A tacrolimus dosage prediction method based on clustering model combination

Abstract

The application relates to the technical field of computers and discloses a tacrolimus dosage prediction method based on a clustering model combination, which comprises a data preprocessing module, a clustering and grouping module, a concentration prediction module, a dosage recommendation module and a display module. A patient grouping modeling end is arranged, seven clinical indexes are integrated by using a clustering algorithm, subgroups with large metabolic characteristic differences are divided, the limitation that patients are regarded as homogeneous groups in a traditional model is broken through, subgroup-specific modeling is realized, the group specificity of dosage prediction is improved, a subgrouping and prediction joint architecture is constructed, a model is independently deployed for each subgroup after subgrouping and dimension reduction, the sample quantity required by a single model is effectively reduced, the failure risk of machine learning under a small sample is reduced, reliable modeling can be realized in hundreds of clinical data, a dynamic prediction modeling end is arranged, memory functions are generated by fusing multi-time point historical data, and the prediction deviation rate caused by initial value disturbance is reduced.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a method for predicting tacrolimus dosage based on a clustering model. Background Technology

[0002] Tacrolimus is a fermentation product isolated from Streptomyces. It is a macrolide antibiotic and a potent immunosuppressant. It is a first-line drug for preventing rejection after solid organ transplantation. As the core immunosuppressant for anti-rejection treatment after organ transplantation, the clinical efficacy of tacrolimus is highly dependent on the precise control of blood drug concentration.

[0003] Currently, tacrolimus dosage prediction based on clustering models cannot effectively distinguish the heterogeneity of key biomedical characteristics in patients. The algorithm relies solely on static physiological indicators for clustering, ignoring genetic polymorphisms and dynamic fluctuations in blood drug concentrations. This can lead to dosage deviations within the same cluster exceeding the therapeutic window threshold, potentially causing rejection or nephrotoxicity. Furthermore, when making dosage decisions, a concentration-efficacy dynamic response mechanism cannot be established. When blood drug concentrations are within the therapeutic window but trending upwards, there is a lack of predictive adjustment strategies based on metabolic phenotypes, and clinical outcome indicators are not correlated, resulting in… More than 26% of patients who met the predicted dose target still experienced subclinical rejection; at the data fusion level, it was unable to coordinate cross-modal medical feature clustering, missed the impact of abnormal bilirubin after liver transplantation on drug protein binding rate, and failed to integrate the time sequence characteristics of medication events, resulting in a dose error rate as high as 37% in multi-drug combination scenarios; in addition, in the cluster boundary area, it was unable to generate a gradual dose transition plan, hard clustering caused dose jumps, and the lack of fuzzy cluster confidence assessment resulted in more than 3 dose direction adjustments within 72 hours for borderline patients, significantly increasing the risk of neurotoxicity or concentration runaway.

[0004] Therefore, a method for predicting tacrolimus dosage based on a clustering model is proposed to solve the above problems. Summary of the Invention

[0005] (a) Technical problems to be solved.

[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting tacrolimus dosage based on a clustering model, which solves the problems mentioned in the background.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting tacrolimus dosage based on a clustering model, comprising the following steps: S1. Perform data preprocessing: Screen key clinical indicators that affect tacrolimus blood concentration, including aspartate aminotransferase, alanine aminotransferase, body mass index, creatinine clearance, drug dosage, pre-test duration of treatment and prothrombin time score. S2. Patient grouping modeling: Based on the selected clinical indicators, density clustering algorithm is used to group the liver transplant patient population and divide them into subgroups with large feature differences. S3. Predicting blood drug concentrations: For each subgroup of patients, a self-memory nonlinear grey Bernoulli model is used to dynamically predict the temporal changes in blood drug concentrations. S4. Optimize dosage output: Based on the deviation between the predicted blood drug concentration and the target therapeutic window, generate individualized drug dosage adjustment suggestions; S5. Obtain the clinical dataset of liver transplant patients, which includes laboratory test indicators, combined medication records and historical blood drug concentration time series data; S6. Perform key indicator screening; S7. After dimensionality reduction based on principal component analysis, density clustering algorithm is used to group patients with a neighborhood radius ε=0.5 and a minimum number of samples MinPts=5, and outputs 4 characteristic subgroups. S8. For each subgroup of patients, extract their blood drug concentration time sequence and construct a self-memory nonlinear grey Bernoulli model. S9. Compare predicted values ​​with the target treatment window.

