Jeein Noh (1), Minji Kwon (2), Jongdae Han (3), Donghwan Lee (4), Bo-Hyung Kim (2,5,6)
(1) Department of Regulatory Science, Graduate School, Kyung Hee University, Republic of Korea, (2) Department of Clinical Pharmacology and Therapeutics, Kyung Hee University Hospital, Republic of Korea, (3) Department of Statistics and Data Science, Korea National Open University, Republic of Korea, (4) Department of Statistics, Ewha Womans University, Republic of Korea, (5) East-West Medical Research Institute, Kyung Hee University, Republic of Korea, (6) Department of Biomedical Science and Technology, Graduate School, Kyung Hee University, Republic of Korea
Objectives: Drug-induced nephrotoxicity accounts for 19-26% of cases of acute kidney injury in hospitalized patients[1]. Nephrotoxicity is closely related with patient morbidity and mortality. Antibiotics are one of the most common drugs of drug-induced nephrotoxicity. Vancomycin is commonly used to treat drug-resistant Gram-positive bacterial infections, such as methicillin-resistant Staphylococcus aureus (MRSA). Colistin is prescribed for the treatment of multidrug-resistant Gram-negative bacterial infections. These antibiotics are known to commonly cause nephrotoxicity[2]. Predicting nephrotoxicity is an important challenge in patient care, but there are no standards for identifying or early detecting drug-induced nephrotoxicity. Therefore, we aimed to develop a machine learning (ML)-based prediction model for kidney injury identification utilizing patient-specific factors, including medication information.
Methods: In the 8-year period of data, there were a total of 6,528 hospitalized patients over the age of 18 who were prescribed antibiotics such as vancomycin, colistin, teicoplanin, meropenem, and ertapenem. Information on subjects who developed nephrotoxicity prior to antibiotic administration was excluded. Additionally, information on cases in which oral vancomycin was administered was excluded. Antibiotic-induced nephrotoxicity was defined as an increase in serum creatinine 0.5 mg/dL or 1.5-fold higher than the baseline. Various ML models encompassing decision tree, random forest, logistic regression and XGBoost have been used to predict nephrotoxicity from patient information for each antibiotic[3,4,5,6]. Variables for model development were patient-specific variables (age, gender, height, weight, smoking and drinking information, vital signs, laboratory variables) and history of other concomitant medications, including the vancomycin, colistin, teicoplanin, and meropenem and other nephrotoxic drugs (adrenaline, ceftazidime, dopamine, linezolid, levofloxacin, metronidazole, midazolam, moxifloxacin, norepinephrine, remifentanil, tazobactam, vasopressin, etc.). To develop the ML model, estimated glomerular filtration rate (eGRF) value was calculated using various equations such as Cockcroft-Gault (CG), Modification of Diet in Renal Disease (MDRD), Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formulas, and used as the baseline variables.
Results: Approximately 30% of all patients met the criteria for nephrotoxicity. Also, nephrotoxicity occurred in approximately 20% and 50% of patients receiving vancomycin and colistin, respectively. The ML model for each antibiotic was developed to predict the presence or absence of nephrotoxicity. In most models, the amount and administration period of the antibiotics were selected as main variables. Concomitant medications such as tazobactam and remifentanil, concomitant diseases such as diabetes mellitus were also selected. Various eGFR values for MDRD or CG were selected differently for each model.
Conclusions: The application of ML model has been shown to be an efficient methodology in predicting antibiotic-induced nephrotoxicity. This methodology is expected to be applicable to predicting nephrotoxicity caused by other drugs including antibiotics.
References:
[1] Mehta R. L. et al. Phenotype standardization for drug-induced kidney disease. Kidney Int. (2015) 88(2), 226–234.
[2] Campbell R. E. et al. Overview of antibiotic-induced nephrotoxicity Kidney Int Rep. (2023) 8, 2211-2225.
[3] Team R. C. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. (2022).
[4] Hothorn T. et al. “partykit: A modular toolkit for recursive partytioning in R.” JMLR (2015) 16(1), 3905-3909.
[5] Liaw A. et al. Classification and Regression by random Forest. R News (2002) 2(3), 18-22.
[6] Chen T. et al. Xgboost: Extreme Gradient Boosting. R package version 1.7.7.1 (2024).
Reference: PAGE 32 (2024) Abstr 11093 [www.page-meeting.org/?abstract=11093]
Poster: Methodology – AI/Machine Learning