Therapeutic drug monitoring system using collaborative multilayer perceptron model

Abstract

The present invention relates to a therapeutic drug monitoring system in which information about a patient who is administered a drug requiring therapeutic drug monitoring is input to a deep learning model generated by learning a collaborative multilayer perceptron model in which multiple multilayer perceptrons are configured in parallel but trained simultaneously, and the system predicts the drug concentration of the patient to which the drug is administered, or determines the appropriate drug administration concentration and timing for each individual patient.

Description

[Technical Field]

[0001] The present invention relates to a therapeutic drug monitoring system that utilizes a collaborative multilayer perceptron model with a novel architecture to predict drug concentrations in patients administered drugs such as vancomycin and aminoglycosides, or to determine appropriate drug administration concentrations and timing for individual patients. [Background technology]

[0002] There are many drugs used in the treatment of patients for which the correct dosage is extremely important, and for these drugs for which the correct dosage is important, therapeutic drug monitoring (TDM) is essential when administering the drug.

[0003] Vancomycin is a typical drug requiring TDM. Vancomycin is a glycopeptide antibiotic that binds to D-Ala-D-Ala amino acids at the cross-linking points of peptides, inhibiting the transport of small peptide units in peptidoglycan, thereby exerting its bactericidal effect. Vancomycin's primary use is to kill multidrug-resistant bacteria, most notably MRSA (Methicillin-resistant Staphylococcus aureus) infections. MRSA is a highly virulent bacterium that can kill hospitalized patients. It is resistant to methicillin, a next-generation penicillin antibiotic, and vancomycin is used to treat it.

[0004] However, as mentioned above, vancomycin has a narrow therapeutic range and individual differences in pharmacokinetic parameters, so therapeutic drug monitoring is used to minimize toxicity and improve therapeutic efficacy.

[0005] When vancomycin is administered, it must pass through the pharmacokinetic (PK) and pharmacodynamic (PD) phases before it reaches its receptor site in the body and exerts its therapeutic effect. During the PK phase, significant individual differences in drug response result in varying drug concentrations in individual patients. Actual drug concentrations are affected by patient factors (age, weight, liver function, renal function, disease state, idiosyncrasies, genetics, nutritional status, metabolic capacity) and drug factors (physicochemical properties of the drug, pharmaceutical properties of the product, dosage, tolerance, drug interactions, pharmacokinetics, dosage form). Therefore, determining the optimal vancomycin dosage and administration interval for each patient using previously proposed methods is not easy. Furthermore, repeated blood sampling to monitor drug concentrations can be a burden for critically ill patients.

[0006] Currently, the most commonly used method for estimating drug concentrations required for TDM is Bayesian statistics. Bayesian statistics uses known population pharmacokinetic variables to estimate the pharmacokinetic variables (Vd, ke, etc.) of the population to which a patient belongs based on basic patient information (serum creatinine, body weight, etc.) and the dose and method of administration of the administered drug. It then creates a simulation graph and predicted steady-state peak and trough values, compares the predicted steady-state peak and trough values ​​with one or two actual measurements, and a clinical pharmacist or other expert analyzes and notifies the patient of whether the concentration is appropriate.

[0007] Since Bayesian statistics is a statistical approach, it has low accuracy for individual patients, and the accuracy is particularly low for patients who deviate from the statistics, for example, patients in abnormal conditions such as critically ill patients. Also, Bayesian statistics has the disadvantage that it is virtually impossible to use new variables because additional statistics must be collected in order to use new variables. Summary of the Invention [Problem to be solved by the invention]

[0008] It is an object of the present invention to provide a therapeutic drug monitoring system that utilizes a collaborative multilayer perceptron model with a novel architecture to predict drug concentrations in patients administered drugs requiring therapeutic drug monitoring (TDM) or to determine appropriate drug administration concentrations and timing for individual patients.

[0009] On the other hand, other objects of the present invention not explicitly stated should be considered within the scope that can be easily inferred from the following detailed description and other effects. [Means for solving the problem]

[0010] In order to achieve the above-mentioned object, the following solution is proposed.

