Precise medication mode construction method and system based on pharmacokinetics and artificial intelligence

Abstract

The invention discloses a precise medication mode construction method and system based on pharmacokinetics and artificial intelligence, and the method comprises the steps: carrying out the multi-stage preprocessing of an obtained drug concentration monitoring data file, and obtaining a structured data set for pharmacokinetics analysis; a drug metabolism concentration prediction model is constructed, a neural network is utilized to learn from the structured data set, pharmacokinetic parameters are generated, prediction concentration is calculated through the pharmacokinetic model, and the drug metabolism concentration prediction model is trained based on the structured data set; drawing a drug concentration metabolism curve of an individual patient based on the constructed drug metabolism concentration prediction model, realizing individual drug administration decision, and realizing drug concentration monitoring; and visually deploying all the algorithms and data to a front-end interface. According to the invention, individualized precise medication can be realized, the safety and effectiveness of drug treatment can be improved, clinical availability and reliability are realized, and the clinical working efficiency and decision scientificity can be improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent medical treatment, and particularly relates to a precision medication mode construction method and system based on pharmacokinetics and artificial intelligence. BACKGROUND

[0002] Personalized precision dosing is an important trend in contemporary medical practice, and its core lies in tailoring treatment strategies according to the unique physiological characteristics of patients (such as gender, age, disease status, etc.). This individualized treatment model not only can significantly improve drug efficacy, but also can effectively control the risk of medication, especially for drugs with small therapeutic index and narrow safety range. It is particularly important to achieve precision dosing. In clinical practice, due to significant biological differences between patients, the pharmacokinetic processes (including absorption, distribution, metabolism, and excretion) of the same drug in different individuals may show significant differences, which makes it difficult for standardized dosing regimens to meet the treatment needs of all patients. Therefore, it is urgent to establish an individualized dosing model based on pharmacokinetic principles to achieve precision dosing.

[0003] Pharmacokinetic research is the theoretical foundation for achieving precision dosing, which analyzes the dynamic changes of drugs in the body and provides a scientific basis for developing reasonable dosing regimens. Traditional pharmacokinetic modeling methods mainly use population pharmacokinetic models or linear dynamic models, such as one-compartment or two-compartment models. These models describe the drug's body process through parameterization, which to some extent provides a reference for clinical dosing. However, such models are usually based on many idealized assumptions and are difficult to accurately reflect the common multi-factor, non-linear drug metabolism characteristics in clinical practice. In particular, when dealing with complex situations such as combined medication, special physiological states (such as pregnancy, organ dysfunction), traditional models have obvious shortcomings in characterizing individual differences and analyzing non-linear dynamics, making it difficult to meet the needs of precision dosing in complex clinical situations.

[0004] With the rapid development of intelligent computing technology, machine learning algorithms (such as ensemble learning, kernel methods, and gradient boosting decision trees) have opened up new avenues for precision medicine research. These algorithms are good at handling high-dimensional complex data and can identify non-linear relationships that traditional analysis methods cannot capture. In pharmacokinetic research, intelligent algorithms can be used to model drug concentration-time curves without relying on strict assumptions, revealing potential complex relationships through data-driven methods. This modeling strategy not only improves prediction accuracy, but also provides better solutions for individualized dosing, effectively overcoming the limitations of traditional models. However, it is worth noting that relying entirely on data-driven concentration prediction methods also has some problems, which may overlook the basic biological principles of pharmacokinetics. Although high prediction accuracy may be achieved, the biological reasonableness and interpretability of the model are often difficult to guarantee, which to some extent limits its widespread application in clinical practice.

[0005] In summary, the existing traditional pharmacokinetic model is difficult to meet the demand of precise drug delivery in complex clinical situations. Although intelligent computing technology can improve prediction accuracy, the concentration prediction method relying entirely on data-driven has problems such as neglecting the basic biological principles of pharmacokinetics and poor model interpretability. Therefore, it is necessary to combine the advantages of traditional pharmacokinetic principles and intelligent algorithms to build a better individualized precise drug delivery model. SUMMARY

[0006] In view of the above, the purpose of the present application is to provide a precise drug delivery mode construction method and system based on pharmacokinetics and artificial intelligence. Based on the verified pharmacokinetic model theoretical framework, the model parameters are modeled and fitted through machine learning or deep learning, solving the blindness of black box model simply relying on data fitting and enhancing the robustness of the model. At the same time, it provides intuitive dynamic concentration change visualization display function of the drug in the human body, provides strong support for clinical decision-making, and constructs a standard interface matching the diagnosis and treatment process to ensure that the model prediction results can be directly converted into executable clinical schemes, which is convenient for doctors to interactively verify and adjust in real time.

