Design of a Vancomycin Clearance Prediction Scheme Based on VPB Combination Model

Abstract

This invention relates to the field of vancomycin pharmacokinetic technology, specifically to the design of a vancomycin clearance prediction scheme based on a VPB ensemble model. This method, based on an ensemble learning strategy, constructs a blending ensemble model to predict vancomycin clearance in adult Chinese patients. First, a variational autoencoder is used to amplify the original sample data to increase data diversity. Then, a particle swarm optimization algorithm is introduced to optimize the parameters of multiple base learners, and the prediction results of the optimized base learners are used as new feature inputs. Finally, a support vector regression machine is used as a meta-learner to integrate and model the above features, forming the final vancomycin clearance prediction model. The VPB model constructed in this invention achieves a determination coefficient R² exceeding 0.9 on both the test and training sets, demonstrating superior prediction accuracy compared to population pharmacokinetic models.

Description

Technical Field

[0001] This invention relates to the field of vancomycin pharmacokinetic technology, and in particular to the design of a vancomycin clearance prediction scheme based on the VPB combination model. Background Technology

[0002] Vancomycin is a broad-spectrum antibiotic primarily used to treat infections caused by Gram-positive bacteria, particularly methicillin-resistant Staphylococcus aureus (MRSA). With the widespread use of vancomycin, its pharmacokinetic characteristics and precise control of blood concentrations have become critical issues in clinical drug therapy. Especially in patients with impaired renal function, those in the intensive care unit (ICU), and those with other complications, vancomycin clearance can vary significantly, impacting efficacy and toxicity. Therefore, understanding and accurately predicting vancomycin clearance is crucial for clinical treatment.

[0003] Vancomycin clearance is typically predicted using individualized pharmacokinetic models that consider the absorption, distribution, metabolism, and excretion (ADME) processes of the drug. Most current studies use one-compartment or two-compartment models to fit the concentration-time curve of vancomycin in the blood. Due to the clinical importance of individualized vancomycin therapy, therapeutic drug monitoring (TDM) is often used in conjunction with factors such as patient weight and renal function to adjust the drug dosage using concentration-time data to achieve optimal therapeutic effects. By monitoring blood drug concentrations and predicting clearance, clinicians can adjust vancomycin dosing regimens in a timely manner.

[0004] With the development of computing technology, an increasing number of studies are using predictive modeling tools to estimate vancomycin clearance and optimize individualized dosing regimens for patients. These tools can combine clinical data with patients' pharmacokinetic characteristics to provide customized treatment plans. In recent years, machine learning and deep learning methods have also been applied to pharmacokinetic parameter estimation. Traditional pharmacokinetic models rely on empirical formulas and clinical data, while machine learning and deep learning methods can learn potential complex relationships from large amounts of data, providing more accurate predictions for individualized treatment. Accurate prediction of vancomycin clearance helps achieve individualized management of drug therapy, thereby improving treatment success rates.

[0005] Currently, most studies on machine learning models for predicting vancomycin clearance rates focus on children or newborns, where the literature is relatively abundant. However, research on adult patients is relatively limited. Therefore, this study aims to fill this gap by exploring and evaluating the application of machine learning models in adult patients, further improving the accuracy and clinical applicability of vancomycin clearance rate prediction. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a vancomycin clearance rate prediction scheme based on the VPB combination model to solve the problems existing in the prior art.

[0007] To achieve the above and other related objectives, this invention provides a vancomycin clearance prediction scheme based on a VPB ensemble model, characterized by combining ensemble model blending with population pharmacokinetic (PPK) principles to improve the prediction accuracy of vancomycin clearance in adult Chinese patients. The vancomycin clearance prediction scheme based on a VPB ensemble model consists of three parts.

[0008] The first part involves augmenting the original sample data using a variational autoencoder (VAE). By learning the latent distribution characteristics of the original data through a deep generative model, new samples that conform to the true distribution pattern are generated, effectively increasing the data scale and diversity, providing richer feature information for subsequent model training, and improving the generalization ability and robustness of the prediction model.