[0011] A therapeutic drug monitoring system according to one embodiment of the present invention is configured to input information about patients who are administered drugs requiring therapeutic drug monitoring into a deep learning model generated by learning a collaborative multilayer perceptron model in which multiple multilayer perceptrons are configured in parallel but trained simultaneously, and predicts the drug concentration of the administered patient or determines the appropriate drug administration concentration and timing for each individual patient.

[0012] In one embodiment, the multi-layer perceptron has a plurality of hidden layers, and a masked attention module is arranged in front of at least one of the hidden layers to further focus on important information and reduce the influence of relatively less meaningful information.

[0013] In one embodiment, the input data of the training data used to train the joint multi-layer perceptron model may be at least one of the total drug dose, the initial drug dose, the total number of drug injections, the drug dose per injection, the average drug injection interval, the interval between the start of drug administration and the measurement of blood drug concentration, age, sex, height, weight, blood creatinine level, whether dialysis is possible, and the amount of blood transfusion between the start of drug administration and the measurement of blood drug concentration, and the output data of the training data may be blood drug concentration.

[0014] In one embodiment, input data for training the joint multi-layer perceptron model may include glomerular filtration rate.

[0015] In one embodiment, the training process of the joint multilayer perceptron model may be characterized by simultaneously inputting training data to each of a plurality of multilayer perceptrons constituting the joint multilayer perceptron model, outputting output data from each multilayer perceptron, combining the output data from each multilayer perceptron to output final output data, and once the final output data has been output, calculating gradients of each multilayer perceptron according to a loss function to train weights. [Effects of the Invention]

[0016] A therapeutic drug monitoring system according to one embodiment of the present invention monitors drug concentrations using a collaborative multilayer perceptron model in which multiple multilayer perceptrons are configured in parallel but trained simultaneously, and can predict drug concentrations more accurately than therapeutic drug monitoring systems that use Bayesian statistics and other artificial intelligence models.

[0017] Furthermore, the therapeutic drug monitoring system according to one embodiment of the present invention uses as learning data variables the total drug dose, the initial drug dose, the total number of drug injections, the drug dose per injection, the average drug injection interval, the interval between the start of drug administration and the measurement of blood drug concentration, age, sex, height, weight, blood creatinine level, blood drug concentration, dialysis, and the amount of blood transfusion between the start of drug administration and the measurement of blood drug concentration. In particular, by using the glomerular filtration rate (eGFR) derived from the blood creatinine level as a variable, more accurate drug concentration predictions can be made.

[0018] On the other hand, even if the effects are not explicitly mentioned here, the effects and provisional effects described in the following specification that are expected by the technical features of the present invention shall be treated as described in the specification of the present invention. [Brief explanation of the drawings]

[0019] [Figure 1] The present invention relates to a therapeutic drug monitoring system, a method for monitoring a drug concentration in a therapeutic drug monitoring apparatus, a method for monitoring a drug concentration in a therapeutic drug monitoring apparatus, a method for monitoring a drug concentration in a therapeutic drug monitoring apparatus, and a method for monitoring a drug concentration in a therapeutic drug monitoring apparatus. DETAILED DESCRIPTION OF THE INVENTION

[0020] Hereinafter, the configurations of the present invention and the effects thereof will be described with reference to the drawings, which illustrate various embodiments of the present invention. In describing the present invention, detailed descriptions of related known functions will be omitted if they are obvious to those skilled in the art and are deemed to obscure the gist of the present invention.

[0021] The term "module" used in this document may include a unit implemented from hardware, software, or firmware, and may be used interchangeably with terms such as logic, logic block, component, or circuit. A module may be an integrally configured component or the smallest unit or part of such component that performs one or more functions.

[0022] In this document, a "module" or a "node" refers to a device that performs tasks such as moving, storing, and converting data using a computing device such as a CPU, AP, etc. For example, a "module" or a "node" may be embodied as a device such as a server, PC, tablet PC, smartphone, etc.

[0023] In this document, the term "deep learning model" means a neural network that is an algorithm modeled on the workings of the human brain and is constructed by learning neural networks, and is to be broadly interpreted as the meaning commonly used in the industry.

[0024] FIG. 1 is a diagram illustrating the general architecture of an intensity multilayer perceptron model of a therapeutic drug monitoring system according to one embodiment of the present invention.

