Auxiliary generation method and system for individual medication blood concentration prediction algorithm

Abstract

The invention discloses an auxiliary generation method and system for an individualized medication blood concentration prediction algorithm, and the method comprises the steps: firstly extracting model parameters of an existing medication blood concentration prediction model, and carrying out the feature coding of the model parameters of each medication blood concentration prediction model through a machine learning technology; the method comprises the following steps of: firstly, performing multi-dimensional energy potential analysis on parameter information of each medication blood concentration prediction model to dynamically aggregate kernel characteristics of the model parameters so as to learn a kernel distribution mode of the model parameters, and further performing model parameter reduction on the basis of the kernel distribution mode. And generating a new medication blood concentration prediction model. Through the intelligent model generation mode, the complexity of model construction can be remarkably reduced, the accuracy of individualized medication blood concentration prediction can be improved, and more scientific and accurate medication guidance is provided for clinicians.

Description

Technical Field

[0001] The present application relates to the field of intelligent prediction, and more specifically, to an auxiliary generation method and system for an individualized medication blood concentration prediction algorithm. Background Art

[0002] In clinical treatment, the blood concentration of a drug is an important factor in determining its efficacy and safety. However, due to changes in the patient's physiological and pathological characteristics, the pharmacokinetic characteristics of the drug may change compared to healthy volunteers, resulting in poor efficacy in some patients due to too low blood concentrations, and adverse reactions in some patients due to too high blood concentrations. Therefore, the standard dosing regimen recommended in the instructions is not necessarily suitable for clinical patients. In clinical practice, it is necessary to further monitor blood drug concentrations to guide clinicians to develop individualized drug dosing regimens based on patient characteristics.

[0003] The prediction model of blood drug concentration after medication refers to a model that predicts the concentration changes of drugs in patients based on the pharmacokinetic properties of drugs through mathematical and statistical methods. Specifically, population pharmacokinetics, by establishing a mathematical model for individualized medication, can calculate the numerical distribution of appropriate dosage by inputting the patient's physiological indicators. Doctors can use the characteristic values ​​of this numerical distribution (such as expected values) to achieve differentiated and precise medication for patients.

[0004] However, the population pharmacokinetic model currently used in practice also has some problems, such as: 1. In the model, the variance of the core parameters is relatively large, which leads to divergent model calculation results. Corresponding to actual patients, this means that there may be relatively large deviations.

[0005] 2. The parameters of these models are basically the results obtained by calculation through some actual medical case data in early years. There are problems such as different regional distribution of the population, measurement equipment deviation, and small number of cases.

[0006] 3. In therapeutic practice, it often happens that the predicted blood drug concentration value is significantly different from the monitored value, which leads to doubts about the applicability of the model.

[0007] Therefore, based on the population pharmacokinetic model, an optimized auxiliary generation method for personalized medication blood drug concentration prediction algorithm is expected to redesign the blood drug concentration prediction model to improve the accuracy and reliability of the prediction. Summary of the invention

[0008] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an auxiliary generation method and system for an individualized medication blood drug concentration prediction algorithm, which first extracts the model parameters of the established blood drug concentration prediction model, and uses machine learning technology to feature encode the model parameters of each medication blood drug concentration prediction model to learn the relationship between the various parameters in the model, and then, through multi-dimensional potential analysis of the parameter information of each medication blood drug concentration prediction model, dynamically aggregate the kernel features of the model parameters to learn the core distribution pattern of the model parameters, and then use this as a basis to restore the model parameters and generate a new medication blood drug concentration prediction model. Through this intelligent model generation method, the complexity of model construction can be significantly reduced, which helps to improve the accuracy of individualized medication blood drug concentration prediction and provide clinicians with more scientific and accurate medication guidance.

[0009] According to one aspect of the present application, a method for assisting in generating an algorithm for predicting blood drug concentration for personalized medication is provided, comprising: Step S1: extracting model parameters of the first to Nth drug blood concentration prediction models to obtain a set of prediction model parameters; Step S2: semantically encoding the model parameters of each prediction model in the set of prediction model parameters to obtain a set of semantic encoding vectors of prediction model parameter information; Step S3: performing kernel feature extraction on the set of semantic encoding vectors of the prediction model parameter information to obtain a distilled representation vector of the prediction model parameter information; Step S4: based on the prediction model parameter information, the distillation representation vector is used to restore the model parameters to obtain the N+1th medication blood concentration prediction model; Step S5: using the training data to determine the prediction accuracy of the N+1th medication blood concentration prediction model; Step S6: If the prediction accuracy of the N+1th medication blood concentration prediction model is the highest, the N+1th medication blood concentration prediction model is determined as the current optimal model.