[0007] To achieve the above-mentioned purpose of the application, the technical solutions provided by the present application are as follows: In a first aspect, the present application provides a precise drug delivery mode construction method based on pharmacokinetics and artificial intelligence, comprising the following steps: Multi-stage preprocessing of the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis; Constructing a drug metabolism concentration prediction model, wherein the neural network is used to learn and generate pharmacokinetic parameters from the structured data set, and the predicted concentration is calculated through the pharmacokinetic model, and the drug metabolism concentration prediction model is trained based on the structured data set; Based on the constructed drug metabolism concentration prediction model, the drug concentration at different sampling points is calculated according to the clinical characteristics and drug dose of the patient, and the individual patient drug concentration metabolism curve is drawn; Based on the constructed drug metabolism concentration prediction model, the drug dose and drug interval are calculated according to the clinical characteristics, target drug duration and target concentration of the patient to realize individual drug decision-making; Based on the constructed drug metabolism concentration prediction model, the drug time is calculated according to the clinical characteristics, drug dose, drug interval and drug concentration threshold of the patient to realize drug concentration monitoring; All the above algorithms and data are visualized and deployed to the front-end interface to realize the login interface, data uploading function, data query function, data downloading function, selection of model and drug species, individual patient drug concentration metabolism curve display, individual drug decision-making and drug concentration monitoring function.

[0008] Preferably, the acquired drug concentration monitoring data files are subjected to multi-stage preprocessing to obtain a structured dataset for pharmacokinetic analysis, including: For the abnormal drug cycle identification problem existing in the clinical records of the drug concentration monitoring data files, a sliding window mechanism is established to traverse all record lines, verify whether the end time of the previous cycle and the start time of the next cycle of the same patient are continuous, and if the time sequence is broken, it is determined as an abnormal separation, and the abnormal separation point and all redundant records of the same cycle traced upward are removed; For the clinical record characteristics of the same treatment cycle scattered in multiple lines in the drug concentration monitoring data files, a hierarchical structure replication mechanism is established to integrate the key field values of the scattered lines into the main record, realizing the vertical aggregation of multiple drug administration records of the same patient; For the case that the data structure of the drug concentration monitoring data files does not meet the requirements of pharmacokinetic analysis, the data structure is reconstructed, the drug administration date and drug concentration detection value are horizontally associated to form a standard two-dimensional table, the backtracking deletion mechanism is used to ensure the time continuity, the field replication is used to establish the explicit association of drug administration and concentration detection, and the null value is filled to generate a structured dataset for pharmacokinetic analysis.

[0009] Preferably, the drug metabolism concentration prediction model includes: One-compartment model, corresponding to two pharmacokinetic parameters, distribution volume , elimination rate , wherein the pharmacokinetic model is represented as: , wherein, represents the predicted concentration at time , and represents the drug dose; Two-compartment model, corresponding to four pharmacokinetic parameters, distribution volume , elimination rate , and , inter-compartment weight factor , wherein the pharmacokinetic model is represented as: , wherein, represents the inter-compartment weight factor, and ; Three-compartment model, corresponding to seven pharmacokinetic parameters, distribution volume , elimination rate , , and , inter-compartment weight factor , , and , wherein the pharmacokinetic model is represented as: , wherein, .

[0010] Preferably, the predicted concentration calculated by the pharmacokinetic model after the pharmacokinetic parameters are learned and generated from the structured dataset using the neural network comprises: The neural network is trained for the clinical characteristics of the patients in the structured dataset, and each neural network corresponds to generate a pharmacokinetic parameter, and the neural network structure of each channel comprises: For a one-compartment model, first, the input layer receives the standardized patient clinical characteristic vector as the input signal of the neuron, a portion of the neurons are randomly discarded using the Dropout layer, the neuron output is layer-normalized using the LayerNorm layer, the non-linear transformation is realized using the GeLU activation function, the hidden layer is connected with the weight matrix, and the model is complicated again using the Dropout layer, the LayerNorm layer and the GeLU activation function. Finally, a single neuron is output, and the value output by each neuron is the generated pharmacokinetic parameter, and the predicted concentration is calculated by the pharmacokinetic model according to the generated pharmacokinetic parameter; For a two-compartment model or a three-compartment model, first, the input layer receives the standardized patient clinical characteristic vector as the input signal of the neuron, the neuron output is layer-normalized using the LayerNorm layer, the non-linear transformation is realized using the GeLU activation function, the hidden layer is connected with the weight matrix, and the model is complicated again using the LayerNorm layer and the GeLU activation function. Finally, a single neuron is output, and the value output by each neuron is the generated pharmacokinetic parameter, and the predicted concentration is calculated by the pharmacokinetic model according to the generated pharmacokinetic parameter.

[0011] Preferably, the value output by each neuron is the generated pharmacokinetic parameter, comprising: For the distribution volume and elimination rate of pharmacokinetics, the original parameter value output by the neuron is respectively subjected to exponential transformation, so as to constrain the value to be positive, so as to obtain the corresponding pharmacokinetic parameter; For the weight factor of pharmacokinetics, the neuron output value is transformed using the sigmoid algorithm or the softmax algorithm, so that the weight factor value is positive and is constrained in the range of 0-1.

[0012] Preferably, a one-compartment model is used for prediction for a simple metabolized drug, a two-compartment model is used for prediction for a drug with obvious tissue distribution, and a three-compartment model is used for prediction for a long-acting or complex distributed drug.