[0009] The second part is: introducing the particle swarm optimization (PSO) algorithm to globally optimize the key parameters of multiple heterogeneous base learners, significantly improving the prediction accuracy and stability of each base model, and splicing and fusing the predicted values ​​output by the optimized base learners as input to a new feature space, thereby constructing a more expressive secondary feature set.

[0010] The third part is: using a support vector regression machine as a meta-learner, taking the concatenated new feature matrix as input, integrating and modeling the output results of the base learner, and constructing the final vancomycin clearance rate integrated prediction model.

[0011] As described above, the vancomycin clearance prediction scheme based on a VPB combination model disclosed in this invention demonstrates excellent performance in predicting vancomycin clearance in adult patients, showing improvements in accuracy and stability compared to single models or traditional methods. This model can optimize drug treatment regimens, improving treatment safety and efficacy. With the accumulation of data and continuous algorithm optimization, this model will be more widely applied in clinical practice. Simultaneously, this invention also provides valuable reference for pharmacokinetic prediction studies of other drugs. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0013] Figure 1 This is a flowchart illustrating the design of a vancomycin clearance prediction scheme based on a VPB combination model, as exemplified by this invention.

[0014] Figure 2 This is a diagram of the VAE model structure in the design of a vancomycin clearance rate prediction scheme based on the VPB combination model, which is an example of the present invention.

[0015] Figure 3 This is a PSO algorithm framework diagram in the design of a vancomycin clearance rate prediction scheme based on the VPB combinatorial model, as an example of the present invention.

[0016] Figure 4 This is a comparison chart of the RMSE of the PPK model and the hybrid model in the design of a vancomycin clearance prediction scheme based on the VPB combination model, which is an example of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0018] refer to Figure 1 This invention proposes a vancomycin clearance rate prediction scheme based on a VPB combination model, comprising the following steps:

[0019] Step 1: Use variational autoencoders to perform unsupervised learning on 1500 simulated patient data sets, extract low-dimensional latent features related to drug clearance rate, generate 15000 samples, expand the data scale and diversity, and provide sufficient sample support for subsequent model training.

[0020] Step 2: Introduce the particle swarm optimization algorithm as a global optimization tool to efficiently search the hyperparameter space of the five base learners. Find the optimal parameter combination through iterative optimization to ensure that each model has high prediction accuracy and generalization ability.

[0021] Step 3: Train five base learners based on the optimized parameters to fully learn the complex patterns in the data. Then, combine the prediction results of each base model on the original data to form a new feature space, providing richer feature representation for secondary modeling.

[0022] Step 4: Using a Support Vector Regression (SVR) machine as the meta-learner, the concatenated new feature matrix is ​​used as input. The outputs of each base model are deeply fused through high-order nonlinear fitting to construct the final blending ensemble prediction model, further improving the overall prediction performance.

[0023] Step 5: On 1500 original datasets, the constructed blending model is compared and validated with the traditional PPK model provided in the literature to evaluate its prediction accuracy, generalization ability and clinical applicability in complex data environments, so as to provide more reliable support for personalized medication decisions.

[0024] refer to Figure 2 This paper demonstrates the VAE model structure in the design of the vancomycin clearance prediction scheme of the VPB combination model proposed in this invention.

[0025] The encoder's task is to map the input data (including patient characteristics such as gender, age, weight, and serum creatinine) to a latent space. The encoder processes the input data using a neural network, outputting the mean and variance of the latent variables, thus describing their probability distribution. Typically, these latent variables follow a standard normal distribution. The core objective of the encoder is to capture, as far as possible, the underlying patterns in the input data, laying the foundation for generating new data samples that conform to these patterns.

[0026] In the latent space of a VAE (Visual Envelope Architecture), a set of low-dimensional latent variables is obtained by learning to compress the input data. These variables can be regarded as the inherent hidden representation of the data. Unlike directly replicating the original data, VAE reveals the structural features behind the data through probabilistic models. Specifically, VAE uses the method of maximizing the lower bound of the data likelihood function to effectively infer the latent variables, enabling the latent space to accurately capture the key feature distributions of the original data.

