A prediction model for valproic acid trough concentration in adults based on machine learning and its application

Abstract

The present invention relates to a prediction model for the valproic acid (VPA) trough concentration in adults based on machine learning, and the prediction model construction method comprises the following steps: S1, data collection and cleaning: using k-nearest neighbor interpolation to fill missing values; S2, model establishment and verification: a total of 12 algorithms were established in the derivative cohort; S3, model interpretation: SHAP value, SHAP summary diagram, global importance ranking diagram and SHAP dependency diagram were used for interpretation; S4, model optimization: it was improved by feature selection, and the influence of these included features on the prediction performance of the model was compared using the mean absolute error (MAE) value. The present invention can predict the trough concentration of VPA by inputting nine features (VPA daily dose, last dose, uric acid (UA), platelets (PLT), combined medication, gender, weight, albumin (ALB) and aspartate aminotransferase (AST)) into the model. The simplified model proposed by the present invention can effectively predict the trough concentration of VPA in adults and can guide clinicians to choose the best medication regimen.

Description

Technical Field

[0001] The present invention relates to the technical field of medical model data processing, and in particular to a machine learning-based prediction model for adult valproic acid trough concentration and its application. Background Art

[0002] Valproic acid is a broad-spectrum anti-epileptic drug. When epileptic symptoms appear, valproic acid can be taken under the guidance of a doctor to avoid taking the drug privately to avoid adverse drug reactions. VPA mainly exerts its anti-epileptic effect by promoting the synthesis and release of γ-aminobutyric acid (GABA), a neurotransmitter that inhibits the activity of the central nervous system. However, the blood concentration range of VPA is usually narrow (50-100 mg / L) and shows considerable individual differences in pharmacokinetics. When the VPA level is >125 mg / L, the incidence of adverse reactions in patients increases, including weight gain, diarrhea, drowsiness, nausea, thrombocytopenia, vomiting, and leukopenia. Pharmacokinetic variability may be affected by factors such as age, weight, sex, pregnancy, daily dose, hepatic and renal insufficiency, combined use of other anti-epileptic drugs, and low albumin levels. These factors lead to different serum VPA concentrations in patients taking the same dose. Given the significant fluctuations in VPA concentrations and the limited therapeutic range, therapeutic drug monitoring (TDM) helps to maximize efficacy and prevent side effects. Currently, TDM for VPA involves blood drawing, which can be painful and uncomfortable for patients. In addition, due to limited medical resources, some hospitals lack TDM platforms, and the sampling and testing process of TDM is time-consuming and costly. In order to reduce patient discomfort and healthcare costs, advanced tools are needed to more effectively predict drug concentrations.

[0003] Chen et al. [1] used the optimal random forest (RF) algorithm to construct a regression model to predict the trough concentration of vancomycin in adult patients with periprosthetic joint infection. Zhu et al. [2] used 15 machine learning (ML) algorithms to develop a model for predicting the dose-adjusted concentration of lamotrigine and found that the ET algorithm performed better than other algorithms. The above studies all used machine learning methods to predict blood drug concentrations and achieved good prediction accuracy. However, there is a lack of model simplification solutions.

[0004] Chinese patent document CN117079743A provides a statin treatment effect prediction model and application, the prediction model includes: low-density lipoprotein LDL standard prediction model, liver enzyme abnormality prediction model, myalgia / creatine kinase abnormality prediction model; the construction method includes: collecting and obtaining clinical data of patients taking statins, performing data pre-screening, data filling, data sampling and feature screening processing, and determining the variables included in the model; the three models are respectively trained using logistic regression, naive Bayes, decision tree, random forest, and gradient boosting decision tree algorithms, and the optimal model is screened for modeling; model verification; sample size verification; using SHAP to explain the model; and the advantages of the optimal model are screened out through model decision curve analysis DCA. The present invention provides a monitoring method based on a prediction model for the rational use of statins, which can provide additional monitoring plans for patients with high risk of adverse reactions, and promote the clinical rational application of statins.