[0010] According to another aspect of the present application, a system for assisting in generating an algorithm for predicting blood drug concentration for personalized medication is provided, comprising: A model parameter extraction module, used to extract model parameters of the first to Nth drug blood concentration prediction models to obtain a set of prediction model parameters; A model parameter information semantic encoding module, used for semantically encoding the model parameters of each prediction model in the set of prediction model parameters to obtain a set of prediction model parameter information semantic encoding vectors; A model parameter information distillation module, used for performing kernel feature extraction on the set of the prediction model parameter information semantic encoding vectors to obtain a prediction model parameter information distillation representation vector; A model parameter restoration module, used to perform model parameter restoration based on the prediction model parameter information distillation representation vector to obtain the N+1th medication blood drug concentration prediction model; A concentration prediction module, used to determine the prediction accuracy of the N+1th medication blood concentration prediction model using training data; The current optimal model generation module is used to determine the N+1th medication blood drug concentration prediction model as the current optimal model if the prediction accuracy of the N+1th medication blood drug concentration prediction model is the highest.

[0011] Compared with the prior art, the present application provides an auxiliary generation method and system for an individualized medication blood concentration prediction algorithm, which first extracts the model parameters of the existing medication blood concentration prediction model, and uses machine learning technology to feature encode the model parameters of each medication blood concentration prediction model to learn the relationship between the various parameters in the model, and then, through multi-dimensional energy potential analysis of the parameter information of each medication blood concentration prediction model, dynamically aggregate the kernel features of the model parameters to learn the core distribution pattern of the model parameters, and then use this as a basis to restore the model parameters and generate a new medication blood concentration prediction model. Through this intelligent model generation method, the complexity of model construction can be significantly reduced, which helps to improve the accuracy of individualized medication blood concentration prediction and provide more scientific and accurate medication guidance for clinicians. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 It is a flowchart of an auxiliary generation method of an individualized medication blood drug concentration prediction algorithm according to an embodiment of the present application; Figure 2 A data flow diagram of an auxiliary generation method of an individualized medication blood drug concentration prediction algorithm according to an embodiment of the present application; Figure 3 It is a flowchart of sub-step S3 of the auxiliary generation method of the personalized medication blood drug concentration prediction algorithm according to an embodiment of the present application; Figure 4It is a block diagram of an auxiliary generation system of an individualized medication blood concentration prediction algorithm according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0015] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0016] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0017] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0019] The population pharmacokinetic model currently used in practice also has some problems, such as: 1. In the model, the variance of the core parameters is relatively large, which leads to divergent model calculation results. Corresponding to actual patients, this means that there may be relatively large deviations.

[0020] 2. The parameters of these models are basically the results obtained by calculation through some actual medical case data in early years. There are problems such as different regional distribution of the population, measurement equipment deviation, and small number of cases.

[0021] 3. In therapeutic practice, it often happens that the predicted blood drug concentration value is significantly different from the monitored value, which leads to doubts about the applicability of the model.

[0022] Therefore, based on the population pharmacokinetic model, an optimized auxiliary generation method for personalized medication blood drug concentration prediction algorithm is expected to redesign the blood drug concentration prediction model to improve the accuracy and reliability of the prediction.

[0023] Based on this, in the technical solution of the present application, a method for assisting in generating an algorithm for predicting blood drug concentration for personalized medication is proposed. Figure 1 The present invention is a flowchart of an auxiliary generation method of an individualized medication blood concentration prediction algorithm according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the auxiliary generation method of the personalized medication blood drug concentration prediction algorithm according to the embodiment of the present application. Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the auxiliary generation method of the personalized medication blood concentration prediction algorithm includes the following steps: S1, extracting the model parameters of the first to N-th medication blood concentration prediction models to obtain a set of prediction model parameters; S2, semantically encoding the model parameters of each prediction model in the set of prediction model parameters to obtain a set of prediction model parameter information semantic encoding vectors; S3, performing kernel feature extraction on the set of prediction model parameter information semantic encoding vectors to obtain a prediction model parameter information distillation representation vector; S4, performing model parameter restoration based on the prediction model parameter information distillation representation vector to obtain the N+1-th medication blood concentration prediction model; S5, using training data to determine the prediction accuracy of the N+1-th medication blood concentration prediction model; S6, if the prediction accuracy of the N+1-th medication blood concentration prediction model is the highest, determining the N+1-th medication blood concentration prediction model as the current optimal model.

[0024] In particular, the S1 extracts the model parameters of the first to Nth medication blood concentration prediction models to obtain a set of prediction model parameters. It should be understood that the existing medication blood concentration prediction models have accumulated a large amount of clinical data and practical experience, and have a certain reference value. Therefore, in the technical solution of the present application, by extracting the model parameters of multiple medication blood concentration prediction models, model parameter analysis can be performed from multiple perspectives, making full use of the prior knowledge and experience of the existing models, thereby helping to improve the comprehensiveness and accuracy of the construction of new medication blood concentration prediction models.