[0025] A therapeutic drug monitoring system according to one embodiment of the present invention is for predicting drug concentrations in patients who have been administered drugs or determining appropriate drug administration concentrations and timing for individual patients, and utilizes a deep learning module generated by training a newly proposed joint multilayer perceptron model.

[0026] Here, the drug may be at least one selected from the group consisting of antiepileptics (anticonvulsants), antiarrhythmics, bronchodilators (antiashmatic drugs), antibiotics (antibiotics), antipsychotics (antimanics (mood stabilizers)), anticancer drugs (antineoplastics), and immunosuppressants (immunosuppressants). More specifically, the antiepileptic drug is at least one selected from the group consisting of phenobarbital, phenytoin (DILANTIN), carbamazepine (TEGRETOL), ethosuximide, primidone, and valproic acid (VPA); the antiarrhythmic drug is at least one selected from the group consisting of digoxin, digitoxin (a cardiac inotropic agent), quinidine, procainamide, and lidocaine; the bronchodilator is at least one selected from the group consisting of theophylline and caffeine; the antibiotic is at least one selected from the group consisting of aminoglycosides (gentamycin, tobramycin, amikacin) and vancomycin; the antipsychotic drug is at least one selected from the group consisting of imipramine, desipramine (antidepressant), pimozide, and clozapine; and the antimanic drug is lithium. The anti-cancer drug may be methotrexate, and the immunosuppressant may be at least one selected from the group consisting of cyclosporine, mycophenolic acid (CELLCEPT, MYCOFORTIC), tacrolimus (PROGRAF), and sirolimus (RAPAMUNE), although the present invention is not limited thereto.

[0027] The joint multilayer perceptron model proposed in the present invention is configured with multiple multilayer perceptrons in parallel, but trained simultaneously. For example, the multilayer perceptron is configured to perform feedforward propagation, in which values ​​transmitted from the input layer are transmitted to all nodes in the hidden layer, and output values ​​from all nodes in the hidden layer are also transmitted to all nodes in the output layer. The multilayer perceptrons constituting the joint multilayer perceptron of the present invention may have a known structure, but the present invention is not limited to a specific structure of the multilayer perceptron. The multilayer perceptron (MLP) used in the joint multilayer perceptron model proposed in the present invention is configured with multiple (e.g., three) hidden layers with a hidden unit size of 32 and using leakyReLU as an activation function, and the multilayer perceptrons (MLPs) are configured to learn by focusing on different inputs. Meanwhile, the joint multilayer perceptron model of the present invention may be configured with at least two multilayer perceptrons in parallel.

[0028] In addition, the joint multilayer perceptron model proposed in the present invention may further include a mask attention module before at least one of the multiple hidden layers constituting the multilayer perceptron model. The mask attention module is a module that focuses more on important information and reduces the influence of relatively less significant information, and is a module that learns weights through learning. The mask attention module can be placed before each hidden layer, but in this patent document, the mask attention module is placed before the first hidden layer of the multilayer perceptron.

[0029] Meanwhile, in the field of deep learning, techniques such as ensemble bagging and ensemble boosting have been proposed. The ensemble bagging technique involves creating several single models, randomly extracting data through a booststrapping process, training the models, and finally predicting data through a voting process. The ensemble boosting technique is similar to bagging in that it extracts initial sample data and generates multiple classifiers, but differs in that it weights the parts of the previous model where errors occurred during the training process. Both ensemble bagging and ensemble boosting techniques have in common that the training process is performed sequentially. However, the joint multilayer perceptron model proposed in the present invention differs from conventional ensemble techniques in that, although an architecture is constructed using multiple different multilayer perceptrons, the different multilayer perceptrons simultaneously train the input training data.

[0030] Furthermore, compared to a multilayer perceptron model with a deeper layer depth (e.g., a 300-layer multilayer perceptron), using a shallower collaborative multilayer perceptron (e.g., 100 three-layer multilayer perceptrons connected in parallel) significantly improves the accuracy of drug concentration prediction.

[0031] In particular, the therapeutic drug monitoring system of the present invention has the advantage of demonstrating significantly improved drug concentration prediction accuracy even in cases of high heterogeneity such as critically ill patients by utilizing a collaborative multilayer perceptron model.

