Individualized drug dose optimization method based on causal traceability

Abstract

The invention discloses an individualized drug dose optimization method based on causal traceability, which improves the accuracy and clinical adaptability of dose regulation and control through a multi-modal space-time alignment and feature decoupling technology, causal reasoning modeling and causal effect quantification and individualized dose optimization, can effectively reduce adverse reactions in the treatment process, and improves the treatment efficiency. The life cycle of a patient is prolonged, the life quality is improved, meanwhile, the long-term medication cost is reduced, and wide clinical application prospects and popularization values are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of individualized dose optimization in nuclear medicine, and relates to an individualized drug dose optimization method based on causal tracing. Background Art

[0002] Radioactive nuclear drugs are widely used in the diagnosis and treatment of diseases in nuclear medicine, especially in aspects such as thyroid cancer, brain imaging, and assessment of cardiac microcirculation. However, due to significant individual metabolic differences, there are large differences in the distribution and metabolism of radioactive nuclear drugs in the body. The current dose optimization models are mainly based on population averages, ignoring individual differences and prone to causing inaccurate treatment doses. Therefore, it is particularly important to establish a dose optimization method based on individual metabolic differences. The existing dose optimization methods face three main challenges: (i) lack of individualized factors, such as key parameters like body weight, renal function, and liver metabolism ability; (ii) lack of causal mechanism modeling, and traditional methods are difficult to reveal the causal relationship between dose, metabolic parameters, and pharmacodynamic response; (iii) insufficient utilization of spatio-temporal dynamic data, and most models ignore the dynamic changes of multimodal medical images in the time and space dimensions, resulting in limited accuracy of dose distribution estimation. Summary of the Invention

[0003] To solve these problems, the present invention provides an individualized drug dose optimization method based on causal tracing, which improves the accuracy and clinical adaptability of dose regulation through multimodal spatio-temporal alignment and feature decoupling technology, causal inference modeling and causal effect quantification, and individualized dose optimization. It can effectively reduce adverse reactions during the treatment process, extend the patient's survival period, improve the quality of life, and at the same time reduce the long-term medication cost, and has broad clinical application prospects and promotion value.

[0004] The technical solution of the present invention is as follows:

[0005] An individualized drug dose optimization method based on causal tracing improves the accuracy and clinical adaptability of dose regulation through multimodal spatio-temporal alignment and feature decoupling technology, causal inference modeling and causal effect quantification, and individualized dose optimization. It can effectively reduce adverse reactions during the treatment process, extend the patient's survival period, improve the quality of life, and at the same time reduce the long-term medication cost, and has broad clinical application prospects and promotion value.

[0006] Specifically, it includes the following steps:

[0007] (1) Obtain the patient's multimodal data, including PET-CT scan images I PET , I CT , blood biochemical indexes B, and physiological parameters P (such as body weight, age, renal function, etc.). PET is used to display the distribution of drugs, and CT is used to display the anatomical structure;

[0008] (2) Use the Transformer model to align I PET (t), I CT (t) in the time dimension to construct a unified spatio-temporal sequence S(t). These data are temporal, meaning they are collected at different time points. To analyze the metabolic process of radioactive drugs in the body, it is necessary to align these data so that data at different time points can be compared and combined to form a complete metabolic pathway. The core of the Transformer lies in the self-attention mechanism, which can learn the relationships between different time points and adjust the representations of different modality data according to these relationships. The formula is as follows:

[0009] S(t) = f Transformer (I PET (t), I CT (t))

[0010] During the spatio-temporal alignment process, the Transformer will compare the PET and CT images at each moment, learn the potential connections between them, and generate unified temporal information. f Transformer represents the Transformer model, and S(t) is a unified spatio-temporal sequence that combines PET and CT data, capable of containing both drug metabolism information and anatomical structure information;

[0011] (3) Use principal component analysis (PCA) and non-negative matrix factorization (NMF) to decouple S(t) to obtain the main metabolic feature matrix M and the noise feature matrix N. PCA is used to extract metabolic features and reduce the dimensionality, making the data representation more concise, which helps reduce the computational complexity and retain the main information of the data as much as possible. Then, NMF is used to ensure that the extracted metabolic feature matrix M and the noise matrix N satisfy the non-negative constraint, making the decoupling result more in line with the actual biological background. During decoupling, the obtained M matrix contains the main metabolic information, and the N matrix contains components related to noise, which is helpful for subsequent metabolic pathway analysis and dose optimization. The formula is as follows:

