditional generative adversarial network to obtain an expanded training set; performing feature screening on the extended training set to obtain a screened feature set; constructing a multi-classification integration model based on the screened feature set, obtaining the multi-classification integration model, and outputting a classification prediction result; and performing interpretability analysis on the prediction result of the multi-classification integration model by adopting an SHAP method, calculating the contribution value of each feature to the prediction result, and generating an evaluation report. According to the risk assessment method provided by the invention, the transparency of a model decision process is realized while the prediction accuracy is ensured, and the scientificity and practicability of AKI risk management in vancomycin treatment are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a method and system for risk assessment of vancomycin-related acute kidney injury. Background Technology

[0002] Vancomycin is an important antibiotic for treating Gram-positive bacterial infections and is widely used to treat serious infectious diseases such as methicillin-resistant Staphylococcus aureus (MRSA) infections and coagulase-negative staphylococcal infections. However, vancomycin has a proven history of nephrotoxicity, with acute kidney injury (AKI) being one of its most common and serious adverse reactions. Literature reports that the incidence of AKI in patients treated with vancomycin ranges from 10% to 40%, significantly impacting patient prognosis and increasing healthcare costs. According to the 2013 international standards published by the Kidney Disease: Improving Global Outcomes (KDIGO), AKI is classified into three severity levels—Grade 1, Grade 2, and Grade 3—based on the degree of serum creatinine elevation, with different clinical intervention strategies corresponding to each grade.

[0003] Existing methods for predicting vancomycin-related acute kidney injury (AKI) risk primarily employ machine learning techniques to analyze clinical data. However, in real-world clinical data, the sample distribution across different severity levels of AKI is extremely imbalanced: approximately 70% of patients have no AKI, about 15% have KDIGO 1, about 10% have KDIGO 2, and only about 5% have KDIGO 3. Traditional methods for addressing sample imbalance, such as random oversampling, are prone to overfitting. Synthetic Minority Oversampling Technique (SMOTE) generates new samples through linear interpolation but fails to consider the complex nonlinear correlations between multidimensional clinical features in medical data. Furthermore, existing methods typically employ purely data-driven feature selection algorithms, relying entirely on statistical indicators to select features. This may lead to the exclusion of key clinical factors with clear pathophysiological mechanisms due to sample characteristics or multicollinearity, resulting in a loss of clinical interpretability of the model.

[0004] Chinese patent CN202311682021.1 discloses a dynamic prediction method and system for acute kidney injury (AKI). This method collects patients' clinical time-series data at preset time intervals, preprocesses the data to obtain clinical features, selects optimal basic clinical features based on multiple feature selection models, uses TSFresh technology to enhance these optimal basic clinical features, and then concatenates the optimal basic clinical features and enhanced clinical features before inputting them into a prediction model for training. The final output is a prediction of the probability of AKI occurrence. However, this patent only predicts whether AKI will occur, but cannot directly predict the risk level of AKI, and therefore cannot provide clinicians with a basis for stratified management and personalized treatment decisions. Summary of the Invention

[0005] In view of this, the present invention proposes a risk assessment method and system for vancomycin-associated acute kidney injury to solve the problems of poor sample imbalance treatment, easy neglect of key risk factors, and insufficient accuracy of AKI risk assessment in the prior art.

[0006] The technical solution of this invention is implemented as follows: This invention provides a risk assessment method for vancomycin-related acute kidney injury, comprising the following steps: S1. Collect clinical data of patients receiving vancomycin treatment, and classify and label patients according to the KDIGO guidelines to obtain a real dataset. Preprocess the real dataset and divide it into training set and test set. S2. A conditional generative adversarial network is used to augment the training set samples. Synthetic samples are generated through adversarial training. After the synthetic samples are validated, they are combined with the training set to obtain the augmented training set. S3. LASSO regression was used to screen features in the expanded training set, while retaining vancomycin trough concentration, pre-vancomycin serum IL-6 concentration and ICU admission status as core risk factors to obtain the screened feature set. S4. Construct a Stacking multi-class ensemble model based on the filtered feature set. The Stacking multi-class ensemble model includes multiple base learners in the first layer and a meta-learner in the second layer. The multi-class ensemble model is trained and outputs the KDIGO hierarchical prediction results. S5. Use the SHAP method to perform interpretability analysis on the prediction results of the multi-class ensemble model, calculate the contribution value of each feature to the prediction results, and generate an evaluation report containing KDIGO hierarchical prediction results and feature contribution values.

[0007] Based on the above technical solutions, preferably, in step S1, the clinical data includes demographic characteristics, basic disease information, vancomycin treatment parameters, laboratory test indicators, and clinical status information; the grading label includes KDIGO level 0, KDIGO level 1, KDIGO level 2, and KDIGO level 3.

[0008] Based on the above technical solutions, preferably, step S2 specifically includes: S21. Construct a conditional generative adversarial network, which includes a generator and a discriminator. The input of the generator is a random noise vector and a conditional label, and the output is a synthetic sample. The input of the discriminator is a real sample or a synthetic sample and its corresponding conditional label, and the output is the probability that the sample is real data. S22. Train a conditional generative adversarial network on real samples. Learn the multidimensional joint probability distribution of real data through adversarial training. After training, use a conditional generation mechanism to generate minority class samples. Use random noise vectors and conditional labels to obtain synthetic samples through the generator. S23. Perform quality verification on the generated synthetic samples, use statistical tests to evaluate the consistency of the distribution of synthetic samples and real samples on each continuous feature, calculate the feature correlation matrix of real data and synthetic data to evaluate the correlation difference, and delete unreasonable synthetic samples through clinical rule filter. Combine the verified synthetic samples with the training set to obtain the expanded training set.

[0009] Based on the above technical solution, preferably, in step S23, the clinical rule filter determines the rationality of the sample by comparing the feature values ​​of the synthesized sample with a preset medical reference database. The comparison process includes: The values ​​of each physiological indicator in the synthetic sample are extracted and compared with the normal or pathological range of the corresponding indicator in the medical reference database to determine whether there are any abnormal values ​​that exceed the physiologically possible range. Multiple features with dependencies are extracted from the synthetic sample, and the logical relationship between the features is calculated to see if it meets the preset physiological constraints. The KDIGO grade label and serum creatinine change value of the synthetic sample are extracted, and the theoretical KDIGO grade of the sample is calculated according to the grade determination criteria defined in the KDIGO guidelines. It is determined whether the theoretical KDIGO grade is consistent with the label, and synthetic samples that do not meet any of the above comparison conditions are deleted.

