Comparative evaluation of 11 machine learning-based risk prediction models for vancomycin-associated acute kidney injury during initial dosing

Comparative evaluation of 11 machine learning-based risk prediction models for vancomycin-associated acute kidney injury during initial dosing | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 19 January 2026 V1 Latest version Share on Comparative evaluation of 11 machine learning-based risk prediction models for vancomycin-associated acute kidney injury during initial dosing Authors : Moeko Iida 0009-0001-5814-6825 , Yasuhiro Horita 0000-0003-1191-8792 , Masato NODA , Yoshinori HISADA , Masaya NAGAMIZU , Kaori TSUZUKI , Sakurako MURAMATSU , … Show All … , Kazuki OHASHI 0000-0002-2606-6468 , Yuki NOMURA , Minami ASAOKA , Tomoyo MIZUNO , Hiroki ASAKURA , Tomoaki HAYAKAWA , Chiharu WACHINO , Masaharu KUDO , Yoshihisa MIMURA , Taketo MIYAMOTO , Masami KAWAHARA , Nobuyuki MORISHITA , Masahiro Kondo , Yuji HOTTA , Atsushi NAKAMURA , and YOKO FURUKAWA-HIBI 0000-0002-2126-5652 [email protected] Show Fewer Authors Info & Affiliations https://doi.org/10.22541/au.176878282.28318775/v1 143 views 98 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Aim : Early vancomycin-associated acute kidney injury (AKI) identification is important during the initial dosing period when non–steady-state pharmacokinetic exposure may influence nephrotoxicity risk. We aimed to conduct a multicentre retrospective cohort study of hospitalised adult patients who received vancomycin at three tertiary care hospitals in Japan between April 2019 and March 2024. Methods : Early area under the concentration–time curve (AUC) metrics, trough concentrations, concomitant medications, and comorbidities were used to develop and compare machine learning models to predict vancomycin-associated AKI. Eleven algorithms were evaluated using a nested cross-validation framework to minimise optimistic bias and prevent data leakage. Results : Among 442 patients (median age 72 years), AKI occurred in 5.7%. Across models, day 2 AUC and trough concentration consistently emerged as the most influential predictors, followed by concomitant use of piperacillin/tazobactam, diuretics, and vasopressors. The random forest model showed the best overall discriminative performance (area under the receiver operating characteristic curve 0.89; area under the precision–recall curve 0.47). Shapley additive explanations revealed nonlinear increases in AKI risk at approximately day 2 AUC ≥ 550 mg·h/L and trough concentration ≥ 20 μg/mL. These findings indicate that early vancomycin exposure, particularly cumulative exposure within the first 48 h, is strongly associated with AKI risk. Conclusion : Machine learning models incorporating early AUC metrics with trough concentrations may provide a clinically informative framework for early risk stratification during vancomycin therapy and support further evaluation of early AUC-guided dosing strategies with trough concentrations as complementary exposure markers. Title Comparative evaluation of 11 machine learning-based risk prediction models for vancomycin-associated acute kidney injury during initial dosing Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher Authors Moeko IIDA 1, 2 , Yasuhiro HORITA 1, 2, 3# , Masato NODA 4 , Yoshinori HISADA 5 , Masaya NAGAMIZU 5 , Kaori TSUZUKI 4 , Sakurako MURAMATSU 6 , Kazuki OHASHI 2, 3 , Yuki NOMURA 2 , Minami ASAOKA 1, 2 , Tomoyo MIZUNO 5 , Hiroki ASAKURA 5 , Tomoaki HAYAKAWA 1, 2, 3 , Chiharu WACHINO 1, 4 , Masaharu KUDO 7 , Yoshihisa MIMURA 1, 2 , Taketo MIYAMOTO 1 , Masami KAWAHARA 6 , Nobuyuki MORISHITA 5 , Masahiro KONDO 1, 4 , Yuji HOTTA 1, 2 , Atsushi NAKAMURA 3 , and Yoko FURUKAWA‐HIBI 1, 2# Affiliations 1 Department of Clinical Pharmaceutics, Graduate School of Medical Sciences, Nagoya City University, Aichi, Japan 2 Department of Pharmacy, Nagoya City University Hospital, Aichi, Japan 3 Division of Infection Prevention and Control, Nagoya City University Hospital, Aichi, Japan 4 Department of Pharmacy, Nagoya City University East Medical Center, Aichi, Japan 5 Department of Pharmacy, Nagoya City University West Medical Center, Aichi, Japan 6 School of Pharmacy, Aichi Gakuin University, Aichi, Japan 7 Department of Pharmacy, Nagoya City University Midori Municipal Hospital, Aichi, Japan Principal investigator statement As this study was non-interventional, no individual was designated as the principal investigator. Corresponding author # These authors contributed equally to this article and share the corresponding authorship. Yoko FURUKAWA‐HIBI, PhD. Department of Clinical Pharmaceutics, Graduate School of Medical Sciences, Nagoya City University, Aichi, Japan; Department of Pharmacy, Nagoya City University Hospital, Aichi, Japan 1 Kawasumi, Mizuho-cho, Mizuho-ku, Nagoya, Aichi 467-8602, Japan E-mail: [email protected] Yasuhiro HORITA, PhD. Department of Clinical Pharmaceutics, Graduate School of Medical Sciences, Nagoya City University, Aichi, Japan One Kawasumi, Mizuho-cho, Mizuho-ku, Nagoya, Aichi 467-8602, JAPAN E-mail: [email protected] Keywords: Vancomycin, acute kidney injury, machine learning, initial dosing design Word count: Table count: 4 Figure count: 3 Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher Bullet point summary What is already known about this subject? • Vancomycin-associated AKI remains a clinically important adverse event, particularly during the early phase. • AUC-guided therapeutic drug monitoring is used to reduce nephrotoxicity; however, the optimal exposure thresholds during the non–steady-state period are not well defined. • Early AUC-based measures have not been used for predicting vancomycin-associated AKI. What this study adds • Early vancomycin exposure is the most influential predictor of AKI across multiple machine learning algorithms. • Trough concentrations provide complementary, non-redundant information when combined with early AUC metrics, improving risk prediction during initial dosing. • Nonlinear risk patterns identified by SHAP analyses suggest clinically relevant exposure thresholds. Aim: Early vancomycin-associated acute kidney injury (AKI) identification is important during the initial dosing period when non–steady-state pharmacokinetic exposure may influence nephrotoxicity risk. We aimed to conduct a multicentre retrospective cohort study of hospitalised adult patients who received vancomycin at three tertiary care hospitals in Japan between April 2019 and March 2024. Methods: Early area under the concentration–time curve (AUC) metrics, trough concentrations, concomitant medications, and comorbidities were used to develop and compare machine learning models to predict vancomycin-associated AKI. Eleven algorithms were evaluated using a nested cross-validation framework to minimise optimistic bias and prevent data leakage. Results: Among 442 patients (median age 72 years), AKI occurred in 5.7%. Across models, day 2 AUC and trough concentration consistently emerged as the most influential predictors, followed by concomitant use of piperacillin/tazobactam, diuretics, and vasopressors. The random forest model showed the best overall discriminative performance (area under the receiver operating characteristic curve 0.89; area under the precision–recall curve 0.47). Shapley additive explanations revealed nonlinear increases in AKI risk at approximately day 2 AUC ≥ 550 mg·h/L and trough concentration ≥ 20 μg/mL. These findings indicate that early vancomycin exposure, particularly cumulative exposure within the first 48 h, is strongly associated with AKI risk. Conclusion: Machine learning models incorporating early AUC metrics with trough concentrations may provide a clinically informative framework for early risk stratification during vancomycin therapy and support further evaluation of early AUC-guided dosing strategies with trough concentrations as complementary exposure markers. Introduction Vancomycin remains an important first-line antibiotic for treating infections caused by methicillin-resistant Staphylococcus aureus (MRSA), Enterococcus spp., and methicillin-resistant coagulase-negative Staphylococcus (MR-CoNS). The 2020 US guidelines for therapeutic drug monitoring (TDM) of vancomycin recommend that traditional trough concentration-based therapeutic monitoring be replaced by an approach guided by the ratio of the area under the concentration–time curve over 24 h to the minimum inhibitory concentration (AUC 24h /MIC) in routine clinical practice. 1 In 2022, AUC-guided dosing of vancomycin was introduced in Japan based on evidence derived from systematic reviews and meta-analyses related to the clinical effectiveness and safety of AUC-guided monitoring. 2-4 The target daily AUC was set at 400–600 mg·h/L during initial dosing to maximise efficacy and reduce the risk of acute kidney injury (AKI). 