文献速递系统性红斑狼疮

预测儿童狼疮肾炎中霉酚酸暴露:三种临床场景的真实世界研究

Predicting mycophenolic acid exposure in paediatric lupus nephritis: a real-world study across three clinical scenarios.
PubMed 文献 · Rheumatology (Oxford) · 2026-01-01

本研究基于154例儿童狼疮肾炎患者(1376次随访)数据,开发机器学习模型预测霉酚酸(MPA)药时曲线下面积(AUC),覆盖治疗窗达标判断、AUC精确估计及剂量调整后暴露预测三种场景。LightGBM预测治疗窗达标的AUC为0.67,随机森林模型在AUC估计中R2=0.84,剂量调整后预测R2=0.51。SHAP分析提示剂量、体重、肾功能和采血时间点为关键预测因子,并开发了网络工具辅助临床决策。

要点

  • 机器学习模型可预测儿童狼疮肾炎患者MPA治疗窗达标(LightGBM的AUC=0.67)。
  • 随机森林结合临床指标与C0.5、C1.5、C4浓度可较准确估计MPA-AUC(R2=0.84)。
  • 剂量调整后暴露预测性能有限(随机森林R2=0.51),需谨慎解读。

核心结论

  • 建立了首个用于儿童狼疮肾炎患者MPA个体化给药的机器学习框架,可弥补治疗药物监测(TDM)的局限性并提升精准医疗水平。

研究设计与样本

  • 真实世界研究,纳入接受MPA治疗的儿童狼疮肾炎患者,分为常规治疗组(组1)和剂量调整组(组2),共154例患者、1376次随访。
  • 使用XGBoost、LightGBM和随机森林等算法构建模型,并以Shapley加性解释(SHAP)方法解释模型。

主要发现

  • 场景1中XGBoost和GBM模型表现优异,LightGBM在预测治疗窗达标方面最优(AUC=0.67,精确率=0.64,召回率=0.66,F1=0.65)。
  • 场景2中结合临床指标与C0.5、C1.5、C4浓度的随机森林模型预测最准确(R2=0.84,平均绝对误差=3.97,均方根误差=5.97)。
  • 场景3纳入99例患者的246次随访记录,随机森林模型表现最佳(R2=0.51,平均绝对误差=20.57)。
  • SHAP分析显示剂量、体重、肾功能和采血时间点是关键预测因子。

关键数值与推荐

  • 场景1预测治疗窗达标:LightGBM的AUC=0.67、精确率=0.64、召回率=0.66、F1=0.65。
  • 场景2估计MPA-AUC:随机森林(含临床指标及C0.5、C1.5、C4浓度)的R2=0.84、平均绝对误差=3.97、均方根误差=5.97。
  • 场景3预测剂量调整后暴露:随机森林的R2=0.51、平均绝对误差=20.57。

临床怎么用

  • 可基于机器学习模型辅助判断儿童狼疮肾炎患者MPA治疗窗是否达标,减少传统TDM的局限。
  • 利用C0.5、C1.5、C4浓度联合临床指标可较准确估计MPA-AUC,支持个体化剂量调整。
  • 研究提供的网络工具可辅助临床决策,但剂量调整后预测准确性有限,需结合临床判断。

局限性

  • 原文未报告局限性相关内容。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

OBJECTIVES: Childhood-onset systemic lupus erythematosus (cSLE) with lupus nephritis (LN) presents therapeutic challenges due to complex mycophenolic acid (MPA) pharmacokinetics. This study developed machine learning (ML)-based models to predict MPA area under the concentration-time curve (MPA-AUC) across three scenarios: determining therapeutic window attainment, precise AUC estimation and forecasting post-dose-adjustment exposure. METHODS: Data were collected from cLN on MPA with two cohorts: routine treatment (group 1) and dose adjustment (group 2). Comprehensive patient data, including demographics, medication details and laboratory test results, were collected. Models were constructed using algorithms such as XGBoost, LightGBM and Random Forest (RF). Model performance was assessed using multiple metrics and the Shapley Additive Explanations (SHAP) method was employed to interpret the models. RESULTS: Data from 154 patients (1376 follow-ups) were included. In Scenario 1, XGBoost and GBM models demonstrated superior performance, with the LightGBM model excelling in predicting therapeutic window attainment (AUC = 0.67, Precision = 0.64, Recall = 0.66, F1 = 0.65). In Scenario 2, the RF model incorporating clinical indicators and C0.5, C1.5, C4 concentrations achieved the most accurate predictions (R2 = 0.84, mean absolute error (MAE) = 3.97 and root mean square error (RMSE) = 5.97). In Scenario 3, which included data from 246 follow-up records of 99 patients, the RF model showed the best performance (R2 = 0.51 and MAE = 20.57). SHAP analysis highlighted dose, weight, renal function and concentration timepoints as key predictors. A web-based tool was developed to support clinical decisions. CONCLUSION: This study establishes the first ML-driven framework for personalized MPA dosing in paediatric LN, addressing TDM limitations and enhancing precision medicine.

来源:Rheumatology (Oxford),查看原文。版权归原作者所有。

引用

  1. Liu B, Chen L, Wang Y, et al. Predicting mycophenolic acid exposure in paediatric lupus nephritis: a real-world study across three clinical scenarios. Rheumatology (Oxford). 2026. PMID: 41701178. doi:10.1093/rheumatology/keag094

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。