[0010] Based on the above technical solutions, preferably, step S3 specifically includes: S31. LASSO regression is used to filter features on the expanded training set. Feature selection is achieved by introducing an L1 regularization term into the loss function. S32. Cross-validation is used to determine the optimal regularization parameter. The training set is divided into several folds. One fold is selected as the validation set and the remaining folds are used as training subsets. For candidate regularization parameter values, the model is fitted on the training subset and the prediction error is calculated on the validation set. The regularization parameter value that minimizes the average validation error is selected as the optimal regularization parameter. S33. Using the optimal regularization parameters, fit the minimum absolute shrinkage and selection operator model on the entire training set to obtain the regression coefficient vector. Select the features with non-zero regression coefficients, while forcibly retaining the vancomycin valley concentration, the serum IL-6 concentration before vancomycin administration, and the ICU admission status to obtain the filtered feature set.

[0011] Based on the above technical solutions, the preferred optimization objective for LASSO regression is: in The number of training samples. For the first The label of each sample For the first The feature vector of each sample For the regression coefficient vector, This is the regularization parameter.

[0012] Based on the above technical solutions, preferably, step S4 specifically includes: S41. Select multiple heterogeneous base learners as the first layer, wherein the base learners include support vector machines, decision trees, random forests, and multinomial logistic regression; S42. Use cross-validation to generate meta-features, divide the training set into several folds, train the base learners using the remaining folds for each fold and predict the samples in that fold, and concatenate the outputs of each base learner into a meta-feature vector. S43. Multinomial logistic regression with class weighting is used as the meta-learner of the second layer. The input is the meta-feature vector and the output is the KDIGO hierarchical probability distribution. The multi-class ensemble model is obtained by minimizing the class weighted cross-entropy loss function. S44. For the test set samples, use the trained base learner to predict the probability vectors respectively and concatenate them into a meta-feature vector. Input the meta-feature vector into the trained meta-learner to obtain the KDIGO level probability distribution, and select the level with the highest probability as the KDIGO level prediction result.

[0013] Based on the above technical solution, preferably, in step S43, the loss function of the meta-learner is: in The number of training samples. For the sample Authentic KDIGO rating labels For the sample The meta-eigenvectors, For level The corresponding weighting coefficients, Predicting samples for meta-learners Belongs to the level The probability, For prediction level Compared to the actual level The distance between them The class distance penalty coefficient; the class weight coefficient for: in The total number of samples in the training set. For level The number of samples in the training set is 4, which represents the total number of categories.

[0014] Based on the above technical solutions, preferably, step S5 specifically includes: S51. Perform interpretability analysis on the multi-class ensemble model, and use the SHAP method to calculate the feature contribution value for each KDIGO level. For samples and levels, the ... The SHAP value of a feature represents the contribution of that feature to the model's prediction of that level; S52. Calculate the SHAP value matrix of each base learner separately, and obtain the SHAP value of the ensemble model by weighting the fusion weights of each base learner in the meta-learner. S53. Extract the corresponding feature contribution values ​​for the predicted KDIGO level and sort them by absolute value to identify the feature that contributes the most to the prediction result. S54. Generate a structured evaluation report.

[0015] This invention provides a risk assessment system for vancomycin-related acute kidney injury, the system being used to perform the risk assessment method described above, comprising: The data processing module is used to collect patient clinical data and perform annotation and preprocessing, and to augment the training data with imbalanced samples. The model building module is used to filter features from the expanded training data and force the retention of core risk factors. Based on the filtered features, a multi-class ensemble model is built and trained. The risk assessment module receives clinical data from patients to be assessed, outputs KDIGO grade prediction results through a trained multi-class ensemble model, and generates an assessment report containing prediction results and feature contribution analysis based on interpretability analysis methods.

[0016] The risk assessment method and system for vancomycin-related acute kidney injury of the present invention have the following advantages over the prior art: (1) The problem of imbalanced clinical data samples is solved by generating adversarial networks through conditional tables. The KDIGO four-class prediction model is constructed using the Stacking ensemble learning framework and combined with SHAP interpretability analysis to generate personalized evaluation reports containing feature contribution information. This technical solution ensures the accuracy of prediction while making the model decision-making process transparent, enabling clinicians to understand the basis of prediction and obtain targeted intervention suggestions. This transforms the machine learning model from a "black box" tool into a reliable clinical decision support system, effectively improving the scientificity and practicality of AKI risk management in vancomycin treatment.

[0017] (2) A Conditional Table Generative Adversarial Network (CTGAN) is used to specifically expand the minority class of KDIGO-level samples. A three-layer clinical rule filtering mechanism, namely medical reference range verification, feature logical relationship verification, and KDIGO standard consistency verification, is used to ensure the medical rationality of the synthesized samples. This method not only solves the limitation of traditional resampling techniques in learning data distribution features, but also avoids the problem of ordinary GANs generating samples that do not conform to clinical rules. The expanded training set achieves relative balance of KDIGO-level samples while maintaining the original data feature distribution, effectively improving the model's ability to identify moderate and severe AKI.

[0018] (3) A two-layer stacking architecture is adopted. The first layer uses four heterogeneous base learners—support vector machine, decision tree, random forest, and multinomial logistic regression—to make preliminary predictions of KDIGO levels. The second layer uses multinomial logistic regression with class weighting and level distance penalty as a meta-learner for fusion. This ensemble strategy reduces the bias and variance of a single model by combining the advantages of different learning algorithms. At the same time, the improved loss function of the meta-learner makes the model pay more attention to minority class samples and penalize serious misclassifications across levels, thus achieving more stable prediction performance and more reasonable level discrimination ability than a single model in multi-class tasks.