[0009] Preferably, the method for screening key clinical indicators in step S1 includes: Analysis of variance was used to assess the correlation between categorical variables and changes in blood drug concentration. Linear regression analysis was used to assess the correlation of continuous variables; The screening criteria were statistical indicators with a p-value less than 0.05.

[0010] Preferably, the density clustering algorithm in step S2 evaluates the clustering quality using the Davies-Bouldin index (DBI), and the formula for calculating DBI is:

[0011] in This is a clustering quality evaluation index, where n is the total number of clusters. Let be the intra-cluster scatter of the i-th and j-th clusters. This indicates the inter-cluster separation degree.

[0012] Preferably, the subgroups divided in step S2 include four categories: Category 1: AST and ALT levels are higher than other subgroups, BMI is normal, and Ccr is low; Category 2: BMI is evenly distributed and normal, but there are significant individual differences in medication dosage; Category 3 and 4: High Ccr values, drug dosage concentrated at 2.5 mg, and blood drug concentration showing an increasing trend.

[0013] Preferably, the SA-NGBM model construction process in step S3 includes: S31. Perform a first-order accumulation operation on the original blood drug concentration sequence; S32. Establish the nonlinear grey Bernoulli equation:

[0014] Where v is the Bernoulli power exponent, a is the development coefficient, and b is the gray action quantity. This is the Bernoulli power index. For background values, The original blood drug concentration sequence was observed at time t. Seven clinical indicators were integrated using a clustering algorithm to divide the patients into subgroups with large differences in metabolic characteristics. This breaks through the limitation of traditional models that treat patients as a homogeneous group and achieves subgroup-specific modeling.

[0015] Preferably, the specific method for introducing the self-remembering mechanism into the SA-NGBM model is as follows: Define the backtracking order p and the time interval. Construct a self-remembering difference integral equation; The prediction equation is obtained by fusing historical multi-time point data using the memory function β(t) and then discretizing it.

[0016] Where p is the backtracking order. For memory coefficient, Interpolation for historical points in time, For historical time indexing.

[0017] Preferably, the optimization objective of the parameters γ, a, and b is to minimize the mean absolute percentage error, and the constraint condition is: γ≠1; Background value weighting coefficient φ∈[0,1]; By using MATLAB to iteratively optimize the global parameter combination, and by constructing a cluster-prediction joint architecture, the model is first clustered and dimensionality reduced, and then the model is deployed independently for each subgroup. This effectively reduces the sample size required for a single model, reduces the risk of machine learning failure under small sample sizes, and enables reliable modeling in hundreds of clinical cases.

[0018] Preferably, the process of generating the predicted blood drug concentration in step S3 includes: Perform an inverse cumulative subtraction operation on the first-order cumulative sequence output by the SA-NGBM model to obtain the original sequence prediction value; The prediction formula is:

[0019] in This represents the predicted value of the original sequence at time t+1. This represents the simulated value of the cumulative sequence at time t.

[0020] Preferably, the dosage optimization method in step S4 includes: When the predicted blood drug concentration is below the target therapeutic window, it is recommended to increase the dose by 1 mg / day; When the predicted value is higher than the target therapeutic window, it is recommended to reduce the dose by 1 mg / day; The adjustment plan needs to be dynamically calibrated in combination with the characteristics of patient subgroups.

[0021] Preferably, in step S2, the density clustering uses a combination of parameters with a neighborhood radius ε=0.5 and a minimum number of samples MinPts=5, and achieves visual clustering after dimensionality reduction based on principal component analysis; in step S3, the SA-NGBM model constructs a self-remembering dynamic equation by introducing a memory function of multi-time point historical data to overcome the sensitivity of traditional models to initial values. The specific method for performing key indicator screening in S6 is as follows: Analysis of variance was performed on categorical variables to screen for prothrombin time score, drug dosage, and pre-test course of medication. Linear regression analysis was performed on continuous variables to screen for aspartate aminotransferase, alanine aminotransferase, body mass index, and creatinine clearance rate. The above indicators are combined to form a 7-dimensional feature vector; S8 also optimizes the Bernoulli power exponent γ, development coefficient a, and gray action b by minimizing the mean absolute percentage error. A memory function with a backtracking order of p=1 is used to fuse data from multiple time points to generate predicted blood drug concentrations. The specific method for S9 is as follows: If the predicted value is below the therapeutic window, it is recommended to increase the output dose by 1 mg / day. If the predicted value is higher than the treatment window, it is recommended to reduce the output dose by 1 mg / day. By introducing a backtracking mechanism in SA-NGBM, a memory function is generated by fusing historical data from multiple time points, eliminating the dependence of traditional NGBM on a single initial value and reducing the prediction bias rate caused by initial value perturbation.