[0013] Preferably, the drug concentration prediction model is constructed based on the clinical characteristics of the patient and the drug dose to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolic curve, comprising: After the clinical characteristics of the patient are data standardized and input into the drug concentration prediction model, the required pharmacokinetic parameters are calculated by the neural network prediction output model, and a number of sampling points based on the pharmacokinetic curve half-life are generated, the pharmacokinetic parameters and the drug dose are input into the pharmacokinetic model to calculate the predicted concentration, and the individual patient drug concentration metabolic curve is drawn based on the sampling points and the predicted concentration.

[0014] Preferably, the drug concentration prediction model is constructed based on the clinical characteristics of the patient and the drug dose to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolic curve, comprising: After the clinical characteristics of the patient are data standardized and input into the drug concentration prediction model, the required pharmacokinetic parameters are calculated by the neural network prediction output model, and a number of sampling points based on the pharmacokinetic curve half-life are generated, the pharmacokinetic parameters and the drug dose are input into the pharmacokinetic model to calculate the predicted concentration, and the individual patient drug concentration metabolic curve is drawn based on the sampling points and the predicted concentration.

[0015] Preferably, the drug concentration prediction model is constructed based on the clinical characteristics of the patient and the drug dose to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolic curve, comprising: After the clinical characteristics of the patient are data standardized and input into the drug concentration prediction model, the required pharmacokinetic parameters are calculated by the neural network prediction output model, and a number of sampling points based on the pharmacokinetic curve half-life are generated, the pharmacokinetic parameters and the drug dose are input into the pharmacokinetic model to calculate the predicted concentration, and the individual patient drug concentration metabolic curve is drawn based on the sampling points and the predicted concentration.

[0016] In a second aspect, the embodiments of the present application also provide a precision medication mode construction system based on pharmacokinetics and artificial intelligence, which is realized by using the precision medication mode construction method based on pharmacokinetics and artificial intelligence described above, comprising: a data set construction module, a pharmacokinetic prediction model modeling module, a pharmacokinetic curve analysis module, an individual drug administration decision module, a drug concentration monitoring module, and a visual interaction module. The data set construction module is used to perform multi-level preprocessing on the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis. The pharmacokinetics prediction model modeling module is configured to construct a drug metabolism concentration prediction model, wherein the neural network is used to learn and generate pharmacokinetic parameters from the structured data set, the predicted concentration is calculated through the pharmacokinetic model, and the drug metabolism concentration prediction model is trained based on the structured data set; The pharmacokinetics curve analysis module is configured to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolism curve based on the constructed drug metabolism concentration prediction model according to the clinical characteristics and the drug dose of the patient; The individual drug administration decision module is configured to calculate the drug dose and the drug interval based on the constructed drug metabolism concentration prediction model according to the clinical characteristics, the target drug administration time length and the target concentration of the patient to realize the individual drug administration decision; The drug concentration monitoring module is configured to calculate the drug administration time based on the constructed drug metabolism concentration prediction model according to the clinical characteristics, the drug dose, the drug interval and the drug concentration threshold of the patient to realize the drug concentration monitoring; The visual interaction module is configured to visualize all the algorithms and data and deploy them to the front-end interface to realize the login interface, the data uploading function, the data query function, the data downloading function, the selection of the model and the drug type, the individual patient drug concentration metabolism curve display, the individual drug administration decision judgment and the drug concentration monitoring function.

[0017] Compared with the prior art, the present application has at least the following beneficial effects: (1) The present application combines the pharmacokinetic model with machine learning and deep learning technology, proposes a parallel neural network parameter processor structure, simultaneously generates multiple key pharmacokinetic parameters, and ensures that the parameters meet the physical constraints of the pharmacokinetic model through different mathematical transformations, breaking the limitations of traditional pharmacokinetic models in high-dimensional nonlinear data analysis, and enabling more detailed and accurate capture of the dynamic characteristics of drugs in the individual body, thereby realizing individualized precision medicine and helping to improve the safety and effectiveness of drug treatment.

[0018] (2) The present application integrates the functions related to the pharmacokinetic model into the neural network, optimizes the model based on the physiological and biological principles of pharmacokinetics by back-propagating the loss function of the machine learning model, so that the model has interpretability and the parameters have pharmacokinetic significance. This method not only improves the accuracy of prediction, but also ensures the interpretability and robustness of the model. This method avoids the "black box" problem that may occur in pure data-driven artificial intelligence models, ensuring the clinical usability and reliability of the model results.

[0019] (3) The application can provide personalized medication recommendations for different populations by reverse deduction of the drug metabolism concentration prediction model to give drug decision, effectively avoid the adverse reaction risk of medication, and significantly improve the safety and tolerance of drug treatment. At the same time, during the drug treatment process, through continuous monitoring of the drug concentration, the regulation effect of the drug on various biochemical indicators and the condition of the human body is explored in depth, which helps doctors to understand the drug efficacy and the physical reaction of the patient in time, and to flexibly adjust the treatment plan according to the actual situation, so as to realize more accurate and effective drug treatment.