[0005] However, there are no reports on a machine learning-based prediction model for valproate trough concentration in adults and its application. Summary of the invention

[0006] The purpose of the present invention is to provide a machine learning-based prediction model for valproic acid trough concentration in adults and its application in view of the deficiencies in the prior art.

[0007] In a first aspect, the present invention provides a prediction model for valproic acid trough concentration in adults based on machine learning, wherein the method for constructing the prediction model comprises the following steps:

[0008] S1. Data collection and cleaning: k-nearest neighbor interpolation was used to fill missing values;

[0009] S2. Model establishment and validation: A total of 12 algorithms were established in the derivative cohort, including four linear models: linear regression, ridge regression, least absolute shrinkage and selection operator, and elastic network, and eight nonlinear models: guided aggregation, adaptive boosting, extreme gradient boosting, category boosting, maximal tree, random forest, decision tree, and gradient boosting regression;

[0010] S3, model explanation: SHAP value, SHAP summary graph, global importance ranking graph and SHAP dependency graph were used for explanation;

[0011] S4. Model optimization: It was improved through feature selection. According to the SHAP global importance ranking chart, the features were added to the model in descending order of their importance. Then, the MAE value was used to compare the impact of these included features on the model prediction performance.

[0012] As a preferred example, the step S1: VPA data set includes demographic information, basic disease information, VPA administration information, combined medication information within 5 days before VPA blood concentration monitoring, laboratory parameters within one week before VPA blood concentration monitoring, and VPA trough concentration.

[0013] As another preferred example, the values ​​are filled by calculating the average value of a given variable from the five nearest neighbors of a given subject. After the data filling is completed, the entire dataset is randomly divided into a derivation cohort and a validation cohort in a ratio of 8:2.

[0014] As another preferred example, the step S2: uses 10-fold cross validation to evaluate the model, and then uses the mean absolute error value in the test set to select the best model.

[0015] As another preferred example, in step S2: based on the 10-fold cross-validation method, the ET algorithm obtained the lowest MAE value in the test set, and the lowest MAE value was 13.08 (10.83-15.32) mg / L. The tree-based integrated algorithm ET was selected to establish a prediction model.

[0016] As another preferred example, the step S2: MAE, RMSE, R 2 , relative accuracy and absolute accuracy were applied to the validation cohort.

[0017] As another preferred example, step S4, model optimization: nine features are selected, namely daily dose, last dose, UA, PLT, concomitant medication, gender, weight, ALB and AST, and the ET model incorporating the nine selected features is compared with the ET model incorporating all feature calculations.

[0018] In the second aspect, the present invention provides an application of a prediction model for the trough concentration of valproic acid in adults based on machine learning. Nine features are input into the model to predict the trough concentration of valproic acid. The nine features are daily dose, last dose, UA, PLT, combination medication, gender, weight, ALB and AST.

[0019] In a third aspect, the present invention provides an information data processing terminal, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the adult valproate trough concentration prediction model based on machine learning.

[0020] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the adult valproate trough concentration prediction model based on machine learning.

[0021] The advantages of the present invention are: the present invention uses SHAP analysis to explore the potential connection between input features and VPA trough concentrations, and screens features. The number of features is reduced from 27 to 9, which improves the accuracy of prediction while reducing the number of model input features. The present invention can predict the trough concentration of valproic acid by inputting nine features (daily dose of valproic acid, last dose, uric acid (UA), platelets (PLT), combined medication, gender, weight, albumin (ALB) and aspartate aminotransferase (AST)) into the model. The simplified model involved in the present invention can effectively predict the trough concentration of valproic acid in adults and can guide clinicians to choose the best medication regimen. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Workflow of data collection, data processing, algorithm selection, and model development.