[0022] (III) Beneficial Effects Compared with existing technologies, this invention provides a method for predicting tacrolimus dosage based on a clustering model, which has the following advantages: 1. In this invention, by setting up a patient group modeling end, a clustering algorithm is used to integrate seven clinical indicators to divide the patients into subgroups with large differences in metabolic characteristics. This breaks through the limitation of traditional models that treat patients as a homogeneous group, and achieves subgroup-specific modeling, thereby improving the group-specificity of dose prediction.

[0023] 2. In this invention, by constructing a clustering-prediction joint architecture, the model is first clustered and dimensionality reduced, and then the model is deployed independently for each subgroup. This effectively reduces the sample size required for a single model, reduces the risk of machine learning failure under small sample sizes, and achieves reliable modeling in hundreds of clinical cases.

[0024] 3. In this invention, by setting up a dynamic prediction modeling end, a backtracking mechanism is introduced into SA-NGBM, and a memory function is generated by integrating historical data from multiple time points. This eliminates the dependence of traditional NGBM on initial values ​​at a single time point and reduces the prediction offset rate caused by initial value disturbances. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Please see Figure 1 A method for predicting tacrolimus dosage based on a clustering model includes the following steps: S1. Perform data preprocessing: Screen key clinical indicators that affect tacrolimus blood concentration, including aspartate aminotransferase, alanine aminotransferase, body mass index, creatinine clearance, drug dosage, pre-test duration of treatment and prothrombin time score. S2. Patient grouping modeling: Based on the selected clinical indicators, density clustering algorithm is used to group the liver transplant patient population and divide them into subgroups with large feature differences. S3. Predicting blood drug concentrations: For each subgroup of patients, a self-memory nonlinear grey Bernoulli model is used to dynamically predict the temporal changes in blood drug concentrations. S4. Optimize dosage output: Based on the deviation between the predicted blood drug concentration and the target therapeutic window, generate individualized drug dosage adjustment suggestions; S5. Obtain the clinical dataset of liver transplant patients, which includes laboratory test indicators, combined medication records and historical blood drug concentration time series data; S6. Perform key indicator screening; S7. After dimensionality reduction based on principal component analysis, density clustering algorithm is used to group patients with a neighborhood radius ε=0.5 and a minimum number of samples MinPts=5, and outputs 4 characteristic subgroups. S8. For each subgroup of patients, extract their blood drug concentration time sequence and construct a self-memory nonlinear grey Bernoulli model. S9. Compare predicted values ​​with the target treatment window.

[0028] The screening methods for key clinical indicators in step S1 include: Analysis of variance was used to assess the correlation between categorical variables and changes in blood drug concentration. Linear regression analysis was used to assess the correlation of continuous variables; The screening criteria were statistical indicators with a p-value less than 0.05.

[0029] The density-based clustering algorithm in step S2 evaluates clustering quality using the Davies-Bouldin index (DBI). The formula for calculating DBI is:

[0030] in This is a clustering quality evaluation index, where n is the total number of clusters. Let be the intra-cluster scatter of the i-th and j-th clusters. This indicates the inter-cluster separation degree.

[0031] The subgroups identified in step S2 include four categories: Category 1: AST and ALT levels are higher than other subgroups, BMI is normal, and Ccr is low; Category 2: BMI is evenly distributed and normal, but there are significant individual differences in medication dosage; Category 3 and 4: High Ccr values, drug dosage concentrated at 2.5 mg, and blood drug concentration showing an increasing trend.

[0032] Step S3, the SA-NGBM model construction process, includes: S31. Perform a first-order accumulation operation on the original blood drug concentration sequence; S32. Establish the nonlinear grey Bernoulli equation:

[0033] Where v is the Bernoulli power exponent, a is the development coefficient, and b is the gray action quantity. This is the Bernoulli power index. For background values, The values ​​are the original blood drug concentration sequences observed at time t.

[0034] The specific method for introducing the self-remembering mechanism into the SA-NGBM model is as follows: Define the backtracking order p and the time interval. Construct a self-remembering difference integral equation; The prediction equation is obtained by fusing historical multi-time point data using the memory function β(t) and then discretizing it.

[0035] Where p is the backtracking order. For memory coefficient, Interpolation for historical points in time, For historical time indexing.