[0020] (4) The application realizes interactive operation through the model construction system, and the system can be easily operated through the front-end interface. In the system, the doctor can flexibly select the applicable model and drug type according to the specific condition of the patient, and quickly generate the drug concentration change prediction curve of the individual patient. This intuitive and convenient visual display mode enables the doctor to clearly understand the metabolism process and concentration change trend of the drug in the patient's body, provides an intuitive and accurate reference basis for clinical medication decision, and helps to improve the clinical work efficiency and the scientific nature of the decision. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0022] Figure 1 is a process schematic diagram of the precise medication mode construction method based on pharmacokinetics and artificial intelligence provided by the embodiment of the application; Figure 2 is a login interface schematic diagram of the individualized precise medication platform provided by the embodiment of the application; Figure 3 is a data upload function interface schematic diagram of the individualized precise medication platform provided by the embodiment of the application; Figure 4 is a data query function interface schematic diagram of the individualized precise medication platform provided by the embodiment of the application; Figure 5 is a data download function interface schematic diagram of the individualized precise medication platform provided by the embodiment of the application; Figure 6 is a selection model and drug type function interface schematic diagram of the individualized precise medication platform provided by the embodiment of the application; Figure 7 is a data selection schematic diagram of the individual patient drug concentration metabolism curve of the individualized precise medication platform provided by the embodiment of the application; Figure 8 is a schematic diagram of chart drawing in an individual patient drug concentration metabolic curve of an individualized precision drug use platform provided by an embodiment of the present application; Figure 9 is a schematic diagram of an interface of a drug administration decision-making function of the individualized precision drug use platform provided by the embodiment of the present application; Figure 10 is a schematic diagram of an interface of a drug concentration monitoring function of the individualized precision drug use platform provided by the embodiment of the present application; Figure 11 is a structural schematic diagram of a precision drug use mode construction system based on pharmacokinetics and artificial intelligence provided by the embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0024] The inventive concept of the present application is that: in view of the fact that the conventional pharmacokinetic model in the prior art has limited ability to process high-dimensional nonlinear data, it is difficult to accurately capture the dynamic changes of drugs in individuals, the pure data-driven artificial intelligence model has a "black box" problem, lacks explainability and clinical reliability, and the drug administration decision lacks individualization, the doctor's operation is inconvenient and the drug concentration change cannot be directly obtained, a conventional pharmacokinetic model is combined with machine learning and deep learning to construct an interactive model system, and precision drug use and convenient operation are realized.

[0025] As shown in Figure 1 , the embodiment provides a precision drug use mode construction method based on pharmacokinetics and artificial intelligence, including the following steps: S1, multi-level preprocessing is performed on the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis.

[0026] In the embodiment, based on the drug concentration monitoring data file exported by the clinical record system, the file header redundant information is identified and skipped, only the effective data area is loaded, the table header misplacement problem is prevented, and the record fragmentation and association fracture problems in the clinical data are solved through a three-level processing procedure.

[0027] First, to address the issue of abnormal drug cycle identification in clinical records within drug concentration monitoring data files, a sliding window mechanism is established to traverse all record rows and verify whether the end time of the previous cycle and the start time of the next cycle for the same patient are continuous. If a time series is broken, it is identified as an abnormal segment, and the abnormal segment point and all redundant records in the same cycle traced upwards are removed. Specifically, based on the break point, all related cycle records are traced upwards, fragmented record chains are deleted in batches, and the row index is reset to construct a seamless treatment timeline, restoring the actual treatment process.

[0028] Secondly, in response to the clinical record characteristics of the same treatment cycle being scattered across multiple lines in the drug concentration monitoring data file, a hierarchical replication mechanism is established. The first complete cycle record is identified as the master node, and subsequent records with consecutive cycle identifiers are identified as subordinate nodes. The key field values ​​of the scattered lines are integrated into the master record, and the treatment plan parameters of the master record are automatically copied to the subsequent concentration detection lines. This establishes a mapping relationship between the treatment plan and blood drug concentration monitoring, enabling the vertical aggregation of multiple dosing records for the same patient.

[0029] Finally, to address the issue that the data structure of drug concentration monitoring data files does not meet the requirements for pharmacokinetic analysis, the data structure was reconstructed. Date, dosage, and duration were defined as treatment parameters; metabolite concentration values ​​were defined as drug concentration labels; and individual patient indicators and biochemical test indicators were used as clinical features. The column sequence was reconstructed according to the medical analysis logic of basic treatment parameters, clinical features, and drug effect concentration. A standard two-dimensional table was formed by horizontally linking the dosing date with the drug concentration detection value. A backtracking deletion mechanism was used to ensure temporal continuity. An explicit association between dosing and concentration detection was established through field copying, and null values ​​were filled to generate a structured dataset for pharmacokinetic analysis, thereby significantly reducing data redundancy and improving analysis efficiency.

[0030] S2. Construct a drug metabolism concentration prediction model. In this model, pharmacokinetic parameters are learned from a structured dataset using a neural network and then calculated using a pharmacokinetic model to obtain the predicted concentration. The drug metabolism concentration prediction model is then trained based on the structured dataset.

[0031] In this embodiment, outliers are first removed from the structured dataset, including missing dose records, invalid concentration values, and samples containing missing values. Based on clinical indicators, patients... Each indicator feature z-score normalization is used to make it unaffected by dimensions, and the scalar eigenvectors are converted into tensor formats, as shown in the following formula: Features The mean, Features Standard deviation: , wherein, and respectively represent the characteristic values before and after standardization.