[0023] Figure 2 The explanation of the model uses Shapley additive explanation (SHAP). (a) SHAP summary plot. Each point represents the SHAP value of a feature instance. Its position on the plot is determined by the SHAP value on the x-axis and the features on the y-axis (ordered by feature importance). The color of each point indicates the value of the feature, with red representing high values ​​and blue representing low values. A higher SHAP value for a feature is associated with a higher VPA trough concentration. (b) A plot of the global importance of each feature ranked by its mean (|SHAP value|).

[0024] Figure 3 The composition of the feature set (i.e., the first 9 features) is selected to optimize the ET model. The green dots represent the MAE values ​​calculated after adding different numbers of features to the ET model.

[0025] Figure 4 Performance comparison of the full variable model (including 27 variables) and the optimized model (including the first 9 variables). MAE: mean absolute error; RMSE: root mean square error; relative accuracy: the error between the predicted value and the measured value is within 20%; absolute accuracy, the difference between the predicted value and the measured value is within 20 mg / L. DETAILED DESCRIPTION

[0026] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0027] Example 1

[0028] Please see Figure 1:The present invention proposes a method for predicting the trough concentration of valproic acid in adults based on machine learning. The method collects demographic information, basic disease information, medication information, combined medication information and patient plasma samples of adult patients using valproic acid, and the K nearest neighbor algorithm (KNN) is used to fill in missing values. Then, the collected data is divided into a derivative cohort and a validation cohort in a ratio of 8:2. 12 machine learning algorithms are used to model in the derivative cohort, and 10-fold cross validation is used to screen the optimal model Extratree according to the mean absolute error (MAE). The covariates are directly fitted to the optimal model, the parameters are re-estimated, a new model is established, and the SHAP method is used for model interpretation. Then, the features are sorted and screened according to their importance, and the top 9 features are selected for modeling to obtain a simplified model. Finally, the simplified model is evaluated using a variety of evaluation indicators. The simplified model involved in the present invention can effectively predict the trough concentration of valproic acid in adults and can guide clinicians to choose the best medication regimen.

[0029] The present invention inputs nine characteristics (daily dose of valproic acid, last dose, uric acid (UA), platelets (PLT), combined medication, gender, body weight, albumin (ALB) and aspartate aminotransferase (AST)) into the model to predict the valproic acid trough concentration.

[0030] 1 Materials and methods

[0031] 1.1 Study population

[0032] A single-center retrospective study was conducted on Chinese adult patients with epilepsy who were treated with valproic acid (VPA) at Jinshan Hospital Affiliated to Fudan University from January 2022 to December 2023. Inclusion criteria were as follows: (1) diagnosed with epilepsy according to the International League Against Epilepsy (ILAE); (2) treated with VPA tablets; (3) aged >18 years; (4) VPA reached a steady-state blood concentration (treatment ≥3 days); (5) VPA trough concentration was monitored before the next dose. Exclusion criteria were as follows: (1) pregnant and lactating women; (2) data missing more than 80%.

[0033] 1.2 Data Collection and Cleaning

[0034] All data were collected using the hospital electronic medical record system and recorded using Excel spreadsheets. The VPA data set included demographic information (sex, age, weight, height, smoking history and drinking history), basic disease information (liver and kidney disease, diabetes, hypertension, cardiovascular and cerebrovascular disease), VPA administration information (administration route, daily dose, last dose and interval time), combined medication information within 5 days before VPA blood concentration monitoring (such as carbamazepine, oxcarbazepine, levetiracetam, lamotrigine, phenobarbital, phenytoin sodium, clonazepam, diazepam, lorazepam, topiramate), VPA blood concentration, and VPA blood concentration. Laboratory parameters [platelet (PLT), white blood cells (WBC), hemoglobin (HB), neutrophil (NE), red blood cell count (RBC), serum creatinine (Scr), uric acid (UA), albumin (ALB), total protein (TP), Na, alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), r-glutamyl transferase (r-GT), creatinine clearance (CLcr)] and VPA trough concentration within the previous week were monitored.