5 The AUC 24h /MIC has been reported to correlate with the early clinical response, reduced 30-d mortality, and lower incidence of treatment failure in infections with Staphylococcus spp. and Enterococcus spp. 6-9 The estimated cutoff points vary depending on the infectious diseases and isolated bacteria, for example, 421 mg·h/L for MRSA bacteremia, 389 mg·h/L for E. faecium bacteremia, and 230 mg·h/L for MR-CoNS bloodstream infections. 10-12 The reported incidence of vancomycin-associated AKI varies from 5% to 43% among individual facilities. 13 In general, AKI is associated with poor clinical outcomes and increased mortality, particularly in critically ill patients. 14 Moreover, nephrotoxicity can extend the hospital stay and increase the cost of additional treatment. 15,16 In addition, the impaired renal function recovery after vancomycin dosing is associated with worsening survival outcomes during hospitalization and within 1 year. 17 A recent multicentre, retrospective cohort study conducted in Japan reported that 77.4% of the patients developed AKI within 7 d during the term of vancomycin administration. 18 Therefore, careful monitoring early in the course of vancomycin administration is warranted to reduce the risk of AKI. To date, several reports have described the thresholds of a steady-state AUC (AUC ss ), which could reduce the risk of nephrotoxicity; however, a safe upper limit of daily AUC at an early stage of vancomycin dosing remains controversial. 19-23,24 We have previously identified and reported optimal cutoff points for day 1 AUC (AUC 0–24h ), day 2 AUC (AUC 24–48h ), 2-d AUC (AUC 0–48h ), and trough levels at the first sampling point. 25 The best predictor for vancomycin-associated AKI in an early stage of administration was day 2 AUC with a threshold of 554.8 mg·h/L. Notably, both day 2 AUC and trough concentration were selected as crucial predictors of AKI in the classification and regression tree (CART) analysis. Although AUC-guided vancomycin dosing is supported globally, 1,5,26,27 the necessity of monitoring trough concentrations has been disputed. 28-30 Recently, blended vancomycin dosing has been associated with a reduced risk of AKI, lower blood sampling volume, and decreased cumulative vancomycin dose. 31 Hence, trough concentrations along with AUC may be a better approach to avoid vancomycin-associated AKI. Machine learning and artificial intelligence have recently demonstrated notable utility, with their application within the clinical pharmacology expanding rapidly. 32 These methodologies are increasingly regarded as essential for optimising the risk-benefit profile in in personalised medicine and model-informed precision dosing (MIPD). 33 Furthermore, Bayesian-based AUC-guided dosing and monitoring are MIPD approaches. 5 While several machine learning models for predicting vancomycin-associated AKI have been published, the majority are predicted on conventional trough concentrations. 34-36 Furthermore, the development of robust models utilising non-steady exposure metrics, such as day 1 AUC and day 2 AUC, remains unestablished. In this study, we conducted a comprehensive, methodologically focused comparison of machine learning-based models for the early prediction of vancomycin-associated AKI. Our analysis incorporated non-steady-state exposure metrics, such as day 1 and day 2 AUC, alongside trough concentrations and established clinical risk factors. Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher Methods Patients and study design A multicentre, retrospective cohort study was conducted among hospitalised patients treated with vancomycin for 2 d or more at Nagoya City University Hospital (NCUH), East Medical Center (EMC), and West Medical Center (WMC), 1800-bed, core medical institutions were involved in Nagoya, Japan, between April, 2019, and March, 2024 (Figure 1) . Patients with at least one blood concentration measurement were included. The main exclusion criteria were as follows: (i) patients aged < 18 years, (ii) patients receiving haemodialysis or undergoing continuous renal replacement therapy, and (iii) patients who experienced AKI in the first 2 d of therapy (Figure 1) . Vancomycin-associated AKI was defined as an increase in serum creatinine (Scr) levels of 0.3 mg/dL over a 48-h period or a > 50% increase above baseline, according to our previous study and the modified Kidney Disease Improving Global Outcomes (KDIGO) criteria. 