[0012] M = WH

[0013] S(t) = M + N

[0014] where W is the feature basis, H is the coefficient matrix, M represents the main features, and N represents noise or other minor components;

[0015] (4) Based on M and P, construct a Bayesian network G = (V, E). The node set V represents drug metabolism variables (such as absorption rate A, volume of distribution V d, clearance rate CL, etc.), the edge set E represents the causal paths between variables, M contains the metabolic features extracted from multimodal data, and P contains the individual data of the patient (such as weight, renal function, age, etc.). The construction of the causal relationship is based on the physiological principles of drug metabolism and is also adjusted and optimized by combining the individual data of the patient (such as age, weight, etc.);

[0016] (5) Identify the key causal variables related to dose decision using the Markov blanket algorithm For each variable in the drug metabolism model, calculate its Markov blanket. By analyzing each variable and its Markov blanket, determine which variables are closely related to the decision of drug dose (such as dose response, treatment effect, side effects, etc.);

[0017] (6) Construct the prior distribution P(V) by combining expert prior knowledge and use Monte Carlo simulation MCMC sampling to optimize the posterior distribution P(V|M,P). The prior distribution reflects our beliefs or assumptions about each variable (such as absorption rate, volume of distribution, clearance rate, etc.) in the drug metabolism process before actual data. This belief or assumption can be based on historical data, literature, knowledge of domain experts, etc. MCMC constructs a Markov chain, and the stationary distribution of this chain is the posterior distribution we need. MCMC generates new samples repeatedly from the current state (i.e., a certain value of the drug metabolism variable) and finally converges to the posterior distribution. The posterior distribution represents our updated beliefs about the drug metabolism variables after observing new data M (such as the metabolic feature matrix) and P (such as the individual data of the patient). The formula is as follows:

[0018]

[0019] where P(M,P|V) represents the probability of observing data M and P given the drug metabolism variable V, P(V) represents our belief about the drug metabolism variable without data, and P(M,P) is used to ensure that the posterior distribution is a valid probability distribution;

[0020] (7) In pharmacology, it is usually assumed that as the dose increases, the pharmacodynamic response will increase, but after exceeding a certain threshold, the pharmacodynamic effect may saturate or the side effects may increase. Therefore, we need to establish a dose-response function:

[0021] E = f(D, A, V d , CL)

[0022] where E represents the pharmacodynamic response, D is the administered dose, A, V d , CL are individual causal variables, the absorption rate (A) of the drug, volume of distribution (V d) Clearance rate (CL). These parameters determine how the drug is distributed, metabolized, and cleared in the body, thus directly affecting the drug efficacy.

[0023] (8) Set the objective function:

[0024] L(D) = α(E max - E(D)) 2 + βR(D)

[0025] where L(D) is the comprehensive function for evaluating at a given dose D, E(D) is the therapeutic effect produced at the given dose, R(D) is the side effect produced at the given dose. The particle swarm optimization (PSO) algorithm is used to iteratively search for the optimal dose D * , so as to maximize the therapeutic effect E(D) and minimize the side effect R(D). In dose optimization, each particle corresponds to a dose value, and each particle has two key attributes: position (corresponding to the drug dose) and velocity (indicating how the particle adjusts its own position). In each iteration, the particle updates its velocity and position. The particle updates its velocity and position according to its own historical optimal position and the optimal position of all particles, so as to continuously explore the possible best solution. Particle swarm optimization can handle complex multi-dimensional search spaces and can effectively avoid local optimal solutions, thus improving the reliability of dose optimization;

[0026] (9) Combine the optimal dose D* with the metabolic prediction curve and output an individualized dose recommendation report for clinical reference, including an individualized medication plan with dose recommendation values, metabolic abnormality warnings, and suggestions for adjusting the dosing interval.