[0035] This study used VPA trough concentration as the target variable. According to statistics, the data missing rate was 2.56%. Therefore, we used k-nearest neighbor (k-NN) interpolation to fill the missing values. Specifically, we calculated the mean of a given variable from the five nearest neighbors of a given subject to fill these values. After the data filling was completed, the entire dataset was randomly divided into a derivation cohort and a validation cohort in a ratio of 8:2.

[0036] 1.3 Model establishment and verification

[0037] A total of 12 algorithms were established in the derivation cohort, including four linear models: linear regression (LR), ridge regression (RidgeR), least absolute shrinkage and selection operator (Lasso), elastic net (EN), and eight nonlinear models: bootstrapped aggregating (Bagging), adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), extratree (ET), random forest (RF), decision tree (DT), and gradient boosted regression (GBR). Due to the limited size of our dataset, we used 10-fold cross validation to evaluate the models, aiming to avoid bias and improve the generalization ability of the models. Then, we used the mean absolute error (MAE) in the test set to select the best model.

[0038] 1.4 Model Explanation

[0039] SHapley Additive exPlanations (SHAP) is a game-theoretic explanation method that can intuitively explain the output of machine learning (ML) models. It explains model predictions by calculating the SHAP value of each feature in each sample, and can evaluate which input variables will have a negative or positive impact on the output of the ML model. In order to understand the impact of individual variables on our choice of model, we use SHAP values, SHAP summary plots, global importance ranking plots, and SHAP dependency plots for explanation.

[0040] 1.5 Model Optimization

[0041] To create a more optimized model, we improved it through feature selection. Feature selection is a method to identify relevant variables for regression by eliminating redundant or irrelevant variables. An optimal feature set should contain the least number of variables while achieving the best predictive performance. According to the SHAP global importance ranking graph, features are added to the model in descending order of their importance. Then, the MAE value is used to compare the impact of these included features on the model's predictive performance.

[0042] 2 Results

[0043] 2.1 Model selection

[0044] Four linear algorithms and eight nonlinear algorithms were selected for VPA trough concentration prediction. Overall, the nonlinear algorithms outperformed the linear algorithms. Based on the 10-fold cross-validation method, the ET algorithm obtained the lowest MAE value [13.08 (10.83-15.32) mg / L] in the test set. Therefore, the tree-based ensemble algorithm ET was selected to establish the prediction model.

[0045] 2.2 Model construction and verification

[0046] We constructed a prediction model for VPA trough concentration using the ET algorithm on the derivation cohort. To evaluate the comprehensive prediction performance of the selected model, MAE, RMSE, R 2 , relative accuracy and absolute accuracy were applied to the validation cohort. The R 2 The value was 0.703, indicating that our method effectively explained the inter-individual concentration variation of VPA. Considering the prediction error, the MAE and RMSE values ​​were used to evaluate the prediction accuracy, and the results were 7.61 and 9.91, respectively. The relative accuracy (±20%) of the ET model was 57.14%, and the absolute accuracy (±20 mg / L) was 90.48%.

[0047] 2.3 Model interpretation and optimization

[0048] SHAP analysis was performed on the ET model to explore the potential associations between input features and VPA trough concentrations. Figure 2 a shows the SHAP summary plot of all features that affect the ET model. Each point corresponds to a row of data in the original dataset. On the y-axis, the features are ranked according to the sum of the SHAP values ​​across all samples. Figure 2 b shows the global importance of each feature. To determine the optimal feature set, we determined the point in the validation cohort where the MAE value was the smallest when adding features to the ET model. As expected, as features were added to the ET model from high to low, the MAE value gradually decreased to the lowest point and then fluctuated slightly. Figure 3The feature selection process is demonstrated, showing that the MAE value reaches the minimum when the first 9 features are included. These features include daily dose, last dose, UA, PLT, concomitant medication, gender, weight, ALB, and AST. Finally, we compared the ET model that included the 9 selected features with the ET model that included all features in the calculation, as shown in Figure 2. Figure 4 The results show that the model using 9 features has a better prediction effect (MAE = 6.82, RMSE = 9.62, R 2 =0.720, relative accuracy is 61.90%, absolute accuracy is 90.48%)

[0049] References

[0050] [1]CHEN YW,LIN XK,HUANG C,et al.Vancomycin trough concentration inadult patients with periprosthetic joint infection:a machine learning-basedcovariate model[J].Br J Clin Pharmacol,2024.