[0019] (4) Based on the SHAP theory, the interpretability analysis of the Stacking ensemble model is carried out, and the contribution of each feature at each KDIGO level is quantified. This method not only solves the inherent "black box" problem of the ensemble model, but also identifies the key factors driving the high-risk prediction of specific patients and their contribution direction, providing clinicians with clear intervention targets, and expanding the model prediction from simple risk classification to decision support information containing explanatory basis. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the risk assessment method for vancomycin-related acute kidney injury according to the present invention. Figure 2 This is a flowchart of the data expansion process in step S2 of the present invention; Figure 3 This is the ROC curve of vancomycin trough concentration and the risk of AKI in this invention; Figure 4 This is an ROC curve showing the relationship between pre-vancomycin serum IL-6 concentration and the risk of AKI in this invention. Figure 5 This is a flowchart of the integrated model construction and output results in step S4 of the present invention; Figure 6 This is a flowchart of the SHAP interpretable analysis process for step S5 of the present invention; Figure 7 This is the ROC curve of the ablation experiment results of this invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the present invention provides a method for risk assessment of vancomycin-related acute kidney injury, comprising the following steps: S1. Collect clinical data of patients receiving vancomycin treatment, and classify and label patients according to the KDIGO guidelines to obtain a real dataset. Preprocess the real dataset and divide it into training set and test set. S2. A conditional generative adversarial network is used to augment the training set samples. Synthetic samples are generated through adversarial training. After the synthetic samples are validated, they are combined with the training set to obtain the augmented training set. S3. LASSO regression was used to screen features in the expanded training set, while retaining vancomycin trough concentration, pre-vancomycin serum IL-6 concentration and ICU admission status as core risk factors to obtain the screened feature set. S4. Construct a Stacking multi-class ensemble model based on the filtered feature set. The Stacking multi-class ensemble model includes multiple base learners in the first layer and a meta-learner in the second layer. The multi-class ensemble model is trained and outputs the KDIGO hierarchical prediction results. S5. Use the SHAP method to perform interpretability analysis on the prediction results of the multi-class ensemble model, calculate the contribution value of each feature to the prediction results, and generate an evaluation report containing KDIGO hierarchical prediction results and feature contribution values.

[0024] In one embodiment of the present invention, in step S1, the clinical data includes demographic characteristics, underlying disease information, vancomycin treatment parameters, laboratory test indicators, and clinical status information; the grading labeling includes KDIGO level 0, KDIGO level 1, KDIGO level 2, and KDIGO level 3. Specifically, demographic characteristics such as age, sex, and weight; underlying medical conditions such as history of diabetes, hypertension, chronic kidney disease, and abnormal liver function; vancomycin treatment parameters including vancomycin trough concentration, cumulative dose, duration of treatment, and route of administration; vancomycin trough concentration was detected using ultra-high performance liquid chromatography-tandem mass spectrometry, recording the trough concentration value after the first steady-state condition was reached; laboratory indicators included baseline serum creatinine, blood urea nitrogen, estimated glomerular filtration rate, serum IL-6, procalcitonin, C-reactive protein, white blood cell count, serum albumin, hemoglobin, platelet count, aspartate aminotransferase, and alanine aminotransferase, with serum IL-6 concentration before vancomycin administration detected using chemiluminescent immunoassay; clinical status information included ICU admission status, use of vasoactive drugs, mechanical ventilation, and concomitant medications including aminoglycoside antibiotics, loop diuretics, nonsteroidal anti-inflammatory drugs, and angiotensin-converting enzyme inhibitors. These characteristics cover the main risk factors for vancomycin-related AKI reported in existing literature, providing a comprehensive predictive basis for the model.

[0025] Patients were categorized into four groups according to the KDIGO 2013 AKI Clinical Practice Guidelines. If a patient's serum creatinine level was ≥26.5 mmol / L within 48 hours of starting vancomycin treatment, this was considered a classification. An increase in serum creatinine of 100 mol / L or more than 50% above baseline within 7 days is considered acute kidney injury (AKI). For patients with AKI, the severity is further graded based on the magnitude of the creatinine increase, and the label is defined as follows: .Label Indicates no AKI occurred; label This indicates a KDIGO grade 1 AKI, defined as an increase in serum creatinine between 26.5 and 44.2. mol / L or 1.5 to 1.9 times higher than baseline; label This indicates KDIGO stage 2 AKI, with serum creatinine increasing 2.0 to 2.9 times from baseline; label This indicates KDIGO grade 3 AKI, with serum creatinine exceeding 353.6. If the blood glucose level is mol / L or increases by 3 times or more from the baseline value, renal replacement therapy may need to be initiated.

[0026] Quality control was implemented for the four-category labeling. A dual-review mechanism was employed, with two researchers independently labeling the same patient according to the KDIGO criteria. Any discrepancies were resolved by a third party. Cases with confounding factors causing changes in creatinine were excluded, specifically including: elevated creatinine due to rhabdomyolysis (with creatine kinase levels exceeding 10 times the upper limit of normal), dilutional decrease in serum creatinine due to massive fluid resuscitation (24-hour fluid intake exceeding 5000 mL), creatinine changes within 72 hours after contrast agent administration, and creatinine changes during hemodialysis or continuous renal replacement therapy. A characteristic vector for each patient was recorded. and the corresponding KDIGO level labels ,in This represents the initial total number of features.

[0027] Preprocess the real dataset. For continuous variables, use Z-score standardization. Continuous features Calculate its mean on the training set. and standard deviation The standardized eigenvalues ​​are For categorical variables such as gender and ICU admission status, one-hot encoding was used to convert them into numerical features. For features with missing values, median imputation was used if the missing rate was less than 5%; otherwise, the feature was removed. The dataset was divided into training and test sets in a 7:3 ratio, and stratified sampling was used to ensure that the sample proportions of the four KDIGO levels in the training and test sets were consistent with the original data distribution. It is particularly important to note that the test set contains only real patient data and does not include subsequently generated synthetic data.

[0028] In one embodiment of the present invention, such as Figure 2 As shown, step S2 specifically includes: S21. Construct a conditional generative adversarial network, which includes a generator and a discriminator. The generator's input is a random noise vector and a conditional label, and its output is a synthetic sample. The discriminator's input is a real sample or a synthetic sample and its corresponding conditional label, and its output is the probability that the sample is real data.