[0027] The decoder then transforms the low-dimensional latent variables back into the original data format. Another neural network maps these latent variables back to a data space of the same dimension as the original input, thereby generating new samples. Using VAE technology, this invention can learn latent patterns from limited patient data and generate new samples highly similar to real data. This not only effectively solves the problem of insufficient data in medical research but also ensures that the generated data does not deviate from real clinical patterns by maintaining the structural characteristics of the original data. This provides high-quality and diverse data support for subsequent model training, significantly improving the accuracy and reliability of the predictive model.

[0028] refer to Figure 3 This paper presents the PSO algorithm framework diagram in the design of the vancomycin clearance rate prediction scheme based on the VPB combinatorial model proposed by this paper.

[0029] In the initialization phase of the PSO algorithm, a set of particles is first randomly generated in the feasible solution space. Each particle represents a combination of hyperparameters to be optimized (such as learning rate, tree model depth, number of neural network layers, etc.), and the position and velocity of each particle are randomly initialized. Simultaneously, the initial position of each particle is set as its individual optimal solution, laying the foundation for subsequent iterative optimization. The search process of all particles must satisfy the preset parameter value range constraints to ensure that the optimization process is carried out within a reasonable parameter space.

[0030] During the iterative optimization phase, the algorithm searches for the optimal solution by continuously updating the velocity and position of particles. Velocity updates comprehensively consider both the particle's historical best position and the global best position of the group, allowing particles to gradually approach the optimal parameter region under the combined guidance of individual experience and social cooperation. Position updates adjust the particle's position in the solution space based on the updated velocity, thereby generating new candidate parameter combinations. After each update, the algorithm re-evaluates the fitness of each particle (i.e., the model's predictive performance on the validation set) and updates the individual optimal solution and the global optimal solution accordingly.

[0031] Through multiple iterations, the particle swarm optimization (PSO) algorithm gradually converges to near the global optimum, ultimately outputting the hyperparameter combination that optimizes the base learner performance. Compared to traditional grid search or random search, the PSO algorithm has stronger global search capabilities and faster convergence speed, effectively avoiding getting trapped in local optima. In this invention, all five base learners optimized by PSO achieved optimal prediction performance, laying the foundation for subsequent ensemble model construction.

[0032] refer to Figure 4 This paper presents a comparison of the RMSE of the PPK model and the hybrid model in the design of the vancomycin clearance prediction scheme based on the VPB combination model proposed in this invention.

[0033] Across all sample batches, the RMSE value of the Blending model was consistently lower than that of the PPK model, and its prediction error remained at a low level as the number of samples increased, demonstrating higher accuracy and stability. In contrast, the RMSE of the PPK model fluctuated more significantly, with higher errors than the Blending model in multiple sample intervals. This reflects the susceptibility of the traditional PPK model to noise and outliers when facing data uncertainty. The Blending model, by integrating the prediction results of multiple base learners, balances bias and variance, reducing prediction volatility.

[0034] To quantitatively evaluate the predictive performance and goodness of fit of the model, this study calculates the following four error indices: mean square error, root mean square error, mean absolute error, and coefficient of determination.

[0035] 1. Mean Squared Error (MSE) is the average of the squares of the model's prediction errors. A smaller MSE value indicates higher model prediction accuracy. The MSE formula is:

[0036]

[0037] Where N is the number of samples; CL pred,i Let C be the predicted clearance rate for the i-th sample; Lobs,i Let be the observed clearance rate of the i-th sample.