[0051] [2]ZHU

[0052] The above is only a preferred embodiment of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and supplements without departing from the method of the present invention. These improvements and supplements should also be regarded as the scope of protection of the present invention.

Claims

1. A prediction model for VPA trough concentration in adults based on machine learning, characterized in that: The method for constructing the prediction model comprises the following steps: S1. Data collection and cleaning: k-nearest neighbor interpolation was used to fill missing values; S2. Model establishment and validation: A total of 12 algorithms were established in the derivative cohort, including four linear models: linear regression, ridge regression, least absolute shrinkage and selection operator, and elastic network, and eight nonlinear models: guided aggregation, adaptive boosting, extreme gradient boosting, category boosting, maximal tree, random forest, decision tree, and gradient boosting regression; S3, model explanation: SHAP value, SHAP summary graph, global importance ranking graph and SHAP dependency graph were used for explanation; S4. Model optimization: It was improved through feature selection. According to the SHAP global importance ranking chart, the features were added to the model in descending order of their importance. Then, the MAE value was used to compare the impact of these included features on the model prediction performance.

2. The adult VPA trough concentration prediction model according to claim 1, characterized in that: The step S1: VPA data set includes demographic information, basic disease information, VPA administration information, combined medication information within 5 days before VPA blood concentration monitoring, laboratory parameters within one week before VPA blood concentration monitoring, and VPA trough concentration.

3. The adult VPA trough concentration prediction model according to claim 2, characterized in that: First, the Euclidean distance is selected to measure the similarity between samples. Then, for each missing sample, its distance from other samples in the data set is calculated, and the five most similar samples are selected as neighbors. Finally, the average value of the neighbor samples is used to fill these values. After the data is filled, the entire data set is randomly divided into a derivation cohort and a validation cohort in a ratio of 8:

2.

4. The adult VPA trough concentration prediction model according to claim 1, characterized in that: The step S2 described above: Use 10-fold cross validation to evaluate the model, and then use the MAE value in the test set to select the best model.

5. The adult VPA trough concentration prediction model according to claim 4, characterized in that: The step S2: Based on the 10-fold cross validation method, the maximal tree algorithm obtained the lowest MAE value in the test set, and the lowest MAE value was 13.08 (10.83-15.32) mg / L. The tree-based integrated algorithm ET was selected to establish a prediction model.

6. The adult VPA trough concentration prediction model according to claim 1, characterized in that: Step S2: MAE, RMSE, R 2 , relative accuracy and absolute accuracy were applied to the validation cohort.

7. The adult VPA trough concentration prediction model according to claim 1, characterized in that: Step S4, model optimization: Nine features were selected, namely daily dose, last dose, UA, PLT, concomitant medication, gender, body weight, ALB and AST. The ET model incorporating the nine selected features was compared with the ET model incorporating all features.

8. The use of the adult VPA trough concentration prediction model based on machine learning according to any one of claims 1 to 7, characterized in that: The trough concentration of valproic acid can be predicted by inputting nine features into the model, wherein the nine features are daily dose, last dose, UA, PLT, concomitant medication, gender, weight, ALB and AST.

9. An information data processing terminal, characterized in that: The information data processing terminal includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the machine learning-based adult valproate trough concentration prediction model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes the steps of the adult valproate trough concentration prediction model based on machine learning as described in any one of claims 1-7.

Images