[0032] For drugs with simple metabolism, a one-compartment model is considered for prediction, which can be calculated efficiently. For drugs with obvious tissue distribution, a two-compartment model is considered for prediction, which has higher precision. For long-acting or complex distribution drugs, a three-compartment model is considered for prediction, which is closest to the physiological characteristics. Based on the three pharmacokinetic models, combined with machine learning and deep learning technology, a neural network is constructed for training of patient clinical characteristics, a modular processor array design is adopted, including multiple independent processing units, each unit performs double-stage operation, GELU activation function realizes nonlinear transformation, and a full connection layer maps 64-dimensional features to scalar parameters. Different designs are made for one-compartment model, two-compartment model and three-compartment model. Through this parallel parameter processor structure, drug metabolism concentration prediction is realized, which can overcome the poor interpretability of traditional "black box" model and significantly reduce the clinical prediction error.

[0033] For the one-compartment model, two neural network channels with the same structure but independent parameters are constructed, each channel corresponds to a pharmacokinetic parameter, avoiding unnecessary coupling between parameters. The neural network structure of each channel includes: first, the input layer receives the standardized patient clinical feature vector As the input signal of neurons, Dropout layer is used to randomly discard a part of neurons to reduce the complex co-adaptation relationship between neurons, thereby improving the generalization ability of the model and preventing overfitting. LayerNorm layer is used to standardize the neuron output to eliminate the magnitude difference between samples. GeLU activation function ( is a standard normal cumulative distribution function) to realize nonlinear transformation, hidden layer connection weight matrix, and again use Dropout layer, LayerNorm layer and GeLU activation function for model complication, and finally output a single neuron. The original parameter values of the neuron outputs of the two channels and are respectively subjected to exponential transformation, and the numerical value is constrained to be positive to ensure the biological rationality that the parameter is greater than zero, and the output distribution volume and elimination rate , the formula is as follows: , , For each sample at each time point and the dose , the predicted concentration is calculated by the pharmacokinetic model : , The total loss LOSS is calculated as shown below, where is an L2 norm penalty term for all weight parameters of the neural network to reduce overfitting, and the weight coefficient is set to 0.01. Based on the AdamW optimizer, the model is trained by iteratively conducting gradient descent through backpropagation, and the mean square error (MSE) of the validation set is continuously monitored. If the validation loss does not decrease to the historical minimum value for 20 consecutive iterations, the training is terminated, and the model parameters with the lowest validation loss are automatically rolled back: , where, denotes the total number of samples, and denote the true concentration value and the predicted concentration value of the i-th sample, respectively.

[0034] For the two-compartment model, four neural network channels with the same structure but independent parameters are constructed, each corresponding to a pharmacokinetic parameter. The neural network structure of each channel includes: first, the input layer receives the standardized patient clinical feature vector as the input signal of the neuron, the LayerNorm layer is used to perform layer normalization on the neuron output to eliminate the magnitude difference between samples, the GeLU activation function ( is the standard normal cumulative distribution function) is used to realize nonlinear transformation, the hidden layer is connected with the weight matrix, and the LayerNorm layer and GeLU activation function are used again to complicate the model, and finally a single neuron is output. The four channels process the input data synchronously and output the original parameter values ~ respectively. For the distribution volume , elimination rate and , a positive constraint mechanism is implemented, as shown in the following formula: , , , For the inter-compartment weight factor , the sigmoid algorithm is used to make the weight value belong to the range (0, 1), and based on the constraint that the inter-compartment weight sum is 1, the inter-compartment weight factor is calculated, as shown in the following formula: , For each sample, each time point and the dose ​​ The total loss LOSS is calculated in the same way as the one-compartment model for subsequent model training.

[0035] For the three-compartment model, seven neural network channels with the same structure but independent parameters are constructed, and each channel corresponds to a pharmacokinetic parameter. The subsequent neural network processing is the same as that of the two-compartment model, and the input data are processed synchronously by the 7 channels, and the original parameters are output respectively The distribution volume Vd, the elimination rate Kel, the inter-compartment weight factor a21, a31 and a12 are implemented with positive constraint mechanism, and the formula is as follows: The weight factor a21, a31 and a12 are implemented with softmax algorithm for weight sum of 1 and weight value belonging to the range of (0, 1), and the formula is as follows: The weight factor a21, a31 and a12 are implemented with softmax algorithm for weight sum of 1 and weight value belonging to the range of (0, 1), and the formula is as follows: For each sample at each time point t and the dose D, the predicted concentration C is calculated by the pharmacokinetic model The total loss LOSS is calculated in the same way as the one-compartment model for subsequent model training. S3, based on the constructed pharmacokinetic concentration prediction model, the drug concentration at different sampling points is calculated according to the clinical characteristics and the dose of the patient, and the individual patient drug concentration metabolism curve is drawn.