[0032] A therapeutic drug monitoring system according to one embodiment of the present invention uses, as input data for training data, at least one of patient information including the total drug dose, the initial drug dose, the total number of drug infusions, the drug dose per infusion, the average drug infusion interval, the interval between the start of drug administration and the measurement of blood drug concentration, age, sex, height, weight, blood creatinine level, whether or not dialysis is required, and the amount of blood transfusion between the start of drug administration and the measurement of blood drug concentration. Furthermore, blood drug concentration is used as output data for training data. In particular, a therapeutic drug monitoring system according to one embodiment of the present invention can more accurately predict drug concentration by using the glomerular filtration rate (eGFR) derived from the blood creatinine concentration as a variable in the training data. The glomerular filtration rate is calculated using the blood creatinine concentration using the Modification of Diet in Renal Disease (MDRD) method and / or the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula. Table 1 below summarizes the input and output data items used in a therapeutic drug monitoring system according to one embodiment of the present invention to evaluate performance.

[0033] [Table 1] For data normalization, the input data for the training data was scaled to the range of [-1, 1] using a min-max scaler, and gender and whether or not dialysis was available were represented as 0 or 1.

[0034] The process of training the joint multilayer perceptron model is as follows: Input data is simultaneously input to each of the multiple multilayer perceptrons that make up the joint multilayer perceptron model, output data is output from each multilayer perceptron, and the output data from each multilayer perceptron is summarized to produce final output data. Once the final output data is output, the gradient of each multilayer perceptron is calculated using a loss function, and weights are trained. As a result, each multilayer perceptron in the joint multilayer perceptron model converges on a different input, thereby improving the accuracy of blood drug concentration predictions in a therapeutic drug monitoring system according to one embodiment of the present invention.

[0035] The performance of a therapeutic drug monitoring system according to an embodiment of the present invention was evaluated using data from critically ill patients who were administered vancomycin.

[0036] The training and test data consisted of 2,406 critically ill patients admitted to the intensive care units (ICUs) of Dongguk University Ilsan Hospital (DUIH) and Kangwon National University Hospital (KNUH) from January 1, 2010 to February 28, 2022. The data from Dongguk University Ilsan Hospital was used as the internal validation dataset, containing 977 patients, while the data from Kangwon National University Hospital was used as the external validation dataset, containing 1,429 patients. The internal validation dataset consisted of 90% training data (879 cases) and 10% test data (98 cases). Patients were critically ill (>18 years old) with a history of vancomycin treatment and had undergone at least one therapeutic drug monitoring (TDM) test for vancomycin administration. In patients with normal renal function who underwent vancomycin TDM multiple times, only the first TDM value was selected and used as data. If the interval between discontinuation and re-administration of vancomycin was 2 weeks or longer, it was considered an independent TDM and used in the analysis. Patients who received oral vancomycin or were under 18 years of age were excluded. Total blood vancomycin concentrations were measured by fluorescent immunoassay (VANC3, Cobas c 702, Roche Diagnostics, IN, USA). The output data to be measured was the minimum blood vancomycin concentration.

[0037] The comparative models used were the PPK model (Comparative Example 1: Matzke GR, McGory RW, Halstenson CE, Keane WF. Pharmacokinetics of vancomycin in patients with various degrees of renal function. Antimicrob Agents Chemother. 1984 Apr;25(4):433-7. PMID:6732213. doi:10.1128 / AAC.25.4.433.), the XGBoost (extreme gradient boosting) model (Comparative Example 2), and the TabNet model (Comparative Example 3). The PPK model is a traditional statistical method. XGBoost is an open-source library for decision tree-based gradient boosting machine learning, known to work well with tabular data and is an important model for distributed training or normalization. The TabNet model is a deep learning model specialized for table datasets, and can optimize TabNet by automatically transforming input variables and selecting necessary variables using additional validation data.

[0038] Ten-fold cross-validation was performed to investigate the generalizability and expected error range of the models of the Examples and Comparative Examples.

[0039] Baseline variables and patient characteristics for the internal and external datasets were expressed as frequencies with percentages or means with standard deviations. Comparisons between datasets were performed using paired t-tests for continuous variables or chi-square tests for categorical variables. Measured serum vancomycin values ​​were used as actual values. The model predictive ability of the examples and comparative examples for the minimum vancomycin concentration was evaluated by calculating the mean absolute error (MAE), root mean square error (RMSE), R-squared, and adjusted R-squared to assess bias and precision. A paired t-test for RMSE was used to confirm significant differences in predictive performance between models [15, 16]. A P-value of ≤0.05 or ≤0.01 was considered statistically significant.