1,25,37 We obtained clinical laboratory data, operation records, types of infectious diseases, detected bacteria, and comorbidities such as type 2 diabetes mellitus (DM) and heart failure from electronic medical records. The following concomitant drugs were considered potential nephrotoxic drugs: piperacillin/tazobactam, aminoglycosides, amphotericin B, levofloxacin, trimethoprim/sulfamethoxazole, acyclovir, diuretics, vasopressors (adrenaline, noradrenaline, dopamine, dobutamine, etilefrine, and phenylephrine), angiotensin-converting enzyme inhibitors, angiotensin receptor blockers, nonsteroidal anti-inflammatory drugs, intravenous contrast agents, ifosfamide, and cisplatin. 1,38 Measurement of serum vancomycin concentrations The initial dosing schedule and blood sampling were performed in compliance with the Japanese TDM guidelines published in 2016 and 2022. 5 Most of the first blood samplings were performed between days 2 and 3. The trough levels were defined as the concentrations at ≤ 30 min before the next dose. Serum vancomycin concentrations were measured using the ARCHITECT iVAN assay kit (Abbott Japan Co., Ltd., Chiba, Japan) at the NCUH and LUMIPULSE G1200 at EMC and WMC (FUJIREBIO Inc., Kanagawa, Japan). 39 Pharmacokinetic analysis Population pharmacokinetic (PK) parameters were estimated using SAKURA-TDM, a Bayesian-based TDM support software that incorporates the population PK model developed by Yasuhara et al. 39,40 The laboratory data obtained before vancomycin dosing were used to derive population and individual PK parameters and calculate the day 1 AUC (AUC 0–24h ) and day 2 AUC (AUC 24–48h ) according to the actual dosing schedule. The 2-d AUC (AUC 0–48h ) was calculated using the following equation: AUC 0–24h + AUC 24–48h . Creatinine clearance and estimated glomerular filtration rate (eGFR) were calculated using the Scr values. Machine learning model development and validation A machine learning model was developed using data from the NCUH, whereas external validation was performed using independent cohorts from the EMC and WMC. We adopted a nested cross-validation (CV) framework consisting of five outer folds and three inner folds to obtain unbiased estimates of model performance and prevent data leakage. The inner CV was used exclusively for hyperparameter tuning, whereas the outer folds were used for the final validation. Each outer loop split the data into training and testing subsets (80:20). Given the low incidence of AKI (approximately 6%), class imbalance was handled using two complementary strategies: (1) random oversampling of the minority class at predefined ratios (0.15, 0.30, and 0.50), 41 and (2) class-weighting schemes (balanced and sqrt) based on inverse class frequencies. 42 Both were applied only within the training partition of the inner CV loop to avoid information leakage. With respect to C5.0, cost-sensitive learning (CSL) and adaptive boosting methods were adapted according to previous reports. 43 The area under the precision–recall curve (AUPRC), computed using the Davis–Goadrich method, was used as the primary tuning metric because of class imbalance. The combination of oversampling ratio, weighting scheme, and hyperparameters that yielded the highest mean AUPRC in the inner cross-validation was selected and retrained on the full training set of each outer fold. The reported metrics included the area under the receiver operating characteristic curve (AUROC), AUPRC, accuracy (95% confidence interval [CI]), precision, recall, and F1 score. All metrics were calculated from the confusion matrix of each outer test set. The optimal probability threshold was determined as the value that maximised the F1 score of the outer folds. The feature importance was extracted from the final ML model using the built-in feature importance function and summarised across the dummy encoded variables. Candidate variables were selected based on previously identified AKI risk factors from our earlier multivariate logistic regression analysis, 25 which provided a clinically grounded filter-type preselection framework. We constructed a least absolute shrinkage and selection operator (LASSO)-reduced model comprising five predictors and compared its performance with that of the full 23-variable model to evaluate the impact of dimensionality reduction (Table S1). This comparison assessed whether a more parsimonious feature set could maintain the predictive performance while improving interpretability. A nine-fold cross-validation was performed to identify the optimal lambda value of 0.025 in the final LASSO regression model. The