[0029] The generator network is a multi-layer fully connected neural network with four hidden layers, each containing 256 neurons. The ReLU activation function is used. The input layer concatenates random noise and one-hot encoded conditional labels. The output layer processes features according to their type: continuous features are activated using the tanh activation function and de-standardized; discrete features are made differentiable using the Gumbel-Softmax technique. Discrete features are then processed. The input can be a real sample or a synthetic sample. and their corresponding conditional tags The output is the probability that the sample is the real data. The discriminator network structure is also a multi-layer fully connected neural network, containing 4 hidden layers with 256 neurons in each layer. The activation function is LeakyReLU, and the output layer uses the sigmoid activation function.

[0030] S22. Train a conditional generative adversarial network on real samples. Learn the multidimensional joint probability distribution of real data through adversarial training. After training, use a conditional generation mechanism to generate minority class samples. Use random noise vectors and conditional labels to obtain synthetic samples through the generator.

[0031] Specifically, the loss function of the generator is: Among them, This represents the distribution of KDIGO levels in real data. The loss function of the discriminator is: in, This represents the actual data distribution.

[0032] exist CTGAN was trained on real samples for 300 epochs with a batch size of 32. The learning rates for both the generator and discriminator were set to 0.5%. The optimizer uses the Adam algorithm. After training, a conditional generation mechanism is used to generate samples for the minority class. For KDIGO levels 2 and 3, samples are generated respectively. and A number of synthetic samples are used to ensure a relatively balanced number of samples at each level in the expanded dataset. The generation process is as follows: sampling random noise. Set conditional tags Synthetic samples are obtained through a generator. .

[0033] S23. Perform quality verification on the generated synthetic samples, use statistical tests to evaluate the consistency of the distribution of synthetic samples and real samples on each continuous feature, calculate the feature correlation matrix of real data and synthetic data to evaluate the correlation difference, and delete unreasonable synthetic samples through clinical rule filter. Combine the verified synthetic samples with the training set to obtain the expanded training set.

[0034] For the Calculate the KS statistic for each continuous characteristic. ,in and In terms of features, real samples and synthetic samples are respectively... The cumulative distribution function on [the plane]. If [the cumulative distribution function on the plane]. A p-value less than 0.05 indicates a significant difference in the feature distribution, requiring adjustment of the CTGAN training parameters and regeneration. Calculate the feature correlation matrix between the real and synthetic data. and Assess the differences in correlation ,Require To ensure that the synthetic data maintains the correlation between features.

[0035] A clinical rule filter is introduced to remove unreasonable synthetic samples. The clinical rule filter determines the reasonableness of a sample by comparing its feature values ​​with a pre-defined medical reference database. The comparison process includes: extracting the values ​​of various physiological indicators from the synthetic sample and comparing these values ​​with the normal or pathological ranges of the corresponding indicators in the medical reference database to determine if there are any abnormal values ​​exceeding the physiologically possible range; extracting multiple features with dependencies from the synthetic sample and calculating whether the logical relationships between the features meet pre-defined physiological constraints, such as the vancomycin trough concentration... Furthermore, given the KDIGO grade of 3, this sample was deleted because low concentrations typically do not cause severe kidney damage; if the serum IL-6 concentration before vancomycin administration was... pg / mL, not in ICU, trough concentration However, samples labeled as KDIGO level 2 or 3 are deleted. The KDIGO level label and serum creatinine change value of the synthetic samples are extracted, and the theoretical KDIGO level of the sample is calculated according to the criteria defined in the KDIGO guidelines. The consistency between the theoretical KDIGO level and the label is determined, and synthetic samples that do not meet any of the above comparison conditions are deleted. Clinical experts conduct random sampling reviews of the synthetic samples that have passed automatic validation. 10% of the synthetic samples are randomly selected, and two clinical pharmacists independently assess their clinical rationality. If the proportion of unreasonable samples exceeds 5%, the CTGAN is retrained.

[0036] The validated synthetic samples are merged with real samples to construct an expanded training dataset. ,in To maximize the total number of samples after expansion, the proportion of synthetic samples is controlled between 50% and 60%. A class-weighted loss function is designed to further enhance the model's learning ability for the minority class, specifically for KDIGO ranking. The weighting coefficient is ,in For level The number of samples in the training set.

[0037] By employing conditional generative adversarial networks for data augmentation, this invention can learn the multidimensional joint probability distribution of real data and generate synthetic samples that are highly consistent with real data in terms of statistical properties and feature correlation. At the same time, the medical rationality of the generated samples is ensured through clinical rule filtering, which effectively solves the problem of insufficient minority class samples in clinical data and provides a balanced data foundation for subsequent model training.

[0038] In one embodiment of the present invention, step S3 specifically includes: S31. LASSO regression is used to select features from the expanded training set. Feature selection is achieved by introducing an L1 regularization term into the loss function.

[0039] LASSO achieves feature selection by introducing an L1 regularization term into the loss function, and its optimization objective is: in The number of training samples. For the first The labels of each sample (for a four-class classification problem, one-vs-rest processing is performed for each category separately), For the first The feature vector of each sample For the regression coefficient vector, These are the regularization parameters. The L1 regularization term is used. The introduction of this feature shrinks the coefficients of some features to 0, thereby achieving feature selection.

[0040] S32. Use cross-validation to determine the optimal regularization parameter. Divide the training set into several folds, select one fold as the validation set in turn, and use the remaining folds as training subsets. Fit the model on the training subset for the candidate regularization parameter values ​​and calculate the prediction error on the validation set. Select the regularization parameter value that minimizes the average validation error as the optimal regularization parameter.

[0041] S33. Using the optimal regularization parameters, fit the minimum absolute shrinkage and selection operator model on the entire training set to obtain the regression coefficient vector. Select the features with non-zero regression coefficients, while forcibly retaining the vancomycin valley concentration, the serum IL-6 concentration before vancomycin administration, and the ICU admission status to obtain the filtered feature set.