[0025] In one example, first, select multiple existing medication blood concentration prediction models to ensure that these models cover different types of drugs and patient groups in order to extract more comprehensive parameter information. For example, nonlinear mixed effect models, Bayesian models, and decision tree models can be selected. Next, for each selected model, extract its core parameters, such as absorption rate constant (Ka), distribution volume (Vd), clearance rate (Cl), and half-life (t1 / 2). The parameter extraction method can be manual extraction or automatic extraction through programming, and the specific method depends on the storage format and data structure of the model. To ensure that the extracted parameters are accurate, the correctness of the parameters can be verified by consulting the original literature or model documentation. In addition, data cleaning techniques can be used to remove outliers or missing values ​​to ensure the quality of the extracted parameters.

[0026] Then, the extracted parameters are standardized to ensure that the parameters of different models are compared and analyzed on the same scale. Standardization methods can include normalization and standardization. Normalization scales the parameter values ​​to between 0 and 1, and standardization converts the parameter values ​​to a standard normal distribution with a mean of 0 and a standard deviation of 1. Using the normalization method, all parameter values ​​are scaled to between 0 and 1. By calculating the minimum and maximum values ​​of each parameter and normalizing the parameters using the normalization formula, the dimensional differences between the parameters of different models can be eliminated, improving the consistency and accuracy of subsequent processing.

[0027] Finally, all extracted and standardized parameters are organized into a set. The parameters of each model can be represented as a vector, and the parameter vectors of all models form a matrix. For example, if there are N models, each with M parameters, the final parameter set can be represented as an N×M matrix. By constructing such a parameter set matrix, subsequent feature encoding, kernel feature extraction, and model parameter restoration can be easily performed.

[0028] Through these steps, we can ensure that the extracted parameter set is of high quality and broad representativeness, providing a solid foundation for generating new blood drug concentration prediction models. The successful implementation of this step is a key link in the entire personalized drug blood drug concentration prediction algorithm-assisted generation method, which helps to improve the accuracy and reliability of the prediction and provide clinicians with more scientific and accurate medication guidance.

[0029] In particular, the S2 semantically encodes the model parameters of each prediction model in the set of prediction model parameters to obtain a set of semantically encoded vectors of prediction model parameter information. In a specific example of the present application, first, each prediction model parameter in the set of prediction model parameters is vectorized to obtain a set of prediction model parameter vectors; that is, in order to learn the potential association pattern between each model parameter in the medication blood concentration prediction model, the present application further vectorizes each prediction model parameter in the set of prediction model parameters. By arranging the model parameters of each medication blood concentration prediction model in a vector form to form a set of prediction model parameter vectors, it is helpful to unify the data format and structure, thereby facilitating the use of deep learning algorithms to identify and learn the parameter characteristics of each medication blood concentration prediction model. Next, a model parameter information encoder based on a multi-layer perceptron model is used to encode each prediction model parameter vector in the set of prediction model parameter vectors to obtain a set of semantically encoded vectors of prediction model parameter information. It should be understood that deep learning algorithms have significant advantages in processing high-dimensional data and identifying complex data association patterns, and the multi-layer perceptron model, as a basic structure of deep learning, can extract features and abstractly represent data through multi-layer nonlinear transformations. Therefore, in the technical solution of the present application, a multi-layer perceptron model is used to construct a model parameter information encoder, and information is encoded for each prediction model parameter vector in the set of prediction model parameter vectors. Through layer-by-layer processing of multiple layers of neurons, the potential association pattern between parameters in the prediction model parameter vector is captured, and the deep semantic information therein is extracted, thereby generating a set of prediction model parameter information semantic encoding vectors.

[0030] Specifically, each prediction model parameter in the set of prediction model parameters is vectorized to obtain a set of prediction model parameter vectors. By arranging the model parameters of each medication blood concentration prediction model in a vector form to form a set of prediction model parameter vectors, it is helpful to unify the data format and structure, thereby facilitating the use of deep learning algorithms to identify and learn the parameter characteristics of each medication blood concentration prediction model.

[0031] Next, in order to semantically encode these parameter vectors, a suitable deep learning model is selected. Commonly used models include multi-layer perceptron (MLP), convolutional neural network (CNN) and recurrent neural network (RNN). In this technical solution, a model parameter information encoder based on the multi-layer perceptron model is selected because the multi-layer perceptron model can extract features and abstractly represent data through multiple layers of nonlinear transformations. Multi-layer perceptron (MLP) is a feedforward neural network consisting of multiple input layers, hidden layers and output layers. Each hidden layer contains multiple neurons, which are nonlinearly transformed through activation functions (such as ReLU, Sigmoid, etc.). The MLP model can capture complex patterns and features in the input data through layer-by-layer processing of multiple layers of neurons.

[0032] To train a multi-layer perceptron model, the following steps are required: Use a normalized and vectorized set of parameters as input data.

[0033] Define the architecture of the multilayer perceptron model, including an input layer, multiple hidden layers, and an output layer. Each hidden layer can have different numbers of neurons and activation functions. Suppose we define a simple multilayer perceptron model with two hidden layers, each with 10 neurons, using the ReLU activation function, and an output layer with 4 neurons (the same number as the input parameters).