risk prediction models for vancomycin-associated AKI were constructed using 11 machine-learning algorithms: CART, C5.0 decision tree with adaptive boosting (C5.0), conditional inference tree (CTree), Naïve Bayes (NB), neural network (NN), support vector machine with radial basis function kernel (SVM-RBF), support vector machine with linear kernel (SVM-LN), random forest (RF), eXtreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and CatBoost. The feature attribution was quantified using Shapley values. The kernel SHAP was computed for the final RF classifiers using the kernelshap package and visualised using Shapviz in R (v4.4.2; R Foundation for Statistical Computing, Vienna, Austria). Variable importance and interaction (VIVI) networks were generated using the Vivid package in R to visualise how the predictors interacted within the RF decision function. Statistical analysis The Shapiro–Wilk test or Kolmogorov–Smirnov test was used to determine the normality of each numerical variable according to the sample size. Statistical significance was assessed using a two-sample t -test, Mann–Whitney U -test, Kruskal–Wallis rank-sum test, or Fisher’s exact test. P values < 0.05 were considered statistically significant. All statistical analyses were performed using R Studio (v.4.4.2; R Foundation for Statistical Computing, Vienna, Austria). Ethics statement This study was conducted in compliance with the Declaration of Helsinki and approved by the Clinical Trial Management Centre of NCUH (management no. 60-22-0117). The requirement for written informed consent was waived by the Institutional Review Board for this retrospective study, with public notification of the study. Results Baseline clinical characteristics A total of 347 and 95 patients were included in model building and external validation, respectively (Tables 1 and S2). Additional information on the concomitant drugs and isolated bacteria is presented in Table S3. The median age and body weight were 72 (interquartile range [IQR], 60–79) years and 52.3 (IQR, 44.2–61.2) kg, respectively, in the model-building cohorts. Overall, 68.3% of the patients were aged ≥ 65 years, and 67.7% were male. The dosing duration until the first TDM was 40.0 (IQR, 36.0–60.0) h. Bacteremia and pneumonia were the most frequent infectious diseases in both cohorts. The proportion of patients admitted to the intensive care unit was 6.1%. At the first sampling point, 18 (5.2%) patients had AKI. Patients were classified into the following stages according to the KDIGO criteria: stage 1 ( n = 10), stage 2 ( n = 4), and stage 3 ( n =4). Two patients had eGFR < 30 mL/min/1.73 m 2 before vancomycin dosing. No significant differences were noted in the baseline demographics or laboratory findings between the AKI and non-AKI groups, except for serum creatinine levels, renal function indices (eGFR and BUN/Scr), and individual PK parameters (Table 1). Regarding concomitant medications (Table S3), piperacillin/tazobactam, diuretics, and vasopressors were used significantly more frequently in the AKI group. MRSA was the most common organism among the isolated pathogens, and was significantly more frequent in the AKI group in the validation cohort ( P = 0.034). Machine learning model building and validation We constructed and evaluated 11 ML algorithms using a nested cross-validation framework to identify the optimal predictive models for vancomycin-associated AKI. The optimised hyperparameters for each algorithm are listed in Table S4. Before model training, LASSO regression was applied to prevent overfitting and enhance interpretability, yielding five key predictors: day 2 AUC, trough concentration, and concomitant use of piperacillin/tazobactam, diuretics, and vasopressors (Figure S1). For comparison, models were developed using the full set of 23 candidate variables without previous feature selection. Among the ML models, the RF model achieved the highest performance for AKI prediction using the full feature set, with a mean AUPRC of 0.47 ± 0.27 and AUROC of 0.89 ± 0.09 (Table 2). When restricted to the LASSO-selected feature set, the overall model performance remained comparable to, or slightly improved, for several algorithms. Notably, RF achieved the highest AUPRC (0.53 ± 0.17), followed by NB (0.50 ± 0.30) and XGBoost (0.48 ± 0.30) (Table 2). Therefore, ensemble learning approaches demonstrated favourable performance characteristics for early AKI prediction within the present