[0036] In the embodiment, the clinical characteristics of the patient whose curve needs to be generated and the injected drug dose are input, the model type to be used (one-compartment model, two-compartment model, three-compartment model) is selected, the patient characteristics are standardized and input into the model, the parameter value required by the model is output, and the pharmacokinetic curve half cycle is calculated, and the formula is as follows,

[0037] ​​​​​​​​​​​​​​​​​​​​​Based on the parameter values, 5 sampling points of half-life are generated, that is The drug concentration at different sampling points is calculated based on the formula of different models: , Based on the sampling points and predicted concentrations, the individual patient drug concentration metabolic curve is drawn.

[0038] S4, based on the constructed drug metabolic concentration prediction model, the drug dose and drug interval are calculated according to the clinical characteristics of the patient, the target drug duration and the target concentration to realize individual drug decision-making.

[0039] In the embodiment, the pre-trained pharmacokinetic model and patient clinical feature standardizer are loaded, the patient clinical features are input, the specified model is input after feature standardization, and the predicted individual pharmacokinetic parameters are output. Based on the pharmacokinetic parameters, the target drug duration The target concentration The drug dose needed to achieve the target concentration is calculated The calculation formula is as follows, when intravenous injection F =1: , Based on the minimum effective concentration and the maximum safe concentration of different drug categories, the most appropriate drug interval is calculated based on the half-life , which is shown as follows: , The calculated drug dose and drug interval are used as a reference for doctors to make individual drug decisions.

[0040] S5, based on the constructed drug metabolic concentration prediction model, the drug time is calculated according to the clinical characteristics of the patient, the drug dose, the drug interval and the drug concentration threshold to realize drug concentration monitoring.

[0041] In the embodiment, the patient clinical features, drug dose, drug interval and drug concentration threshold are input, the specified model is input after feature standardization, and the predicted individual pharmacokinetic parameters are output. Based on the pharmacokinetic model and parameters, the numerical solution of time under the specified drug concentration threshold can be calculated using the Newton iteration method, which is used for drug concentration monitoring.

[0042] S6, all the above algorithms and data are visualized and deployed to the front-end interface to realize the login interface, data uploading function, data query function, data downloading function, selection of model and drug category, individual patient drug concentration metabolic curve display, individual drug decision-making and drug concentration monitoring function.

[0043] In the embodiment, the front end is built by Layui and TypeScript to construct an adaptive single-page application, and concurrent initialization of the multi-pathology prediction component is realized. The backend is developed based on the Django framework, and provides sequential and parallel model inference interfaces to adapt to different user scenarios. The prediction engine uses TorchServe to host the neural network model, and supports drug metabolism kinetic model and multi-category drug prediction functions. Specifically, the front-end visualization interface is as shown in Figures 2-10 .

[0044] In a specific application, the account password is first input to log in to the main page (as shown in Figure 2 ), and the processed csv data is uploaded to the system through the data upload function (as shown in Figure 3 ). The data uploaded successfully can be queried for the corresponding drug injection record of the patient ID (as shown in Figure 4 ), and downloaded to the computer through the data download method (as shown in Figure 5 ).

[0045] After the data is uploaded, the drug type (including methotrexate, vancomycin, topiramate, alanine, and oxcarbazepine) and model type (including one-compartment model, two-compartment model, and three-compartment model) to be predicted can be selected, and different algorithms are used for different drugs to start drug metabolism concentration prediction (as shown in Figure 6 ). After the prediction is completed, the drug metabolism concentration curve of each patient can be viewed to view the metabolism process of the drug concentration changing with time (as shown in Figure 7 and Figure 8 ). And by inputting the patient ID, the individual drug administration decision recommendation can be realized based on the patient's basic information stored in the database (as shown in Figure 9 ), and the drug concentration monitoring can be realized by inputting the drug administration mode and patient information (as shown in Figure 10 ).

[0046] Based on the same inventive concept, as shown in Figure 11 , the embodiment of the present application also provides a precision medicine mode construction system 110 based on pharmacokinetics and artificial intelligence, comprising: a data set construction module 111, a pharmacokinetic prediction model modeling module 112, a pharmacokinetic curve analysis module 113, an individual drug administration decision module 114, a drug concentration monitoring module 115, and a visualization interaction module 116.

[0047] The data set construction module 111 is used for multi-level preprocessing of the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis.

[0048] The pharmacokinetic prediction model modeling module 112 is configured to construct a drug metabolism concentration prediction model, wherein the pharmacokinetic parameters are learned and generated from the structured data set by using a neural network, the predicted concentration is calculated by a pharmacokinetic model, and the drug metabolism concentration prediction model is trained based on the structured data set.

[0049] The pharmacokinetic curve analysis module 113 is configured to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolism curve according to the clinical characteristics and the drug dose of the patient based on the constructed drug metabolism concentration prediction model.

[0050] The individual drug administration decision module 114 is configured to calculate the drug dose and the drug interval to achieve the individual drug administration decision according to the clinical characteristics, the target drug administration time length and the target concentration of the patient based on the constructed drug metabolism concentration prediction model.

[0051] The drug concentration monitoring module 115 is configured to calculate the drug administration time to achieve the drug concentration monitoring according to the clinical characteristics, the drug dose, the drug interval and the drug concentration threshold of the patient based on the constructed drug metabolism concentration prediction model.