[0042] To ensure the clinical interpretability of the model and the reliability of decision-making, three core risk factors are mandatory to be retained, namely vancomycin trough concentration. Pre-treatment serum IL-6 concentration and ICU admission status Even if their LASSO coefficients are small, they are not excluded. In one specific embodiment, these three factors were identified as independent risk factors for vancomycin-related AKI by multivariate logistic regression analysis, with significant statistical significance and a clear pathophysiological mechanism. Specifically, as... Figure 3 As shown, the risk of AKI was significantly increased when the vancomycin trough concentration was greater than 18.00 μg / mL (OR=24.784, 95%CI: 5.127-119.870), indicating that drug exposure is a direct indicator of nephrotoxicity; Figure 4 As shown, a pre-vancomycin serum IL-6 concentration ≥133.85 pg / mL increased the risk of AKI by more than 7 times (OR=7.144, 95%CI: 2.032-25.114), reflecting that a high level of inflammation exacerbates the nephrotoxicity of vancomycin. Patients admitted to the ICU had a 10-fold increased risk of AKI (OR=10.197, 95%CI: 2.783-37.364), as ICU patients often have high-risk factors such as hemodynamic instability and multiple organ dysfunction. The final selected feature set is denoted as... ,in To retain the number of features, It includes three core risk factors.

[0043] During the feature selection phase, LASSO regression determined the optimal regularization parameter λ using five-fold cross-validation. The mean squared error of the cross-validation was minimized when λ was 0.05. Under this parameter, LASSO regression selected 12 features with non-zero coefficients from the initial 32 features, including age, baseline serum creatinine, duration of treatment, cumulative dose, serum albumin, use of vasoactive drugs, history of diabetes, mechanical ventilation, procalcitonin, C-reactive protein, use of loop diuretics, and use of aminoglycoside antibiotics.

[0044] It is worth noting that the coefficient for vancomycin trough concentration in the LASSO regression was 0.18 (non-zero), the coefficient for serum interleukin-6 was 0.12 (non-zero), and the coefficient for ICU admission status was 0 (compressed to zero). Although ICU admission status was excluded by LASSO, it was forcibly retained in this invention based on the aforementioned medical evidence. The final feature set contains 15 features, namely the 12 features selected by LASSO plus the 3 core factors forcibly retained (with duplicates removed).

[0045] By comparing the performance of models using only LASSO to select features (12) and those using the full feature set (15, including the core factors that are forcibly retained), the macro-average AUC of the latter is 0.729, while that of the former is 0.701, demonstrating the effectiveness of the strategy of forcibly retaining core factors.

[0046] By integrating pure data-driven LASSO feature selection with a mandatory retention mechanism of clinical prior knowledge, this invention reduces feature dimensionality and avoids overfitting while ensuring that key clinical risk factors inevitably participate in model prediction, thereby improving the clinical interpretability and decision reliability of the model.

[0047] In one embodiment of the present invention, such as Figure 5 As shown, step S4 specifically includes: S41. Select multiple heterogeneous base learners as the first layer, including support vector machines, decision trees, random forests, and multinomial logistic regression.

[0048] Support Vector Machines use radial basis function kernels, the kernel function of which is in the form of: ,in For kernel parameters, a one-to-many strategy is adopted for multi-class classification problems, and the penalty parameter is optimized through grid search. and kernel parameters .

[0049] The decision tree uses the CART algorithm, employing Gini impurity as the splitting criterion, and nodes... Gini impurity is defined as ,in For nodes KDIGO Level The sample proportion is set with hyperparameters set to maximum depth 5, minimum leaf node sample count 10, and minimum split sample count 20.

[0050] Random forests consist of 100 decision trees. They improve generalization by using a bagging ensemble strategy and random feature selection, randomly selecting features at each node during splitting. These features were selected as candidate splitting features.

[0051] Multinomial logistic regression for KDIGO levels The predicted log odds for this level are The logarithmic odds are converted into probabilities using the softmax function. L2 regularization is used to prevent overfitting.

[0052] S42. Use cross-validation to generate meta-features, divide the training set into several folds, train base learners using the remaining folds for each fold and predict the samples in that fold, and concatenate the outputs of each base learner into a meta-feature vector.

[0053] training set Randomly divided into 5 equal-sized folds For the first One fold, use the remaining four folds. Train four base learners, then apply them to folding. The samples in the training set are used for prediction. This process is repeated 5 times so that each sample in the training set receives a prediction result based on cross-validation. Four base learners generate four prediction probability vectors. For each sample... , No. Each base learner ( Output the four-class probability vector ,in as a base learner Predicted Samples Belongs to KDIGO level The probability satisfies The outputs of the four base learners are concatenated to form a meta-feature vector. .

[0054] S43. A class-weighted multinomial logistic regression is used as the meta-learner for the second layer. The input is the meta-feature vector and the output is the KDIGO hierarchical probability distribution. The model is trained by minimizing the class-weighted cross-entropy loss function to obtain a multi-class ensemble model.

[0055] The model form of the meta-learner is ,in Let $k$ be the weight parameter vector for rank $k$. The meta-learner is trained by minimizing the improved class-weighted cross-entropy loss function: in The number of training samples. For the sample Authentic KDIGO rating labels For level The corresponding weighting coefficients, Predicting samples for meta-learners Belongs to the level The probability, For prediction level Compared to the actual level The distance between them This represents the rank distance penalty coefficient. The category weight coefficient is also mentioned. ,in The total number of samples in the training set. For level The number of samples in the training set.

[0056] This loss function introduces a class distance penalty term on top of the traditional class-weighted cross-entropy. The purpose of this penalty is to: for true labels... For samples with a true rank of 0, the model will be penalized more severely if it assigns a higher probability to the incorrect category that is further away from the true rank. For example, for a sample with a true rank of 0, if the model incorrectly predicts it as rank 3, the penalty will be greater than if it incorrectly predicts it as rank 1. This design aligns with the ordered nature of the KDIGO classification system, guiding the model to predict neighboring ranks rather than ranks with larger gaps when classification errors occur, thus reducing more serious misclassifications in clinical practice. (Rank distance penalty coefficient) Cross-validation determined that the typical value range is 0.1 to 0.5. Through this improved loss function, the present invention, while maintaining sensitivity to minority classes, fully utilizes the ordered nature of KDIGO ranks, thereby improving the accuracy and clinical applicability of multi-class prediction.

[0057] S44. For the test set samples, use the trained base learner to predict the probability vectors respectively and concatenate them into a meta-feature vector. Input the meta-feature vector into the trained meta-learner to obtain the KDIGO level probability distribution, and select the level with the highest probability as the KDIGO level prediction result.

[0058] By constructing a multi-class ensemble model based on the Stacking strategy, this invention can integrate data features captured from different perspectives by multiple heterogeneous base learners, and automatically learn the optimal fusion weights through a meta-learner, thereby improving the prediction accuracy and robustness of KDIGO levels.