[0027] Advantages of the present invention:

[0028] Through multi-modal spatio-temporal alignment and feature decoupling technology, causal inference modeling and causal effect quantification, and individualized dose optimization, the present invention improves the accuracy and clinical adaptability of dose regulation, can effectively reduce adverse reactions during the treatment process, prolong the patient's survival period, improve the quality of life, and at the same time reduce the long-term medication cost, and has broad clinical application prospects and promotion value. Description of the Drawings

[0029] Figure 1 is the execution flowchart of the present invention. Detailed Embodiments

[0030] The present invention will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0031] As Figure 1 shown, an individualized drug dose optimization method based on causal tracing specifically includes the following steps:

[0032] (1) Obtain multi-modal data of the patient, including the PET-CT scan image I PET , I CT , blood biochemical index B, physiological parameters P (such as body weight, age, renal function, etc.). PET is used to show the distribution of drugs, and CT is used to show the anatomical structure;

[0033] (2) Use the Transformer model to align I PET (t), I CT (t) in the time dimension to construct a unified spatio-temporal sequence S(t). These data are temporal, which means they are collected at different time points. In order to analyze the metabolic process of radioactive drugs in the body, it is necessary to align these data so that data at different time points can be compared and combined to form a complete metabolic path. The core of the Transformer lies in the self-attention mechanism, which can learn the mutual relationships between different time points and adjust the representations of different modal data according to these relationships. The formula is as follows:

[0034] S(t) = f Transformer (I PET (t), I CT (t))

[0035] During the spatio-temporal alignment process, the Transformer will compare the PET and CT images at each moment, learn the potential connections between them, and generate unified temporal information. f Transformer represents the Transformer model, and S(t) is a unified spatio-temporal sequence that integrates PET and CT data, which can contain both drug metabolism information and anatomical structure information;

[0036] (3) Use principal component analysis PCA and non-negative matrix factorization NMF to decouple S(t) to obtain the main metabolic feature matrix M and the noise feature matrix N. Extract metabolic features and reduce the dimension through PCA, making the representation of the data more concise, which helps to reduce the computational complexity and retain the main information of the data as much as possible. Then, use NMF to ensure that the extracted metabolic feature matrix M and the noise matrix N satisfy the non-negative constraint, making the decoupling result more in line with the actual biological background. During decoupling, the obtained M matrix contains the main metabolic information, and the N matrix contains components related to noise, which helps subsequent metabolic path analysis and dose optimization. The formula is as follows:

[0037] M = WH

[0038] S(t) = M + N

[0039] where W is the feature basis, H is the coefficient matrix, M represents the main features, and N represents noise or other secondary components;

[0040] (4) Based on M and P, construct a Bayesian network G=(V, E). The node set V represents drug metabolism variables (such as absorption rate A, volume of distribution V d , clearance rate CL, etc.), and the edge set E represents the causal paths between variables. M contains metabolic features extracted from multimodal data, while P contains individual patient data (such as body weight, renal function, age, etc.). The construction of causal relationships is based on the physiological principles of drug metabolism and is also adjusted and optimized in combination with individual patient data (such as age, body weight, etc.);

[0041] (5) Use the Markov blanket algorithm to identify key causal variables related to dose decision-making For each variable in the drug metabolism model, calculate its Markov blanket. By analyzing each variable and its Markov blanket, determine which variables are closely related to the decision-making of drug dose (such as dose response, treatment effect, side effects, etc.);

[0042] (6) Construct a prior distribution P(V) by combining expert prior knowledge and use Monte Carlo simulation MCMC sampling to optimize the posterior distribution P(V|M, P). The prior distribution reflects our beliefs or assumptions about each variable (such as absorption rate, volume of distribution, clearance rate, etc.) in the drug metabolism process before actual data. This belief or assumption can be based on historical data, literature, knowledge of domain experts, etc. MCMC constructs a Markov chain, and the stationary distribution of this chain is the posterior distribution we need. MCMC generates new samples repeatedly from the current state (i.e., a certain value of drug metabolism variables) and finally converges to the posterior distribution. The posterior distribution represents our updated beliefs about drug metabolism variables after observing new data M (such as metabolic feature matrix) and P (such as individual patient data). The formula is as follows:

[0043]

[0044] where P(M, P|V) represents the probability of observing data M and P given drug metabolism variable V, P(V) represents our belief about drug metabolism variables without data, and P(M, P) is used to ensure that the posterior distribution is a valid probability distribution;

[0045] (7) In pharmacology, it is usually assumed that as the dose increases, the pharmacological response will increase, but after exceeding a certain threshold, the pharmacological effect may saturate or the side effects may increase. Therefore, we need to establish a dose-response function:

[0046] E = f(D, A, V d , CL)

[0047] where E represents the pharmacodynamic response, D is the administered dose, A, V d , CL are individual causal variables, the absorption rate (A), volume of distribution (V d ), clearance rate (CL) of the drug. These parameters determine how the drug is distributed, metabolized, and cleared in the body, thus directly affecting the pharmacodynamic effect;