[0040] Table 2 below summarizes the patient characteristics of the internal data (training data and test data) and external data.

[0041] [Table 2] There were no significant differences in basic characteristics between the internal and external data, except that the external data contained slightly more elderly patients who were shorter and had slightly worse renal function calculated using the CKD-EPI method.

[0042] Table 3 below shows the results of therapeutic drug monitoring of vancomycin concentration using the comparative example and the example.

[0043] [Table 3] As can be seen from Table 3, the performance of Examples 1 and 2, which used a joint multilayer perceptron model, was significantly superior to the other comparative examples for all data. In particular, Comparative Example 4 used a 300-layer deep multilayer perceptron model, and it can be seen that Examples 1 and 2, which used 100 multilayer perceptrons with a depth of 3 layers in parallel, showed significantly better performance than Comparative Example 4.

[0044] On the other hand, the best performance was achieved in Example 2, in which the mask_attention module was placed before the first hidden layer of the multiple multilayer perceptrons that make up the joint multilayer perceptron model.

[0045] The performance of Examples 1 and 2 includes cases with abnormal patterns such as the high TDM values ​​observed in some severely ill patients, which means that the joint multilayer perceptron model of Examples 1 and 2 is more suitable for actual therapeutic drug monitoring of vancomycin.

[0046] In conclusion, a vancomycin therapeutic drug monitoring system according to one embodiment of the present invention uses a deep learning model formed by learning a joint multilayer perceptron model, in which multiple multilayer perceptrons are configured in parallel but trained simultaneously. That is, by inputting information about a patient who is administered a drug requiring therapeutic drug monitoring (e.g., the input data described above) into the generated deep learning model, it is possible to predict the drug concentration of the administered patient without having to draw blood multiple times from the patient, or to determine the appropriate drug administration concentration and timing for each individual patient.

[0047] The therapeutic drug monitoring system described above may be implemented as a program (or application) including a computer-executable algorithm or on the web. The program may be stored and provided on a non-transitory computer-readable medium. Here, a non-transitory computer-readable medium refers to a medium that stores data semi-permanently and is readable by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. More specifically, the various applications or programs described above may be stored and provided on a non-transitory computer-readable medium, such as a CD, DVD, hard disk, Blu-ray disc, USB, memory card, ROM, etc.

[0048] It should be reiterated that the scope of protection of the present invention is not limited to the description and expression of the embodiments explicitly described above, and that the scope of protection of the present invention is not limited by obvious modifications or substitutions in the technical field to which the present invention pertains.

Claims

1. A deep learning model is generated by learning a joint multilayer perceptron model in which multiple multilayer perceptrons are configured in parallel but trained simultaneously, and information on a patient who is administered a drug that requires therapeutic drug monitoring is input to the deep learning model to predict the drug concentration of the patient who has been administered the drug; The training process of the joint multilayer perceptron model comprises simultaneously inputting the same training data to a plurality of multilayer perceptrons constituting the joint multilayer perceptron model, outputting output data from each multilayer perceptron, combining the output data from each multilayer perceptron to output final output data, and once the final output data has been output, calculating the gradient of each multilayer perceptron according to a loss function to train weights, and each of the multilayer perceptrons is configured to train by focusing on a different input, and the multilayer perceptron has a plurality of hidden layers, and a masked attention module is arranged before a first hidden layer of the plurality of hidden layers to further focus on important information and reduce the influence of relatively less meaningful information.

2. 2. The therapeutic drug monitoring system of claim 1, wherein the input data for training the joint multilayer perceptron model is at least one of the total drug dose, the initial drug dose, the total number of drug injections, the drug dose per injection, the average drug injection interval, the interval from the start of drug administration to the measurement of blood drug concentration, age, sex, height, weight, blood creatinine level, whether dialysis is possible, and the amount of blood transfusion from the start of drug administration to the measurement of blood drug concentration, and the output data of the training data is blood drug concentration.

3. 10. The therapeutic drug monitoring system of claim 1, wherein input data for training the joint multi-layer perceptron model includes glomerular filtration rate.

Images