[0052] The visual interaction module 116 is configured to visualize all the algorithms and data to the front-end interface to realize the login interface, the data uploading function, the data query function, the data downloading function, the selection of the model and the drug type, the individual patient drug concentration metabolism curve display, the individual drug administration decision judgment and the drug concentration monitoring function.

[0053] It should be noted that the precise medication mode construction system based on pharmacokinetics and artificial intelligence provided by the above embodiment and the precise medication mode construction method based on pharmacokinetics and artificial intelligence belong to the same inventive concept, and the specific implementation process is described in detail in the precise medication mode construction method based on pharmacokinetics and artificial intelligence, which will not be repeated here.

[0054] The specific embodiments described above have explained the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not used to limit the present application. Any modification, supplement and equivalent replacement within the principle range of the present application should be included in the protection scope of the present application.

Claims

1. A method for constructing a precision medicine model based on pharmacokinetics and artificial intelligence, characterized in that, The method comprises the following steps: Multi-level preprocessing of the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis; Building a drug metabolism concentration prediction model, wherein the neural network is used to learn and generate pharmacokinetic parameters from the structured data set, and the predicted concentration is calculated through the pharmacokinetic model, and the drug metabolism concentration prediction model is trained based on the structured data set; Based on the constructed drug metabolism concentration prediction model, the drug concentration at different sampling points is calculated according to the clinical characteristics and the dose of the patient, and the individual patient drug concentration metabolism curve is drawn; Based on the constructed drug metabolism concentration prediction model, the dose and the interval of the drug are calculated according to the clinical characteristics, the target drug duration and the target concentration of the patient to realize individual drug decision-making; Based on the constructed drug metabolism concentration prediction model, the drug time is calculated according to the clinical characteristics, the dose of the drug, the interval of the drug and the drug concentration threshold to realize drug concentration monitoring; All the above algorithms and data are visualized and deployed to the front-end interface to realize the login interface, data uploading function, data query function, data downloading function, selection of model and drug type, individual patient drug concentration metabolism curve display, individual drug decision-making and drug concentration monitoring function. 2.The method of claim 1, wherein the method comprises: collecting a plurality of pharmacokinetic data of a plurality of drugs from a plurality of patients; and training a pharmacokinetic model using the collected pharmacokinetic data. The multi-level preprocessing of the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis comprises: For the abnormal problem of drug cycle identification in the clinical record in the drug concentration monitoring data file, a sliding window mechanism is established to traverse all record lines, and it is verified whether the end time of the previous cycle and the start time of the next cycle of the same patient are continuous. If the time sequence is broken, it is determined as an abnormal separation, and the abnormal separation point and all redundant records of the same cycle traced upward are removed; For the clinical record characteristics of the same treatment cycle scattered in multiple lines in the drug concentration monitoring data file, a hierarchical structure replication mechanism is established to integrate the key field values of the scattered lines to the main record, realizing the longitudinal aggregation of the multiple drug records of the same patient; For the case that the data structure of the drug concentration monitoring data file does not meet the requirements of pharmacokinetic analysis, the data structure is reconstructed, the drug date and drug concentration detection value are horizontally associated to form a standard two-dimensional table, the backtracking deletion mechanism is used to ensure the time continuity, the explicit association of drug and concentration detection is established through field replication, and the structured data set for pharmacokinetic analysis is generated by filling the null values. 3.The method of claim 1, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The drug metabolism concentration prediction model comprises: One-compartment model, corresponding to two pharmacokinetic parameters, volume of distribution , elimination rate where the pharmacokinetic model is represented by: ， wherein, denotes the time predicted concentration, denotes the dose administered; two-compartment model, corresponding to four pharmacokinetic parameters, volume of distribution , elimination rate and , inter-compartment weight factor wherein the pharmacokinetic model is represented by: ， wherein represents an inter-chamber weight factor, and ; three-compartment model, corresponding to seven pharmacokinetic parameters, volume of distribution , elimination rate , and , inter-compartment weight factor , and , wherein the pharmacokinetic model is represented by: ， wherein . 4.The method of claim 3, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The use of neural network to learn and generate pharmacokinetic parameters from the structured data set, and the predicted concentration is calculated through the pharmacokinetic model, comprises: The neural network is constructed to train the clinical characteristics of the patient in the structured data set, each neural network corresponds to generate a pharmacokinetic parameter, and the neural network structure of each channel comprises: For a one-compartment model, first, the input layer receives the standardized patient clinical feature vector as the input signal of the neuron, a portion of the neurons is randomly discarded by the Dropout layer, the neuron output is layer-normalized by the LayerNorm layer, the non-linear transformation is realized by the GeLU activation function, the hidden layer is connected with the weight matrix, and the model is complicated again by using the Dropout layer, the LayerNorm layer and the GeLU activation function, and finally a single neuron is output. The value output by each neuron corresponds to the generated pharmacokinetic parameter, and the predicted concentration is calculated according to the generated pharmacokinetic parameter through the pharmacokinetic model; For a two-compartment model or a three-compartment model, first, the input layer receives the standardized patient clinical feature vector as the input signal of the neuron, the neuron output is layer-normalized by the LayerNorm layer, the non-linear transformation is realized by the GeLU activation function, the hidden layer is connected with the weight matrix, and the model is complicated again by using the LayerNorm layer and the GeLU activation function, and finally a single neuron is output. The value output by each neuron corresponds to the generated pharmacokinetic parameter, and the predicted concentration is calculated according to the generated pharmacokinetic parameter through the pharmacokinetic model. 5.The method of claim 4, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The value output by each neuron corresponds to the generated pharmacokinetic parameter, including: For the distribution volume and elimination rate of pharmacokinetics, the original parameter value output by the neuron is exponentially transformed, so that the value is positive, to obtain the corresponding pharmacokinetic parameter; For the weight factor of pharmacokinetics, the neuron output value is transformed by using the sigmoid algorithm or the softmax algorithm, so that the weight factor value is positive and is constrained in the range of 0-1. 6.The method for constructing a precision medicine model based on pharmacokinetics and artificial intelligence according to claim 3, characterized in that, For a simple metabolized drug, a one-compartment model is used for prediction, for a drug with obvious tissue distribution, a two-compartment model is used for prediction, and for a long-acting or complex distributed drug, a three-compartment model is used for prediction. 7.The method of claim 1, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The drug metabolism concentration prediction model is constructed based on the clinical characteristics and the drug dose of the patient to calculate the drug concentration at different sampling points and draw the individual patient drug concentration metabolism curve, including: After the patient's clinical characteristics are standardized, they are input into the drug metabolism concentration prediction model, the required pharmacokinetic parameters are calculated by the neural network prediction output model, and a number of sampling points based on the pharmacokinetic curve half-life are generated. The pharmacokinetic parameters and the drug dose are input into the pharmacokinetic model to calculate the predicted concentration, and the individual patient drug concentration metabolism curve is drawn based on the sampling points and the predicted concentration. 8.The method of claim 1, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The drug metabolism concentration prediction model is constructed based on the clinical characteristics, target drug administration time and target concentration of the patient to calculate the drug dose and drug interval to realize individual drug administration decision, including: The clinical characteristics of the patient are input into the drug metabolism concentration prediction model after data standardization, the required pharmacokinetic parameters are calculated by the neural network prediction output model, the pharmacokinetic parameters, target drug administration time and target concentration reached in the time are input into the pharmacokinetic model to calculate the required drug administration dose and drug administration interval as a reference for the doctor to make individual drug administration decisions. 9.The method of claim 1, wherein the method comprises: determining a pharmacokinetic model of the patient based on the pharmacokinetic data; and determining a pharmacodynamic model of the patient based on the pharmacokinetic data and the pharmacodynamic data. The drug metabolism concentration prediction model is constructed, and the drug administration time is calculated according to the clinical characteristics of the patient, the drug administration dose, the drug administration interval and the drug concentration threshold to achieve drug concentration monitoring, which comprises: The clinical characteristics of the patient are input into the drug metabolism concentration prediction model after data standardization, the required pharmacokinetic parameters are calculated by the neural network prediction output model, the pharmacokinetic parameters, drug administration dose, drug administration interval and drug concentration threshold can be calculated by Newton iteration method to obtain the numerical solution of the time under the specified drug concentration threshold, which is used for drug concentration monitoring.