[0048] (8) Set the objective function:

[0049] L(D) = α(E max - E(D)) 2 + βR(D)

[0050] where L(D) is the comprehensive function used to evaluate the given dose D, E(D) is the therapeutic effect produced at the given dose, R(D) is the side effect produced at the given dose. The particle swarm optimization (PSO) algorithm is used to iteratively search for the optimal dose D* to maximize the therapeutic effect E(D) and minimize the side effect R(D). In dose optimization, each particle corresponds to a dose value, and each particle has two key attributes: position (corresponding to the drug dose) and velocity (indicating how the particle adjusts its position). In each iteration, the particle updates its velocity and position. The particle updates its velocity and position based on its own historical optimal position and the optimal position of all particles, thus continuously exploring the possible best solutions. Particle swarm optimization can handle complex multi-dimensional search spaces and can effectively avoid local optimal solutions, thereby improving the reliability of dose optimization;

[0051] (9) Combine the optimal dose D* with the metabolic prediction curve and output an individualized dose recommendation report for clinical reference, including an individualized medication plan with dose recommendation values, metabolic abnormality warnings, and suggestions for adjusting the dosing interval.

[0052] The above is only the preferred embodiment of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and deformations can be made, and these improvements and deformations should also be regarded as the protection scope of the present invention.

Claims

1. An individualized drug dosage optimization method based on causal traceability, characterized in that It includes the following steps: (1) Obtain multimodal data of the patient, including PET-CT scan image I PET , I CT , blood biochemical index B, physiological parameter P. PET is used to display the distribution of drugs, and CT is used to display anatomical structures; (2) Use the Transformer model to align I PET (t) and I CT (t) in the time dimension to construct a unified spatio-temporal sequence S(t), and the formula is as follows: S(t) = f Transformer (I PET (t), I CT (t)) During the spatio-temporal alignment process, the Transformer compares the PET and CT images at each moment, learns the potential connections between them, and generates unified temporal information, where f Transformer represents the Transformer model, and S(t) is a unified spatio-temporal sequence that integrates PET and CT data and can simultaneously contain drug metabolism information and anatomical structure information; (3) Decouple the multi-modal data S(t) after spatio-temporal alignment by using principal component analysis (PCA) and non-negative matrix factorization (NMF) to obtain the main metabolic feature matrix M and the noise feature matrix N. The formula is as follows: M = WH S(t) = M + N where W is the feature basis, H is the coefficient matrix, M represents the main features, and N represents noise or other minor components; (4) Based on M and P, construct a Bayesian network G = (V, E). The node set V represents drug metabolism variables, and the edge set E represents the causal paths between variables. M contains the metabolic features extracted from the multi-modal data, and P contains the individual data of the patient; (5) Identify key causal variables related to dose decision using the Markov blanket algorithm For each variable in the drug metabolism model, calculate its Markov blanket. By analyzing each variable and its Markov blanket, determine which variables are closely related to the decision of drug dose; (6) Construct a prior distribution P(V) by combining expert prior knowledge, and use Monte Carlo Markov chain (MCMC) sampling to optimize the posterior distribution P(V|M, P). The formula is as follows: where P(M, P|V) represents the probability of observing data M and P given the drug metabolism variable V, P(V) represents the belief in the drug metabolism variable without data, and P(M, P) is used to ensure that the posterior distribution is a valid probability distribution; (7) Establish a dose-response function: E = f(D, A, V d , CL) where E represents the pharmacodynamic response, D is the administered dose, A, V d , CL is an individual causal variable, A represents the absorption rate of the drug, V d represents the volume of distribution, and CL represents the clearance rate; (8) Set the objective function: L(D) = α(E max - E(D)) 2 + βR(D) where L(D) is a comprehensive function used to evaluate the given dose D, E(D) is the treatment effect produced at the given dose, and R(D) is the side effect produced at the given dose. Use the particle swarm optimization (PSO) algorithm to iteratively search for the optimal dose D*. In dose optimization, each particle corresponds to a dose value, and each particle has two key attributes: position and velocity. In each iteration, the particle updates its velocity and position. The particle updates its velocity and position based on its own historical optimal position and the optimal position of all particles, so as to continuously explore the possible best solution; (9) Combine the optimal dose D* with the metabolic prediction curve and output an individualized dose recommendation report for clinical reference, including an individualized medication plan containing the dose recommendation value, metabolic abnormality warning, and dosing interval adjustment suggestion.

Images