| Background:Nephrotic syndrome is one of the most common glomerular diseases in childhood and an important risk factor for end-stage renal disease in patients.Clinically,glucocorticoids(steroids)are the first choice for the treatment of primary nephrotic syndrome,but steroids dependence or resistance still occurs in a small number of patients,and long-term ineffective/ineffective use of steroids may delay the disease or increase the incidence of adverse effects.For these patients who fail to respond to steroids therapy,additional immunosuppressive agents such as tacrolimus(TAC)are required.However,TAC suffers from large individual differences in pharmacokinetics and difficulty in dosing.Therefore,there is an urgent need to predict the response of children to steroids therapy before administration and to change the treatment regimen in a timely manner;and to explore the factors affecting the large individual differences in TAC pharmacokinetics and to establish a pharmacokinetic prediction model for TAC to guide individualized drug administration.In recent years,machine learning,one of the branches of artificial intelligence,has been widely used in the medical field,including diagnosis,treatment,and prevention of diseases.Unlike traditional logistic regression and specialized population pharmacokinetic(PPK)methods,machine learning has the advantage of being able to handle large,complex,multidimensional data and build good prediction models.In addition,machine learning has the advantage of finding algorithmic models with the best prediction performance and the interpretability in PPK theory can exactly complement each other.Aims:(1)Using clinical data from children diagnosed with nephrotic syndrome in electronic medical record systems over the past decade,a predictive model for glucocorticoid efficacy in children with nephrotic syndrome was established based on machine learning methods.(2)Incorporate clinical features and genetic variables that may affect the pharmacokinetic parameters of TAC,establish a TAC PPK model for children with nephrotic syndrome,and further use machine learning methods to establish a predictive model for TAC pharmacokinetic parameters.Methods:(1)994 patients diagnosed with childhood nephrotic syndrome and treated with glucocorticoids alone between January 2009 and May 2020 were retrospectively included using the electronic medical record system,and an additional 50 patients were collected for external validation of the model.Clinical data including demographic data,blood tests,urine tests,and other clinical data before and after glucocorticoids therapy of the children in the electronic medical record system were collected.A predictive model of glucocorticoid efficacy was built and validated using machine learning,and finally the model was interpreted using the SHAP method.(2)141 children with nephrotic syndrome from August 2013 to December 2018were included,and blood samples of TAC trough and partial near peak concentrations were collected.The blood concentrations of TAC were determined by enzyme immunoassay;the genes CYP3A5,MYH9 and LAMB2 that may affect the pharmacokinetics of TAC were determined;and a PPK model was established using a nonlinear mixed-effects approach.Based on this,machine learning method was used to establish the predicton model of TAC clearance rate.Results:(1)Based on clinical data at predose,one month,and two months of dosing,with efficacy at 1-and 2-months post-dose as the outcome.We developed a prediction model for steroids efficacy in pediatric nephrotic syndrome patients under five scenes.The external validation results showed that the AUC values of the models for the five scenes ranged from 0.634 to 0.828,with the best predictive power of using variables before and after one month of administration to predict 2-month efficacy with steroids,using the random forest algorithm,and the maximum AUC value is0.828.In addition,machine learning modeling results showed that first dose of glucocorticoid,preadministration urinaryβ2 microglobulin/creatinine,creatinine,and cystatin C were important variables for predicting steroids efficacy.(2)A one-compartment model of first-order absorption and elimination adequately described the pharmacokinetics of TAC.Age,co-administration of Wuzhi capsules,CYP3A5*3/*3 genotype and CTLA4 rs4553808 genotype were significantly affecting the clearance of TAC.Among the six machine learning models,the Lasso algorithm model performed the best(R~2=0.42).Conclusions:(1)Based on approximately 10 years of electronic medical record data before and after steroids administration.For the first time,we used machine learning to build prediction models of steroids efficacy in pediatric nephrotic syndrome patients under five scenes.These five different scenes simulate different actual treatment situations,and the clinic can use these models to predict whether patients are steroids resistant in a timely and accurate manner before and during treatment,and then intervene in advance to improve the efficiency of drug treatment and avoid disease progression.(2)A population pharmacokinetic combined with machine learning approach was used to develop a clearance prediction model for tacrolimus in pediatric patients with nephrotic syndrome.The model incorporated genetic variables affecting the clearance of tacrolimus,and the detection of patient-related genotypes should be considered clinically to provide a reference for optimizing the initial dose administered and to achieve the goal of individualized treatment. |
PDF Full Text Request