[0036] The optimization objective for parameters γ, a, and b is to minimize the mean absolute percentage error, with the following constraints: γ≠1; Background value weighting coefficient φ∈[0,1]; The globally optimal parameter combination can be obtained through iterative optimization using MATLAB.

[0037] The process of generating the predicted blood drug concentration in step S3 includes: Perform an inverse cumulative subtraction operation on the first-order cumulative sequence output by the SA-NGBM model to obtain the original sequence prediction value; The prediction formula is:

[0038] in, This represents the predicted value of the original sequence at time t+1. This represents the simulated value of the cumulative sequence at time t.

[0039] The dosage optimization methods for S4 include: When the predicted blood drug concentration is below the target therapeutic window, it is recommended to increase the dose by 1 mg / day; When the predicted value is higher than the target therapeutic window, it is recommended to reduce the dose by 1 mg / day; The adjustment plan needs to be dynamically calibrated in combination with the characteristics of patient subgroups.

[0040] include: Data preprocessing module: configured to perform clinical indicator screening and data standardization; Clustering module: Integrates the DBSCAN algorithm and outputs patient subgroup classification results; Concentration prediction module: Deploy the SA-NGBM model to generate time-series predicted values ​​of blood drug concentrations; Dosage recommendation module: Outputs dosage adjustment plan based on the deviation between the prediction result and the treatment window; Display module: Visualizes clustering, prediction curves, and medication recommendations.

[0041] In step S2, density clustering uses a combination of parameters: neighborhood radius ε=0.5 and minimum sample size MinPts=5. Visual clustering is achieved after dimensionality reduction based on principal component analysis. In step S3, the SA-NGBM model constructs a self-remembering dynamic equation by introducing a memory function of multi-time point historical data to overcome the sensitivity of traditional models to initial values. The specific method for performing key indicator screening in S6 is as follows: Analysis of variance was performed on categorical variables to screen for prothrombin time score, drug dosage, and pre-test course of medication. Linear regression analysis was performed on continuous variables to screen for aspartate aminotransferase, alanine aminotransferase, body mass index, and creatinine clearance rate. The above indicators are combined to form a 7-dimensional feature vector; S8 also optimizes the Bernoulli power exponent γ, development coefficient a, and gray action b by minimizing the mean absolute percentage error. A memory function with a backtracking order of p=1 is used to fuse data from multiple time points to generate predicted blood drug concentrations. The specific method for S9 is as follows: If the predicted value is below the therapeutic window, it is recommended to increase the output dose by 1 mg / day. If the predicted value is higher than the therapeutic window, it is recommended to reduce the output dose by 1 mg / day.

[0042] Implementation 1: Application in liver transplant patients Step 1: Clinical Data Acquisition and Feature Extraction Multidimensional clinical indicators were collected from the patient population, including liver function parameters, physiological characteristics, and medication records. A dual-path screening mechanism, employing categorical variable ANOVA and continuous variable linear regression, was used to extract core indicators affecting changes in blood drug concentration. After screening, a feature vector was generated, covering key dimensions of metabolic capacity, weight status, and treatment progress.

[0043] Step 2: Patient Subgroup Classification After dimensionality reduction using principal component analysis, the preprocessed data was grouped into patient clusters using a density clustering algorithm. A specific neighborhood radius and minimum sample threshold were set to identify natural cluster structures within the feature space; ultimately, mutually exclusive feature subgroups were output, with each group exhibiting differences in liver function levels and drug metabolism rates.

[0044] Step 3: Time-series modeling of blood drug concentration Independent prediction models were constructed for each patient subgroup: 1. Data initialization: Input the patient's historical blood drug concentration time series; 2. Parameter optimization: The power exponent, development coefficient, and gray action of the Bernoulli equation are determined through iterative optimization. 3. Self-memory fusion: Loading a backtracking mechanism to dynamically integrate historical data from multiple time points using memory functions; 4. Generation Prediction: Solve the self-memory difference integral equation and output the predicted concentration values ​​at future time points.

[0045] Step 4: Individualized Dosage Adjustment Compare the predicted concentration values ​​with the target therapeutic window: 1. When the predicted value remains below the lower limit of the therapeutic window, it is recommended to increase the daily dose in a stepwise manner, taking into account the metabolic characteristics of the patient's subgroup. 2. When the predicted value approaches or exceeds the upper limit of the therapeutic window, a phased dose reduction regimen is recommended based on the renal function status of the subgroup; 3. Maintenance dosage decisions should take into account coagulation function score subgroup-specific indicators.

