Hu Ting, Ding Xian, Han Feifei, An Zhuoling
Department of Pharmacy, Beijing Chao-Yang Hospital, Capital Medical University, Beijing, China.
Front Pharmacol. 2025 Jun 12;16:1549500. doi: 10.3389/fphar.2025.1549500. eCollection 2025.
Appropriate vancomycin trough levels are crucial for ensuring therapeutic efficacy while minimizing toxicity. The aim of this study is to identify clinical factors that influence the steady-state trough concentration of vancomycin and to establish a machine learning model for accurately predicting vancomycin's steady-state trough concentration.
This study is a single-center, retrospective, observational investigation involving 546 hospitalized patients who received intravenous vancomycin therapy. A total of 57 clinical indicators were collected from the subjects. Random forest models were constructed and validated using internal and external datasets, with performance compared to a Bayesian PopPK model.
The random forest model incorporated a comprehensive set of clinical indicators, including creatinine clearance, C-reactive protein (CRP), B-type natriuretic peptide (BNP), high-density lipoprotein cholesterol (HDL-C), and daily vancomycin dose, collected 48 hours before steady-state concentration assessment. The random forest regression model achieved correlation coefficients of 0.94 for the training set and 0.81 for the test set, respectively. The random forest classification model demonstrated impressive accuracy rates of 0.99 for the training set and 0.84 for the test set. External validation further confirmed the model's generalization capabilities, with a predictive accuracy of 0.83, surpassing the Bayesian PopPK model's 0.57 accuracy.
This study presents a robust random forest model that predicts vancomycin steady-state trough concentrations with high accuracy, offering a significant advantage over existing Bayesian PopPK model. By integrating diverse clinical indicators, the model supports personalized medicine approaches and has the potential to improve clinical outcomes by facilitating more precise dosing strategies.
合适的万古霉素谷浓度对于确保治疗效果同时将毒性降至最低至关重要。本研究的目的是确定影响万古霉素稳态谷浓度的临床因素，并建立一个用于准确预测万古霉素稳态谷浓度的机器学习模型。
本研究是一项单中心、回顾性、观察性调查，涉及546例接受静脉万古霉素治疗的住院患者。从受试者中收集了总共57项临床指标。使用内部和外部数据集构建并验证随机森林模型，并将其性能与贝叶斯群体药代动力学（PopPK）模型进行比较。
随机森林模型纳入了一组全面的临床指标，包括肌酐清除率、C反应蛋白（CRP）、B型利钠肽（BNP）、高密度脂蛋白胆固醇（HDL-C）以及在稳态浓度评估前48小时收集的每日万古霉素剂量。随机森林回归模型在训练集和测试集上的相关系数分别为0.94和0.81。随机森林分类模型在训练集和测试集上的准确率分别达到了令人印象深刻的0.99和0.84。外部验证进一步证实了该模型的泛化能力，预测准确率为0.83，超过了贝叶斯PopPK模型的0.57准确率。
本研究提出了一个强大的随机森林模型，能够高精度预测万古霉素稳态谷浓度，比现有的贝叶斯PopPK模型具有显著优势。通过整合多种临床指标，该模型支持个性化医疗方法，并有可能通过促进更精确的给药策略来改善临床结果。