[0034] In particular, S3 performs kernel feature extraction on the set of semantic encoding vectors of the prediction model parameter information to obtain a distilled representation vector of the prediction model parameter information. Taking into account that different medication blood concentration prediction models usually have different prediction accuracy and scope of application, in order to fully integrate the advantages of each model, the core distribution pattern of the model parameters is extracted from the set of semantic encoding vectors of the prediction model parameter information. This application proposes a kernel feature extraction method based on multi-dimensional potential analysis, which dynamically aggregates parameter information by performing multi-dimensional potential analysis on the semantic encoding vectors of each prediction model parameter information, so as to learn the core distribution pattern of the model parameters. In a specific example of the present application, such as Figure 3 As shown, S3 includes: S31, performing multi-dimensional feature potential analysis on the set of semantic encoding vectors of the prediction model parameter information to obtain a set of superposition state potential factors of the prediction model parameter information; S32, based on the set of superposition state potential factors of the prediction model parameter information, performing kernel feature aggregation on the set of semantic encoding vectors of the prediction model parameter information to obtain the distillation representation vector of the prediction model parameter information.

[0035] Specifically, the S31 performs a multi-dimensional feature potential analysis on the set of semantic encoding vectors of the prediction model parameter information to obtain a set of superposition state potential factors of the prediction model parameter information. In an embodiment of the present application, first, the positional mean vector of the set of feature vectors of the prediction model parameter information is calculated to obtain the initial aggregate representation vector of the prediction model parameter information; that is, the set of feature vectors of the prediction model parameter information is calculated by positional mean to construct a typical example reflecting the overall characteristics of the prediction model parameters, capturing the overall distribution pattern of the prediction model parameters, thereby obtaining the initial aggregate representation vector of the prediction model parameter information. Next, the static potential factor of each prediction model parameter information feature vector in the set of the prediction model parameter information feature vector is calculated to obtain a set of static potential factors of the prediction model parameter information; for each prediction model parameter information feature vector, its static potential factor is calculated to reveal the information density it carries, and the inherent properties and information importance of each prediction model parameter information feature vector when it is independent of other vectors are evaluated. Then, the dynamic potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors is calculated relative to the prediction model parameter information initial aggregation representation vector to obtain a set of prediction model parameter information dynamic potential factors; considering that each prediction model parameter information feature vector has a certain influence on the overall model parameter distribution, therefore, the relative relationship between the data is captured by further calculating the dynamic potential factor of each prediction model parameter information feature vector relative to the prediction model parameter information initial aggregation representation vector, and its interaction and influence with the overall distribution pattern of the prediction model parameters is measured, so as to have a deeper understanding of how the parameter information of each prediction model affects its overall distribution. Subsequently, based on the set of the prediction model parameter information static potential factors and the set of the prediction model parameter information dynamic potential factors, the superposition state potential factors of each prediction model parameter information feature vector in the set of the prediction model parameter information feature vectors are determined to obtain the set of the prediction model parameter information superposition state potential factors. Here, in order to comprehensively consider the static and dynamic properties of each prediction model parameter information feature vector, the static energy potential factor and dynamic energy potential factor of each prediction model parameter information feature vector are further integrated to form a superposition energy potential factor. This comprehensive indicator is used to reflect the composite influence of each prediction model parameter information feature vector in the entire set, and to more comprehensively evaluate its contribution to the core distribution pattern of model parameters.

[0036] Among them, the specific process of calculating the static energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors to obtain the set of prediction model parameter information static energy potential factors includes: calculating the fourth-order central moment of the prediction model parameter information feature vector divided by the fourth power of its characteristic variance to obtain the prediction model parameter information static energy potential factor.

[0037] More specifically, the specific process of calculating the dynamic energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors relative to the prediction model parameter information initial aggregate representation vector to obtain a set of prediction model parameter information dynamic energy potential factors includes: calculating the position-weighted sum of the prediction model parameter information feature vector and the prediction model parameter information initial aggregate representation vector to obtain the prediction model parameter information dynamic energy potential representation vector; calculating the square of the L2 norm of the point-added vector between the prediction model parameter information dynamic energy potential representation vector and the bias vector to obtain the prediction model parameter information dynamic energy potential factor.

[0038] More specifically, based on the set of static energy potential factors of the prediction model parameter information and the set of dynamic energy potential factors of the prediction model parameter information, the specific process of determining the superposition state energy potential factors of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors to obtain the set of superposition state energy potential factors of the prediction model parameter information includes: using the squares of the static energy potential factor and the dynamic energy potential factor of the prediction model parameter information feature vector as exponents, respectively, calculating the exponential function value with base e to obtain the exponential static energy potential factor and the exponential dynamic energy potential factor, and adding the exponential static energy potential factor and the exponential dynamic energy potential factor to obtain the superposition state energy potential factor of the prediction model parameter information.