[0046] Step 5: Clinical efficacy verification 1. Comparison with traditional prediction models shows that: 2. Cluster modeling improves the consistency of concentration trends, especially for patients with metabolic disorders; 3. The self-remembering mechanism effectively suppresses prediction bias caused by initial value fluctuations, thus enhancing the stability of the time series; 4. The dosage adjustment plan is precisely matched with the subgroup characteristics, resulting in a decrease in the incidence of postoperative rejection and adverse drug reactions.

[0047] Implementation 2: Application in kidney transplant patients The clinical procedure for implementing this method in the long-term immunosuppression management of kidney transplant recipients is divided into three consecutive stages. First, in-depth analysis of the patient's pharmacokinetic characteristics is performed, integrating postoperative renal function dynamic monitoring indicators, combined drug interaction matrix and gene polymorphism detection results. The core variable set is extracted through regularized feature selection technology, focusing on capturing the correlation characteristics between the trend of cyclosporine concentration-time curve change and the glomerular filtration rate decay rate.

[0048] Subsequently, a metabolic phenotype intelligent grouping engine was activated, employing a density clustering algorithm to perform nonlinear dimensionality reduction mapping while preserving the clinical feature topology. This adaptively identified three patient subgroups: high-metabolism, standard clearance, and delayed excretion. The delayed excretion subgroup was identified by abnormal bile acid-binding protein levels. A cross-modal prediction-monitoring dual-channel architecture was established for this high-risk group. The main prediction channel loaded a self-memory grey model with a forgetting factor, dynamically fusing the most recent five blood drug concentration monitoring values ​​through a memory function to generate a concentration prediction trajectory. The auxiliary safety channel analyzed the fluctuation patterns of renal tubular injury markers in real time, automatically triggering a constraint reinforcement mechanism when N-acetyl-β-D-glucosidase levels were abnormally elevated.

[0049] The dosing decision system generates a dosing regimen based on the deviation between the predicted concentration and the personalized treatment window. For patients with a continuously declining estimated glomerular filtration rate, a protective adjustment logic is initiated to prioritize maintaining the trough concentration within the lower limit of the target window. At the same time, a drug-food interaction compensation factor is embedded, and the final output is a complete medical order recommendation including the starting dose, dose titration gradient, and next monitoring time window.

[0050] Clinical trials have shown that this technology system breaks through the blind spots of traditional kidney transplant patient medication: it avoids subclinical rejection caused by immunodeficiency, effectively inhibits the progression of nephrotoxicity of calcineurin inhibitors, and improves the efficiency of physician decision-making.

[0051] Implementation 3: Child Heart Transplant Patient Scenario This method is applied in the management of pediatric heart transplant recipients to form a special population adaptation technology pathway. The clinical operation begins with the dynamic acquisition of multi-dimensional physiological parameters, integrating the child's growth and development curves, cardiac output changes, and cytochrome P450 phenotype test reports. A virtual twin is constructed using adversarial generative network technology to simulate the developmental correlation of pharmacokinetic parameters.

[0052] The core feature selection focused on the nonlinear relationship between weight gain and apparent volume of distribution, and used a temporal convolutional network to extract the coupling features between fluctuations in liver enzyme activity and troughs in blood drug concentration. Subsequently, a growth-adaptive clustering engine was used to embed age-correcting factors in Euclidean space to dynamically divide the population into two subgroups: a fast-metabolic developmental group and a slow-metabolic stable group. The fast-metabolic group was characterized by a surge in drug clearance rate during puberty.

[0053] A developmental-dependent prediction model was constructed for this subgroup, using bone age index as an adaptive adjustment variable for memory retracement order. During growth spurts, the retracement window was automatically shortened to capture recent dramatic changes in metabolic rate. Simultaneously, a physiological development noise filtering module was introduced to suppress interference from abnormal concentration monitoring values ​​caused by hormonal fluctuations during puberty. The dosage optimization system dynamically expands the safety margin of the treatment window based on serum troponin T levels and uses fuzzy decision trees to generate multi-objective dosing regimens: when the quarterly increase in body surface area exceeds a threshold, a cross-step dosage linkage mechanism is activated; and calcineurin inhibitors with minimal toxicity release constraints are implanted in high-risk patients with coronary artery disease after heart transplantation.