[0059] In one embodiment of the present invention, such as Figure 6 As shown, step S5 specifically includes: S51. Perform interpretability analysis on the multi-class ensemble model, and use the SHAP method to calculate the feature contribution value for each KDIGO level. For samples and levels, the ... The SHAP value of a feature represents the contribution of that feature to the model's prediction of that level.

[0060] For multi-class classification tasks, SHAP calculates the feature contribution value for each KDIGO level separately. For each sample... and level No. SHAP value of each feature This indicates that the feature is predicted as a grade by the model. The contribution of SHAP value. The theoretical definition of SHAP value is: in For feature set, For features not included Feature subset, To use only a subset of features Time model for levels The predicted value.

[0061] S52. Calculate the SHAP value matrix of each base learner separately, and obtain the SHAP value of the ensemble model by weighting the fusion weights of each base learner in the meta-learner.

[0062] For the Each base learner and sample Calculate the SHAP matrix , of which elements Representation base learner Chinese characteristics For level The SHAP contribution value. Weighted fusion is performed based on the fusion weights of each base learner in the meta-learner. For levels... From meta-learner parameters Extracting the base learner The weights of the output probabilities are denoted as... Base learner In level The normalized weights on are The SHAP value of the ensemble model is obtained through a weighted average: .

[0063] S53. Extract the corresponding feature contribution values ​​for the predicted KDIGO level and sort them by absolute value to identify the feature that contributes the most to the prediction result.

[0064] For each sample in the test set, calculate its SHAP value matrix at the four KDIGO levels. The sign and magnitude of the SHAP value have a clear interpretive meaning: Representation of features Add model prediction samples belonging to the level The probability, This indicates that the feature reduces the probability of that level. The magnitude indicates the strength of the contribution. This applies to the predicted KDIGO level. Extract the SHAP value corresponding to this level. Sort by absolute value and identify the top 10 features that contribute the most to the prediction results.

[0065] S54. Generate a structured evaluation report. The report includes the following core content modules: The first module displays the KDIGO rank prediction results. It shows the KDIGO rank predicted by the model. and their corresponding prediction probabilities It also displays the complete probability distribution across four levels to help doctors understand the confidence level of the prediction. If the predicted probability is close to 50%, the report will indicate high uncertainty, and it is recommended to combine it with other clinical information for a comprehensive assessment.

[0066] The second module is the core risk factor status analysis. It displays the current values, clinical reference thresholds, and degrees of deviation for three core factors. For vancomycin trough concentration... Report its current measurement (unit: (g / mL), with a reference threshold of 18.0 g / mL. g / mL, calculate the degree of deviation .if If the trough concentration exceeds the recommended upper limit, the risk of nephrotoxicity increases; if Further indications suggest that the trough concentration was significantly exceeded, and immediate dose adjustment is recommended. Regarding the serum IL-6 concentration before vancomycin administration... (Unit: pg / mL), with a reference threshold of 133.85 pg / mL, calculate the degree of deviation. ,if Elevated inflammatory markers may increase the risk of kidney damage. This is relevant for ICU admission status. ,if In the ICU, the patient is marked as being in critical condition with multiple high-risk factors; if... The patient was not in the ICU and their overall condition was relatively stable.

[0067] The third module is SHAP feature contribution analysis, specifically for the predicted KDIGO levels. Display the SHAP contribution value of each feature. Sort the features by absolute SHAP value from largest to smallest, and list the top 10 features with the largest contributions. For each feature, report its current value, SHAP contribution, and direction of contribution. Highlight the SHAP values ​​of the three core factors, using different colors or markers to distinguish them from other features. For example, if the SHAP value of vancomycin trough concentration... This is interpreted as the valley concentration being correlated with the predicted KDIGO level. It has a significant positive contribution, but this factor increases the risk.

[0068] The fourth module provides personalized clinical intervention recommendations. Stratified management recommendations are generated based on the predicted KDIGO grade. If the predicted grade is 0, the patient is considered to have a low risk of AKI, and routine monitoring of serum creatinine (every 48-72 hours) is recommended, with attention focused on maintaining trough concentrations within the recommended range. If the predicted grade is 1, the patient is considered to have a risk of mild AKI, and close monitoring of serum creatinine (every 24-48 hours) is recommended, along with assessment of whether vancomycin dosage adjustment is necessary, increased hydration therapy, and avoidance of concomitant use of other nephrotoxic drugs. If the predicted grade is 2, the patient is considered to have a higher risk of moderate AKI, and daily monitoring of serum creatinine is recommended, along with consideration of reducing the vancomycin dosage or extending the dosing interval, switching to other antibiotics if necessary, and strengthening renal function protection measures. If the predicted grade is 3, the patient is considered to have a high risk of severe AKI, and immediate reduction or discontinuation of vancomycin and switching to other antibiotics is recommended, along with daily monitoring of serum creatinine and urine output, close monitoring of electrolyte and acid-base balance, initiation of renal replacement therapy if necessary, and consultation with a nephrologist. Personalized recommendations are generated by combining core factor status and SHAP contribution analysis. For example, if the trough concentration is significantly elevated and SHAP contributes the most, it is recommended to adjust the vancomycin dose first; if the inflammatory marker IL-6 is significantly elevated, it is recommended to actively control infection and inflammatory response; if the patient is in the ICU and is using vasoactive drugs, it is recommended to optimize hemodynamic management.

[0069] The report is presented in a visual format to improve the reading experience and decision-making efficiency for clinicians. A risk level dashboard uses different colors to represent four KDIGO levels (green for level 0, yellow for level 1, orange for level 2, and red for level 3), with a pointer indicating the predicted level and displaying the predicted probability. A SHAP feature contribution waterfall plot is created, with the horizontal axis representing the SHAP value and the vertical axis representing the feature name. Starting from the baseline value, the SHAP contribution of each feature is sequentially stacked to reach the predicted value, clearly showing how each feature gradually influences the prediction result. The status of core factors is displayed using radar charts or bar charts, providing a visual comparison of the deviation between the current value and the threshold.

[0070] By employing the SHAP method for interpretability analysis and generating structured assessment reports, this invention not only provides accurate predictions of KDIGO levels but also quantifies the specific contributions of each feature to the prediction results, particularly highlighting the role of core risk factors. This provides clinicians with transparent and credible decision support information, enhancing the acceptability and application value of the model in clinical practice.