[0039] In summary, in the above embodiment, a multi-dimensional feature potential analysis is performed on the set of semantic encoding vectors of the prediction model parameter information to obtain a set of superposition state potential factors of the prediction model parameter information, including: performing a multi-dimensional feature potential analysis on the set of semantic encoding vectors of the prediction model parameter information using the following multi-dimensional feature potential analysis formula to obtain a set of superposition state potential factors of the prediction model parameter information; wherein the specific formula is: ; ; ; ; ; in, represents a set of feature vectors of parameter information of the prediction model, , , and Respectively represent the first, second, and third in the set of the prediction model parameter information feature vector and prediction model parameter information feature vector, is the number of feature vectors of parameter information of the prediction model, is the initial aggregation representation vector of the prediction model parameter information, Indicates the The first feature vector of the prediction model parameter information eigenvalues, and Respectively represent the The characteristic mean and characteristic variance of the characteristic vector of the prediction model parameter information, represents the length of the prediction model parameter information feature vector, , and Respectively represent the The static potential factor, dynamic potential factor and superposition potential factor of the prediction model parameter information feature vector, represents the bias vector, and represents different weight parameters, represents dot product, It means calculating the square of L2 norm of a vector.

[0040] Specifically, the S32, based on the set of superposition state potential factors of the prediction model parameter information, performs kernel feature aggregation on the set of semantic encoding vectors of the prediction model parameter information to obtain the distillation representation vector of the prediction model parameter information. In an embodiment of the present application, first, the set of superposition state potential factors of the prediction model parameter information is input into the potential factor gated weighting network to obtain a set of potential weight coefficients of the prediction model parameter information feature; that is, based on the gated mask mechanism, each generated superposition state potential factor is normalized and weighted, so as to perform weighted aggregation on the set of feature vectors of the prediction model parameter information, thereby explicitly marking the prediction model parameter information that has a greater impact on the core distribution pattern of the model parameters, filtering out the feature representation that contributes less to the overall pattern, and obtaining the distillation representation vector of the prediction model parameter information, thereby realizing the accurate extraction of the core distribution pattern of the model parameters, and providing a solid data foundation for the subsequent model parameter restoration and the generation of new models. Finally, based on the set of potential weight coefficients of the prediction model parameter information feature, the position-weighted sum of the set of feature vectors of the prediction model parameter information is calculated to obtain the distillation representation vector of the prediction model parameter information. That is, based on the set of potential weight coefficients of the prediction model parameter information feature, the position-weighted sum of the set of prediction model parameter information feature vectors is calculated to obtain the prediction model parameter information distillation representation vector.

[0041] Among them, the specific process of inputting the set of energy potential factors of the superposition state of the prediction model parameter information into the energy potential factor gated weighting network to obtain the set of characteristic energy potential weight coefficients of the prediction model parameter information includes: after inputting the set of energy potential factors of the superposition state of the prediction model parameter information into the sigmoid function for normalization processing, the normalized set of energy potential factors of the superposition state of the prediction model parameter information is masked and inactivated based on the mask threshold to obtain the set of characteristic energy potential weight coefficients of the prediction model parameter information.

[0042] In summary, in the above embodiment, based on the set of superimposed state energy potential factors of the prediction model parameter information, kernel feature aggregation is performed on the set of semantic encoding vectors of the prediction model parameter information to obtain the distilled representation vector of the prediction model parameter information, including: based on the set of superimposed state energy potential factors of the prediction model parameter information, kernel feature aggregation is performed on the set of semantic encoding vectors of the prediction model parameter information using the following kernel feature aggregation formula to obtain the distilled representation vector of the prediction model parameter information, wherein the kernel feature aggregation formula is: ; ; ; ; in, Indicates A normalized prediction model parameter information superposition state energy potential factor, is the mask function, represents the mask threshold, For the The information characteristic energy potential weight coefficient of the prediction model parameter, A distilled representation vector representing the prediction model parameter information.

[0043] In particular, the S4 performs model parameter reduction based on the prediction model parameter information distillation representation vector to obtain the N+1 medication blood concentration prediction model. In a specific example of the present application, the prediction model parameter information distillation representation vector is input into a decoder-based model parameter reducer to obtain the N+1 medication blood concentration prediction model. In the technical solution of the present application, in order to achieve the reduction of model parameters and the generation of new models, the present application uses a decoder to construct a model parameter reducer, and decodes and regresses the prediction model parameter information distillation representation vector to reduce it to specific model parameters, thereby generating the N+1 medication blood concentration prediction model. In a specific example of the present application, the decoder uses a recursive neural network (RNN) structure, and through the layer-by-layer expansion of the recursive neural network, the prediction model parameter information distillation representation vector is gradually reduced to a complete model parameter sequence, and its cyclic memory ability is used to maintain the dependency relationship between the generated model parameters, thereby ensuring that the newly generated medication blood concentration prediction model parameters have logical consistency and clinical applicability.