[0054] The final output is an intelligent drug administration guidance system that includes a parental interface, time synchronization reminders, and emergency poisoning plans. Clinical application has validated that this approach successfully overcomes a dual challenge in the field of pediatric transplantation: it addresses the issue of pharmacokinetic parameter drift caused by growth and development while avoiding the systemic drug risks associated with traditional adult regimens in children.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting tacrolimus dosage based on a clustering model, characterized in that, Includes the following steps: S1. Data Preprocessing: Screening key clinical indicators that affect tacrolimus blood concentration, including aspartate aminotransferase, alanine aminotransferase, body mass index, creatinine clearance, drug dosage, pre-test duration of treatment and prothrombin time score. S2. Patient grouping modeling: Based on the selected clinical indicators, density clustering algorithm is used to group the liver transplant patient population and divide them into subgroups with large feature differences. The subgroups defined in S2 include four categories: Category 1: AST and ALT levels are higher than other subgroups, BMI is normal, and Ccr is low; Category 2: BMI is evenly distributed and normal, but there are significant individual differences in medication dosage; Category 3 and Category 4: Ccr values ​​are relatively high, the dosage is concentrated at 2.5 mg, and the blood drug concentration shows an increasing trend; S3. Blood drug concentration prediction: For each subgroup of patients, a self-memory nonlinear grey Bernoulli model is used to dynamically predict the temporal changes in blood drug concentration. S4. Dosage Optimization Output: Based on the deviation between the predicted blood drug concentration and the target therapeutic window, generate individualized drug dosage adjustment suggestions; In step S2, density clustering uses a combination of parameters: neighborhood radius ε=0.5 and minimum number of samples MinPts=5. Visual clustering is achieved after dimensionality reduction based on principal component analysis. In step S3, the SA-NGBM model constructs a self-remembering dynamic equation by introducing a memory function of multi-time point historical data to overcome the sensitivity of traditional models to initial values.

2. The method for predicting tacrolimus dosage based on a clustering model as described in claim 1, characterized in that, The screening method for key clinical indicators in step S1 includes: Analysis of variance was used to assess the correlation between categorical variables and changes in blood drug concentration. Linear regression analysis was used to assess the correlation of continuous variables; The screening criteria were statistical indicators with a p-value less than 0.

05.

3. The method for predicting tacrolimus dosage based on a clustering model as described in claim 1, characterized in that, The density clustering algorithm in step S2 evaluates clustering quality using the Davies-Bouldin index (DBI). The formula for calculating DBI is: ； in This is a clustering quality evaluation index, where n is the total number of clusters. Let be the intra-cluster scatter of the i-th cluster, as shown in the table. To indicate the separation degree between clusters.

4. The method for predicting tacrolimus dosage based on a clustering model as described in claim 1, characterized in that, The SA-NGBM model construction process in step S3 includes: S31. Perform a first-order accumulation operation on the original blood drug concentration sequence; S32. Establish the nonlinear grey Bernoulli equation: ； Where v is the Bernoulli power exponent, a is the development coefficient, and b is the gray action quantity. This is the Bernoulli power index. For background values, The values ​​are the original blood drug concentration sequences observed at time t.

5. The method for predicting tacrolimus dosage based on a clustering model according to claim 4, characterized in that, The specific method for introducing the self-remembering mechanism into the SA-NGBM model is as follows: Define the backtracking order p and the time interval. Construct a self-remembering difference integral equation; The prediction equation is obtained by fusing historical multi-time point data using the memory function β(t) and then discretizing it. ； Where p is the backtracking order. For memory coefficient, Interpolation for historical points in time, For historical time indexing.

6. The method for predicting tacrolimus dosage based on a clustering model as described in claim 4, characterized in that, The optimization objective of the parameters γ, a, and b is to minimize the mean absolute percentage error, with the constraint that γ ≠ 1. Background value weighting coefficient φ∈[0,1]; The globally optimal parameter combination can be obtained through iterative optimization using MATLAB.

7. The method for predicting tacrolimus dosage based on a clustering model as described in claim 1, characterized in that, The process of generating the predicted blood drug concentration in step S3 includes: Perform an inverse cumulative subtraction operation on the first-order cumulative sequence output by the SA-NGBM model to obtain the original sequence prediction value; The prediction formula is: ； in This represents the predicted value of the original sequence at time t+1. This represents the simulated value of the cumulative sequence at time t.

8. The tacrolimus dosage prediction method based on clustering model combination according to claim 1, wherein the dosage optimization method in step S4 includes: When the predicted blood drug concentration is below the target therapeutic window, it is recommended to increase the dose by 1 mg / day; When the predicted value is higher than the target therapeutic window, it is recommended to reduce the dose by 1 mg / day; The adjustment plan needs to be dynamically calibrated in combination with the characteristics of patient subgroups.

Images