[0071] To verify the contribution of each technical step of this invention to the model performance, the following ablation experiment was designed: Example 1: The complete technical solution of the present invention is adopted, including CTGAN data augmentation, Stacking ensemble learning and feature selection strategy of mandatory retention of core factors.

[0072] Comparative Example 1: Without CTGAN data augmentation, the original training set was used directly for model training, and the remaining steps remained unchanged.

[0073] Comparative Example 2: CTGAN data augmentation was used, but only a single base learner (random forest) was used for prediction, without using the Stacking ensemble learning strategy.

[0074] Comparative Example 3: CTGAN data augmentation and Stacking ensemble learning were used, but core factors were not forcibly retained during feature selection, and the automatic selection results of the LASSO algorithm were relied upon entirely.

[0075] The four schemes were evaluated using the same test set, and ROC curves (Receiver Operating Characteristic Curves) were plotted as follows: Figure 7 As shown, the ROC curve plots the false positive rate (FPR) on the x-axis and the true positive rate (TPR) on the y-axis. The area under the curve (AUC) is used to evaluate the model's classification performance. The AUC value ranges from 0 to 1; a value closer to 1 indicates better model performance. An AUC of 0.5 corresponds to the baseline level of random guessing.

[0076] Depend on Figure 7It can be seen that Example 1 exhibits the best ROC curve performance, with an AUC value of 0.729. This result indicates that the complete technical solution can identify all positive samples with a low false positive cost, demonstrating excellent clinical application value. Comparative Example 3 has an AUC value of 0.711. The comparison shows that forcibly retaining core clinical factors can improve the AUC value, verifying the necessity of combining domain knowledge with machine learning. Comparative Example 2 has an AUC value of 0.684, indicating that Stacking ensemble learning effectively improves the model's generalization ability by fusing the prediction results of multiple base learners. Comparative Example 1 has the lowest AUC value, only 0.621. The lack of CTGAN data augmentation led to a significant decrease in the AUC value, demonstrating the significant role of data augmentation technology in alleviating sample imbalance and improving model performance.

[0077] In summary, the ablation experiment results validated the effectiveness of each technical step in this invention: CTGAN data augmentation is a key foundation for improving model performance, Stacking ensemble learning further enhances model stability, and the forced retention of core clinical factors ensures the medical interpretability and practicality of the model.

[0078] In one embodiment of the present invention, the risk assessment system is implemented based on a client-server architecture. The server side deploys the model training and inference engine, and the client provides a user interface, including a data processing module, a model building module, and a risk assessment module.

[0079] The data processing module comprises a data acquisition unit, a labeling unit, and an augmentation unit. The data acquisition unit automatically extracts patient demographic characteristics, underlying disease information, vancomycin treatment parameters, laboratory test results, and clinical status information from the Hospital Information System (HIS) and stores the data in a structured database. The labeling unit automatically reads the patient's baseline serum creatinine and the highest serum creatinine value after treatment according to the KDIGO 2013 AKI diagnostic criteria, calculates the creatinine elevation, and outputs KDIGO grade labels. The augmentation unit calls a pre-trained CTGAN model to generate synthetic data for minority class samples in the training set. After validation with a clinical rule filter, this synthetic data is merged with real data to form an augmented training set. The data preprocessing unit performs Z-score standardization on continuous variables, one-hot encoding on categorical variables, handles missing values, and divides the training and test sets in a 7:3 ratio.

[0080] The model building module includes a feature selection unit and a model training unit. The feature selection unit uses the LASSO regression algorithm, employing 10-fold cross-validation to determine the optimal regularization parameters, selecting features with non-zero regression coefficients, while forcibly retaining three core risk factors: vancomycin trough concentration, pre-vancomycin serum IL-6 concentration, and ICU admission status. The model training unit constructs a Stacking multi-class ensemble model. The first layer contains four base learners: support vector machine, decision tree, random forest, and multinomial logistic regression, using 5-fold cross-validation to generate meta-features. The second layer uses class-weighted multinomial logistic regression as the meta-learner, and training is completed through an optimized and improved loss function. After training, the model parameters and weights are stored in a serialized format in the model library.

[0081] The risk assessment module comprises a data receiving unit, a prediction unit, and a report generation unit. The data receiving unit receives clinical data of the patient to be assessed from the physician via a web or mobile interface, performing format verification and data integrity checks. The prediction unit loads the pre-trained ensemble model, standardizes the input data, and extracts features. It then outputs the predicted probability distributions for four KDIGO levels through a base learner and a meta-learner, selecting the level with the highest probability as the prediction result. The report generation unit calls the SHAP interpreter to calculate the contribution value of each feature to the prediction result, sorts them by contribution, identifies key risk factors, and combines the prediction level, core factor status, and feature contribution analysis to generate a structured assessment report containing a risk level dashboard, a SHAP waterfall chart, and personalized clinical recommendations. This report is output in PDF or HTML format for physician review.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for risk assessment of vancomycin-associated acute kidney injury, characterized by, The method comprises the following steps: S1, collecting clinical data of patients receiving vancomycin treatment, marking the patients according to the KDIGO guidelines to obtain a real data set, preprocessing the real data set and dividing it into a training set and a test set; S2, using a conditional generative adversarial network to perform data augmentation on the training set samples, generating synthetic samples through adversarial training, and combining the synthetic samples with the training set to obtain an augmented training set after verification of the synthetic samples; S3, using LASSO regression to perform feature selection on the augmented training set, while retaining the vancomycin trough concentration, the serum IL-6 concentration before vancomycin administration and the ICU admission state as core risk factors, to obtain a filtered feature set; S4, constructing a Stacking multi-classification ensemble model based on the filtered feature set, the Stacking multi-classification ensemble model comprising a plurality of base learners in the first layer and a meta-learner in the second layer, training the multi-classification ensemble model to output a KDIGO classification prediction result; S5, using the SHAP method to perform explainability analysis on the prediction result of the multi-classification ensemble model, calculating the contribution value of each feature to the prediction result, and generating an evaluation report containing the KDIGO classification prediction result and the feature contribution value.

2. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 1, wherein, In step S1, the clinical data includes demographic characteristics, underlying disease information, vancomycin treatment parameters, laboratory examination indicators and clinical state information; the classification annotation includes KDIGO 0, KDIGO 1, KDIGO 2 and KDIGO 3.

3. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 1, wherein, Step S2 specifically comprises: S21, constructing a conditional generative adversarial network, the conditional generative adversarial network comprising a generator and a discriminator, the input of the generator being a random noise vector and a conditional label, and the output being a synthetic sample, the input of the discriminator being a real sample or a synthetic sample and its corresponding conditional label, and the output being the probability that the sample is real data; S22, training the conditional generative adversarial network on real samples, learning the multi-dimensional joint probability distribution of real data through adversarial training, generating a few-class sample using a conditional generation mechanism after training, and obtaining a synthetic sample through the generator using a random noise vector and a conditional label; S23, verifying the quality of the generated synthetic sample, evaluating the distribution consistency of the synthetic sample and the real sample on each continuous feature using statistical tests, calculating the feature correlation matrix of the real data and the synthetic data to evaluate the correlation difference, and deleting unreasonable synthetic samples through a clinical rule filter, and combining the verified synthetic samples with the training set to obtain an augmented training set.

4. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 3, wherein, In step S23, the clinical rule filter judges the rationality of the sample by comparing the feature values of the synthetic sample with a preset medical reference database, and the comparison process comprises: The values of the physiological indexes in the synthetic sample are extracted, the values are compared with the normal range or pathological range of the corresponding indexes in the medical reference database, and it is judged whether there is an abnormal value exceeding the physiological possible range; a plurality of features with a dependent relationship in the synthetic sample are extracted, and it is calculated whether the logical relationship between the features meets the preset physiological constraint condition; and the KDIGO grade label and the serum creatinine change value of the synthetic sample are extracted, the theoretical KDIGO grade of the sample is calculated according to the grade determination standard defined in the KDIGO guide, it is judged whether the theoretical KDIGO grade is consistent with the label, and the synthetic sample that does not satisfy any of the above comparison conditions is deleted.

5. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 1, wherein, Step S3 specifically includes: S31, using LASSO regression to perform feature screening on the expanded training set, and realizing feature selection by introducing an L1 regularization term in the loss function; S32, using cross-validation to determine the optimal regularization parameter, dividing the training set into several folds, and sequentially selecting one fold as the validation set and the remaining folds as the training subsets; for the candidate regularization parameter value, fitting the model on the training subsets and calculating the prediction error on the validation set, and selecting the regularization parameter value that minimizes the average validation error as the optimal regularization parameter; S33, using the optimal regularization parameter to fit the least absolute shrinkage and selection operator model on the entire training set to obtain a regression coefficient vector, selecting the features with non-zero regression coefficients, and simultaneously forcing to retain the vancomycin valley concentration, the serum IL-6 concentration before vancomycin administration, and the ICU admission state, to obtain a screened feature set.

6. A method for risk assessment of vancomycin-associated acute kidney injury as claimed in claim 5 wherein, The optimization objective of LASSO regression is: wherein is the number of training samples, is the label of the th sample, is the feature vector of the th sample, is the vector of regression coefficients, is the regularization parameter.

7. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 1, wherein, Step S4 specifically includes: S41, selecting a plurality of heterogeneous base learners as the first layer, the base learners including support vector machines, decision trees, random forests, and polynomial logistic regression; S42, using cross-validation to generate meta-features, dividing the training set into several folds, and for each fold, training the base learners using the remaining folds and predicting the samples in the fold, and splicing the outputs of each base learner into a meta-feature vector; S43, using a polynomial logistic regression with class weighting as the meta-learner of the second layer, inputting the meta-feature vector, and outputting the KDIGO grading probability distribution, training by minimizing the class-weighted cross-entropy loss function, and obtaining a multi-classification ensemble model; S44, for the test set samples, using the trained base learners to respectively predict and obtain probability vectors and splice them into a meta-feature vector, inputting the meta-feature vector into the trained meta-learner to obtain the KDIGO grade probability distribution, and selecting the grade with the maximum probability as the KDIGO grading prediction result.

8. A method for risk assessment of vancomycin-associated acute kidney injury according to claim 7, wherein, In step S43, the loss function of the meta-learner is: wherein is the number of training samples, is a sample with a true KDIGO stage label, is a meta-feature vector for a sample , is a stage corresponding weight coefficient, is a probability that a meta-learner predicts a sample belongs to a stage , is a distance between a predicted stage and a true stage , is a stage distance penalty coefficient; and the class weight coefficients are: wherein is the total number of samples in the training set, is the rank is the number of samples in the training set, and 4 is the total number of classes.

9. The method for risk assessment of vancomycin-associated acute kidney injury according to claim 1, wherein, Step S5 specifically includes: S51, performing explainability analysis on the multi-classification ensemble model, using the SHAP method to calculate the feature contribution value for each KDIGO grade, and the SHAP value of the i-th feature for the sample and the grade indicates the contribution of the feature to the model predicting the grade; S51, performing explainability analysis on the multi-classification ensemble model, using the SHAP method to calculate the feature contribution value for each KDIGO grade, and the SHAP value of the i-th feature for the sample and the grade indicates the contribution of the feature to the model predicting the grade; S52, calculating the SHAP value matrix of each base learner respectively, and weighting to obtain the SHAP value of the ensemble model according to the fusion weight of each base learner in the meta-learner; S53, extracting the corresponding feature contribution value for the predicted KDIGO grade and sorting according to the absolute value size to identify the feature with the largest contribution to the prediction result; S54, generating a structured evaluation report.

10. A system for risk assessment of vancomycin-associated acute kidney injury, characterized by, The system is used for executing the risk assessment method as claimed in any one of claims 1-9, comprising: a data processing module, configured to collect clinical data of patients, and perform labeling and preprocessing, and to perform expansion processing on training data with sample imbalance; a model construction module, configured to perform feature screening on the expanded training data, and to force retention of core risk factors, and to construct and train a multi-classification ensemble model based on the screened features; a risk assessment module, configured to receive clinical data of a patient to be assessed, to output a KDIGO grade prediction result through the trained multi-classification ensemble model, and to generate an assessment report containing the prediction result and feature contribution analysis based on an explainability analysis method.

Images