[0038] 2. Root Mean Square Error (RMSE) measures the magnitude and variance of the prediction error. A smaller RMSE value indicates higher model prediction accuracy. The RMSE formula is:

[0039]

[0040] 3. Mean Absolute Error (MAE) directly reflects the average magnitude of prediction error. It is insensitive to outliers and is a direct indicator for evaluating prediction accuracy. The MAE formula is:

[0041]

[0042] 4. Coefficient of Determination (R²) 2 R is used to evaluate the proportion of data variation explained by the model. The closer the value is to 1, the stronger the model's ability to explain data variation. 2 The formula is:

[0043]

[0044] As shown in the table below, the prediction performance of each base learner model varies on the dataset, while the Blending ensemble model constructed in this invention demonstrates significant advantages in all evaluation metrics. Specifically, the coefficient of determination R0 of the Blending model on the training and test sets... 2 All values ​​exceeded 0.9, indicating strong data fitting ability and the ability to effectively capture the potential changes in vancomycin clearance rate. Regarding error metrics, the RMSE of the Blending model on the training and test sets were 0.6014 and 0.6468, respectively; the MSE was 0.3617 and 0.4183, respectively; and the MAE was 0.5154 and 0.5572, respectively. All error values ​​were significantly lower than other single-base models. These results demonstrate that the Blending model not only has higher prediction accuracy but also excellent generalization ability, maintaining stable and reliable predictive performance on data, providing more accurate data support for personalized clinical dosing decisions.

[0045]

[0046] Furthermore, using a dataset containing 1,500 original samples, the performance of the proposed hybrid model was compared with that of the vancomycin clearance (CL) prediction model based on PPK. As shown in the table below, the VPB model significantly improved upon the PPK model in all key pharmacokinetic parameters. The R² value of the VPB model reached 0.9660 (compared to 0.9125 for the PPK model). Error metrics all showed a decreasing trend: the root mean square error (RMSE) decreased by 37.6% (0.3503 vs. 0.5616), and the mean square error (MSE) decreased by 61.1% (0.1227 vs. 0.3154).

[0047]

[0048] In summary, the VPB model proposed in this invention demonstrates superior performance in predicting vancomycin clearance rates in adult Chinese patients, showing significant improvements in both prediction accuracy and stability compared to traditional PPK models and single machine learning models. This model can provide clinical pharmacists with more reliable support for individualized dosing decisions, helping to optimize vancomycin dosing regimens, reduce the risk of treatment failure or adverse reactions due to fluctuations in blood drug concentrations, and thus improve the safety and efficacy of clinical treatment. With the continuous accumulation of clinical data and iterative optimization of the algorithm, this model is expected to play a wider application role in individualized anti-infective therapy.

[0049] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A vancomycin clearance rate prediction scheme based on a VPB combinatorial model, characterized in that: A variational autoencoder is used to perform unsupervised learning on the original patient data, learn the latent distribution of the data, and generate simulated samples to expand the training dataset and solve the problem of insufficient sample size.

2. The vancomycin clearance prediction scheme based on the VPB combination model as described in claim 1, characterized in that: Based on the expanded dataset, a particle swarm optimization algorithm is introduced to perform global hyperparameter optimization on multiple base learners, including deep learning models and machine learning models, to capture the complex nonlinear relationship between vancomycin clearance rate and patient characteristics.

3. The vancomycin clearance prediction scheme based on the VPB combination model as described in claim 2, characterized in that: The prediction results of multiple optimized base learners are spliced ​​and fused to form a new feature space. The support vector regression machine is used as a meta-learner to fit the fused features and construct the final blending ensemble prediction model, namely the VPB model.

4. The vancomycin clearance prediction scheme based on the VPB combination model as described in claim 3, characterized in that: The VPB model was used to predict vancomycin clearance in adult Chinese patients. Its prediction accuracy and stability were superior to traditional population pharmacokinetic models and single machine learning models. The coefficient of determination R2 on both the training and test sets exceeded 0.9, and the error index was reduced.

5. The vancomycin clearance prediction scheme based on the VPB combination model as described in claim 4, characterized in that: The VPB model is applied to clinical personalized dosing decision support. By inputting patient characteristic data, it outputs a vancomycin clearance rate prediction value, which is used to optimize the vancomycin dosing regimen and improve the safety and effectiveness of treatment.

Images