Individualized medication model of vancomycin in patients with renal
insufficiency using machine learning techniques
Abstract
Objectives: Inappropriate dose of vancomycin can cause nephrotoxicity,
which should be avoided in clinical, particularly in patients with renal
insufficiency. We aim to use machine learning techniques to explore
important variables influencing vancomycin dose in patients with renal
insufficiency and establish an individualized medication model to
benefit these patients. Methods: Vancomycin administration cases in
patients with renal insufficiency were collected at Xinhua Hospital
affiliated to Shanghai Jiaotong University School of Medicine from May
2018 to March 2022. Sequential Forward Selection was used for feature
selection. Eight machine learning algorithms were compared the
predictive performance, including XGBoost, LightGBM, CatBoost, GBDT, RF,
SVC, KNN, and Logistic Regression. The one with the best predictive
performance was chosen to calculate the importance score of modeling
variables and establish the individualized medication model. Dose
subgroups were divided into 500 mg, 1000 mg, 1500 mg and 2000 mg.
Subgroup analysis based on the modeling variables were conducted.
Results: This study included 237 eligible patients with 351 vancomycin
cases. Six important variables were screened out, including gender,
weight, AUC, uric acid, creatinine and total protein. CatBoost was
chosen with the best prediction performance (accuracy=0.59) for
modeling. The individualized medication model had precision over 53%
and recall rate over 50% among all dose subgroups. The prediction of
1500 mg vancomycin had the best precision (65%), recall rate (71%) and
F1-score (0.68). Conclusion: The individualized medication model of
vancomycin for patients with renal insufficiency has good predictive
performance, which can help clinicians make better decision of
vancomycin use.