[0044] Here, the role of the decoder model is to restore specific model parameters from the prediction model parameter information distillation representation vector. Common decoder models include recursive neural networks (RNNs), multi-layer perceptrons (MLPs), etc. In the technical solution of the present application, a decoder based on a multi-layer perceptron model is selected because the multi-layer perceptron model can extract and restore features of data through multi-layer nonlinear transformations.

[0045] Preferably, in the technical solution of the present application, in the process of performing kernel feature extraction on the set of the prediction model parameter information semantic coding vectors to obtain the prediction model parameter information distillation representation vector, since each prediction model parameter information semantic coding vector in the set of the prediction model parameter information semantic coding vectors respectively represents the semantic coding features of the model parameters of each prediction model, but since the model parameters of each prediction model themselves have convergence inconsistency and strong unexplainability, it will cause the prediction model parameter information distillation representation vector obtained based on the kernel feature extraction to have decoding space uncertainty, that is, it will affect the accuracy of the model parameters of the N+1th medication blood concentration prediction model obtained by inputting the model parameter reducer based on the decoder.

[0046] Based on this, in a preferred embodiment, in the process of inputting the prediction model parameter information distillation representation vector into the decoder-based model parameter reducer to obtain the N+1 medication blood drug concentration prediction model, the prediction model parameter information distillation representation vector is optimized for feature expression, and the process includes: Arranging the prediction model parameter information distillation representation vector based on the eigenvalue size to obtain a prediction model parameter information distillation sequence response encoding vector; Calculating a transposed vector of the prediction model parameter information distillation representation vector to obtain a prediction model parameter information distillation representation transposed vector; Multiplying the prediction model parameter information distillation order response encoding vector by the prediction model parameter information distillation representation transposed vector to obtain a prediction model parameter information distillation order topological encoding matrix; Calculate the hyperbolic sine function value of the sum of squares of all eigenvalues ​​of the prediction model parameter information distillation representation vector to obtain a topological scaling factor; ; in, The first eigenvalues, represents the length of the distillation representation vector of the prediction model parameter information, represents the hyperbolic sine function, represents the topological scaling factor; Multiplying the topological scaling factor by the prediction model parameter information distillation order topological coding matrix to obtain a prediction model parameter information distillation order topological coding scale adaptation matrix; The prediction model parameter information distillation representation vector is multiplied by the prediction model parameter information distillation order topology coding scale adaptation matrix to obtain an optimized prediction model parameter information distillation representation vector.

[0047] That is, by mapping discrete tensors into a topological space defined by the scalar product of vectors, meaningful measurement of the statistical correlation of attribute clusters within a stable domain can be achieved. On this basis, a representation space with a non-coupled framework is constructed through dynamic coordinate mapping with the representation tensor, and the multi-layer topology of the representation tensor is morphologically integrated within the representation space based on vector queries. This can avoid reducing the information fidelity of the representation tensor due to loose morphology, thereby increasing the reliability of the judgment output obtained by the decision module of the discriminant attribute tensor, thereby improving the accuracy and robustness of the model parameters of the N+1th medication blood concentration prediction model obtained by the model parameter reducer based on the decoder.

[0048] In particular, the S5 uses training data to determine the prediction accuracy of the N+1 medication blood drug concentration prediction model. In one example, the training data contains known inputs (such as dosage, time, etc.) and outputs (such as actual measured blood drug concentrations). Through these data, the performance of the model under known conditions can be evaluated. The training data is used to help adjust the parameters of the model so that it better fits the known data. That is, the performance of the N+1 medication blood drug concentration prediction model is verified using a training data set with known inputs and outputs to detect the prediction accuracy of the N+1 medication blood drug concentration prediction model to ensure that the performance of the model in actual applications meets expectations.

[0049] In the auxiliary generation method of the personalized medication blood concentration prediction algorithm, using training data to determine the prediction accuracy of the N+1 medication blood concentration prediction model is a key step. This process involves multiple sub-steps, including data preparation, model training, model evaluation and result analysis. The following are the detailed implementation steps: First, prepare datasets for training and evaluation. These datasets usually include patients' medication records, blood drug concentration measurements, and other relevant features (such as age, weight, liver and kidney function, etc.). The dataset should be divided into training and test sets so that independent evaluation can be performed after training the model. Suppose there is a dataset containing multiple patient medication records, each record includes the following information: patient ID, medication dosage, administration time, blood drug concentration measurement, age, weight, liver and kidney function.

[0050] The training data set is used to train the N+1 medication blood concentration prediction model. During the training process, the model will learn the mapping relationship between input features (such as medication dosage, administration time, age, weight, liver and kidney function, etc.) and blood drug concentration measurement values.

[0051] After training is completed, the prediction accuracy of the model is evaluated using the test data set. Evaluation indicators usually include mean square error (MSE), mean absolute error (MAE), coefficient of determination (R²), etc.