10. A precision medicine mode construction system based on pharmacokinetics and artificial intelligence, which is realized by using the precision medicine mode construction method based on pharmacokinetics and artificial intelligence according to any one of claims 1-9. It comprises: A data set construction module, a pharmacokinetic prediction model modeling module, a pharmacokinetic curve analysis module, an individual drug administration decision module, a drug concentration monitoring module, and a visual interaction module; The data set construction module is used for multi-level preprocessing of the obtained drug concentration monitoring data file to obtain a structured data set for pharmacokinetic analysis; The pharmacokinetic prediction model modeling module is used for constructing a drug metabolism concentration prediction model, wherein the neural network is used to learn and generate pharmacokinetic parameters from the structured data set, and the pharmacokinetic model is used to calculate the predicted concentration, and the drug metabolism concentration prediction model is trained based on the structured data set; The pharmacokinetic curve analysis module is used for calculating the drug concentration at different sampling points based on the constructed drug metabolism concentration prediction model according to the clinical characteristics of the patient and the drug administration dose, and drawing the individual patient drug concentration metabolism curve; The individual drug administration decision module is used for calculating the drug administration dose and drug administration interval based on the constructed drug metabolism concentration prediction model according to the clinical characteristics of the patient, the target drug administration time and the target concentration to achieve individual drug administration decision; The drug concentration monitoring module is used for calculating the drug administration time based on the constructed drug metabolism concentration prediction model according to the clinical characteristics of the patient, the drug administration dose, the drug administration interval and the drug concentration threshold to achieve drug concentration monitoring; The visual interaction module is used for visualizing all the algorithms and data to the front-end interface to realize the login interface, data uploading function, data query function, data downloading function, model and drug selection, individual patient drug concentration metabolism curve display, individual drug administration decision judgment and drug concentration monitoring function.

Images