[0052] Based on the evaluation results, analyze the prediction accuracy of the model. Common evaluation indicators and their significance are as follows: Mean Squared Error (MSE): Measures the average of the squared errors between the predicted values ​​and the true values. The smaller the MSE value, the smaller the prediction error of the model.

[0053] Mean Absolute Error (MAE): Measures the average absolute error between the predicted value and the true value. The smaller the MAE value, the smaller the prediction error of the model.

[0054] Coefficient of determination (R²): A measure of the proportion of variability explained by the model. The closer the R² value is to 1, the better the model fit is.

[0055] These results show that the model performs well on the test data set, has a small prediction error, and can well explain the variability of the data. Through the above steps, we can use the training data to determine the prediction accuracy of the N+1 medication blood concentration prediction model, thereby ensuring the reliability and effectiveness of the model in practical applications.

[0056] In particular, in S6, if the prediction accuracy of the N+1th medication blood concentration prediction model is the highest, the N+1th medication blood concentration prediction model is determined as the current optimal model. That is, by comparing the prediction results of the N+1th medication blood concentration prediction model with those of the existing prediction model under the same conditions, the performance improvement of the new model is evaluated. If the prediction accuracy of the N+1th medication blood concentration prediction model is the highest, the N+1th medication blood concentration prediction model is determined as the current optimal model.

[0057] In summary, the auxiliary generation method of the personalized medication blood concentration prediction algorithm according to the embodiment of the present application is explained, which first extracts the model parameters of the existing medication blood concentration prediction model, and uses machine learning technology to feature encode the model parameters of each medication blood concentration prediction model to learn the relationship between the various parameters in the model, and then, through multi-dimensional potential analysis of the parameter information of each medication blood concentration prediction model, dynamically aggregate the kernel features of the model parameters to learn the core distribution pattern of the model parameters, and then use this as a basis to restore the model parameters and generate a new medication blood concentration prediction model. Through this intelligent model generation method, the complexity of model construction can be significantly reduced, which helps to improve the accuracy of personalized medication blood concentration prediction and provide more scientific and accurate medication guidance for clinicians.

[0058] Furthermore, a system for assisting in generating an algorithm for predicting blood drug concentrations for individualized medication is also provided.

[0059] Figure 4FIG. 1 is a block diagram of an auxiliary generation system for an individualized medication blood drug concentration prediction algorithm according to an embodiment of the present application. Figure 4 As shown, according to the embodiment of the present application, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm includes: a model parameter extraction module 310, which is used to extract the model parameters of the first to N-th medication blood concentration prediction models to obtain a set of prediction model parameters; a model parameter information semantic encoding module 320, which is used to semantically encode the model parameters of each prediction model in the set of prediction model parameters to obtain a set of prediction model parameter information semantic encoding vectors; a model parameter information distillation module 330, which is used to perform kernel feature extraction on the set of prediction model parameter information semantic encoding vectors to obtain a prediction model parameter information distillation representation vector; a model parameter restoration module 340, which is used to perform model parameter restoration based on the prediction model parameter information distillation representation vector to obtain the N+1-th medication blood concentration prediction model; a concentration prediction module 350, which is used to determine the prediction accuracy of the N+1-th medication blood concentration prediction model using training data; a current optimal model generation module 360, which is used to determine the N+1-th medication blood concentration prediction model as the current optimal model if the prediction accuracy of the N+1-th medication blood concentration prediction model is the highest.

[0060] As described above, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm according to the embodiment of the present application can be implemented in various wireless terminals, such as a server having an auxiliary generation algorithm for the personalized medication blood concentration prediction algorithm. In one possible implementation, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm can also be one of the many hardware modules of the wireless terminal.

[0061] Alternatively, in another example, the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm and the wireless terminal may also be separate devices, and the auxiliary generation system 300 of the personalized medication blood concentration prediction algorithm may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0062] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for assisting the generation of an algorithm for predicting the blood concentration of individualized medication, characterized in that: include: Step S1: extracting model parameters of the first to Nth blood drug concentration prediction models to obtain a set of prediction model parameters; Step S2: semantically encoding the model parameters of each prediction model in the set of prediction model parameters to obtain a set of semantic encoding vectors of prediction model parameter information; Step S3: performing kernel feature extraction on the set of semantic encoding vectors of the prediction model parameter information to obtain a distilled representation vector of the prediction model parameter information; Step S4: based on the prediction model parameter information, the distillation representation vector is used to restore the model parameters to obtain the N+1th medication blood concentration prediction model; Step S5: using the training data to determine the prediction accuracy of the N+1th medication blood concentration prediction model; Step S6: If the prediction accuracy of the N+1th medication blood concentration prediction model is the highest, the N+1th medication blood concentration prediction model is determined as the current optimal model.

2. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 1, characterized in that: The step S2 comprises: Vectorizing each prediction model parameter in the set of prediction model parameters to obtain a set of prediction model parameter vectors; A model parameter information encoder based on a multi-layer perceptron model is used to encode each prediction model parameter vector in the set of prediction model parameter vectors to obtain a set of prediction model parameter information semantic encoding vectors.

3. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 2, characterized in that: The step S3 comprises: Performing a multi-dimensional feature potential analysis on the set of semantic encoding vectors of the prediction model parameter information to obtain a set of superposition state potential factors of the prediction model parameter information; Based on the set of superposition state energy potential factors of the prediction model parameter information, kernel feature aggregation is performed on the set of semantic encoding vectors of the prediction model parameter information to obtain a distillation representation vector of the prediction model parameter information.

4. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 3, characterized in that: A multi-dimensional feature potential analysis is performed on the set of semantic encoding vectors of the prediction model parameter information to obtain a set of superposition state potential factors of the prediction model parameter information, including: Calculating the positional mean vector of the set of prediction model parameter information feature vectors to obtain an initial aggregate representation vector of the prediction model parameter information; Calculating the static energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors to obtain a set of prediction model parameter information static energy potential factors; Calculating the dynamic energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors relative to the prediction model parameter information initial aggregation representation vector to obtain a set of prediction model parameter information dynamic energy potential factors; Based on the set of static energy potential factors of the prediction model parameter information and the set of dynamic energy potential factors of the prediction model parameter information, the superposition state energy potential factors of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors are determined to obtain the set of superposition state energy potential factors of the prediction model parameter information.

5. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 4, characterized in that: Calculating the static energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors to obtain a set of prediction model parameter information static energy potential factors, including: The fourth-order central moment of the prediction model parameter information feature vector is calculated and divided by the fourth power of its feature variance to obtain the prediction model parameter information static energy potential factor.

6. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 5, characterized in that: Calculating the dynamic potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors relative to the prediction model parameter information initial aggregation representation vector to obtain a set of prediction model parameter information dynamic potential factors, including: Calculating the position-weighted sum of the prediction model parameter information feature vector and the prediction model parameter information initial aggregation representation vector to obtain a prediction model parameter information dynamic energy potential representation vector; The square of the L2 norm of the point-added vector between the prediction model parameter information dynamic energy potential representation vector and the bias vector is calculated to obtain the prediction model parameter information dynamic energy potential factor.

7. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 6, characterized in that: Based on the set of static energy potential factors of the prediction model parameter information and the set of dynamic energy potential factors of the prediction model parameter information, determining the superposition state energy potential factor of each prediction model parameter information feature vector in the set of prediction model parameter information feature vectors to obtain the set of superposition state energy potential factors of the prediction model parameter information, including: The squares of the static energy potential factor and the dynamic energy potential factor of the characteristic vector of the prediction model parameter information are respectively used as exponents, and the exponential function values ​​with base e are calculated to obtain the exponential static energy potential factor and the exponential dynamic energy potential factor, and the exponential static energy potential factor and the exponential dynamic energy potential factor are added and summed to obtain the superposition state energy potential factor of the prediction model parameter information.

8. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 7, characterized in that: Based on the set of superposition state energy potential factors of the prediction model parameter information, kernel feature aggregation is performed on the set of semantic encoding vectors of the prediction model parameter information to obtain a distillation representation vector of the prediction model parameter information, including: Inputting the set of the prediction model parameter information superposition state energy potential factors into the energy potential factor gating weighting network to obtain a set of prediction model parameter information characteristic energy potential weight coefficients; Based on the set of potential weight coefficients of the prediction model parameter information feature, a position-weighted sum of the set of prediction model parameter information feature vectors is calculated to obtain the prediction model parameter information distillation representation vector.

9. The method for assisting in generating an algorithm for predicting the blood concentration of individualized medication according to claim 8, characterized in that: The step S4 comprises: The prediction model parameter information distillation representation vector is input into a decoder-based model parameter reducer to obtain the N+1th medication blood concentration prediction model.

10. An auxiliary generation system for an individualized medication blood concentration prediction algorithm, characterized in that: include: A model parameter extraction module, used to extract model parameters of the first to Nth drug blood concentration prediction models to obtain a set of prediction model parameters; A model parameter information semantic encoding module, used for semantically encoding the model parameters of each prediction model in the set of prediction model parameters to obtain a set of prediction model parameter information semantic encoding vectors; A model parameter information distillation module, used for performing kernel feature extraction on the set of the prediction model parameter information semantic encoding vectors to obtain a prediction model parameter information distillation representation vector; A model parameter restoration module, used to perform model parameter restoration based on the prediction model parameter information distillation representation vector to obtain the N+1th medication blood drug concentration prediction model; A concentration prediction module, used to determine the prediction accuracy of the N+1th medication blood concentration prediction model using training data; The current optimal model generation module is used to determine the N+1th medication blood drug concentration prediction model as the current optimal model if the prediction accuracy of the N+1th medication blood drug concentration prediction model is the highest.

Images