dataset. Among the ensemble tree-based algorithms, LightGBM and CatBoost achieved the highest AUPRCs on the validation data using the full feature set (0.65 and 0.66, respectively) (Table 3), with high AUROC values (LightGBM = 0.79, CatBoost = 0.89) (Table 3). Similarly, NB performed well (AUPRC = 0.69, AUROC = 0.94), demonstrating a stable performance across the evaluated cohorts (Table 3). The performance gap between the training and validation datasets was small for most algorithms. The overall model performance remained stable or improved for several algorithms after LASSO-based variable selection. Although NB demonstrated the highest AUPRC (0.67) in the LASSO-based validation dataset, RF demonstrated a comparable discriminative ability (AUPRC = 0.61) with superior AUROC and F1 scores, indicating better overall predictive stability (Table 3). Collectively, these findings suggest that a reduced set of predictors can retain sufficient information for AKI risk prediction, with minimal loss of discriminative performance. Across the 11 machine learning models, the day 2 AUC consistently displayed the highest feature importance, regardless of whether variable selection was performed using LASSO regression (Tables S5 and S6). In addition, both trough levels and 2-day AUC were identified as key secondary predictors, highlighting the significance of early vancomycin exposure across the models. Among the concomitant drugs, piperacillin/tazobactam demonstrated a relatively high importance. In contrast, laboratory variables, such as CrCl, serum albumin, and BUN/Scr, exhibited moderate contributions, whereas demographic or background variables (age, sex, and body weight) and other concomitant medications, such as diuretics and vasopressors, generally exhibited lower importance (Table S5). A leave-one-feature-out analysis was conducted using the RF algorithm (Table 4) to further evaluate the relative contribution of the day 2 AUC and trough levels to model performance. The exclusion of either variable consistently reduced AUPRC and F1 scores across both training and validation datasets, especially in the LASSO-based model. On the validation data, the exclusion of day 2 AUC caused the most notable deterioration in model performance, with AUPRC decreasing from 0.64 to 0.62 in the full feature model and from 0.61 to 0.47 in the LASSO-based model (Table 4). Removal of trough concentration had minimal impact on the full feature model (AUPRC decreased from 0.64 to 0.65); however, it caused a more distinct decline in the LASSO-based model (AUPRC decreased from 0.61 to 0.54). These findings indicate that both the day 2 AUC and trough concentration provide complementary and non-redundant information, reflecting distinct aspects of vancomycin exposure relevant to AKI risk prediction. SHAP analysis of the RF model using all 23 variables identified day 2 AUC as the most influential variable, followed by the 2-d AUC, trough levels, and piperacillin/tazobactam co-administration (Figure 2). The ranking remained consistent in the LASSO-reduced RF, including five variables: day 2 AUC > trough levels > piperacillin/tazobactam > diuretics > vasopressors; however, the magnitude of the effect of each feature increased (Figure S2). SHAP dependence plots demonstrated monotonic, nonlinear increases in AKI risk with higher exposure metrics, displaying inflection zones around day 2 AUC 550 mg·h/L, 2-d AUC 1100 mg·h/L, and trough level 20 μg/mL (Figures 3A, 3B, 3C). In addition, the BUN/Scr ratio exhibited a threshold-like positive pattern (Figure 3E). For binary covariates, the SHAP distributions shifted upward in the presence of piperacillin/tazobactam, diuretics, or vasopressors, indicating an elevated risk of AKI with these cotreatments (Figure S3). The variable importance and variable interaction (VIVI) network provided a graphical overview of inter-feature dependencies based on conditional SHAP interaction values. The network revealed dense interconnections among the early exposure indices (day 1 AUC, day 2 AUC, 2-d AUC, and trough levels) and renal function markers (Scr, BUN, and CrCl) in the full 23-variable RF model (Figure S4A), suggesting a strong collinearity and synergistic influence on AKI predictions. Day 2 AUC demonstrated the largest node degree and interaction strength, aligning with its leading marginal SHAP importance. In contrast, the LASSO-restricted model (Figure S4B) yielded a more compact structure, with a day 2 AUC and trough forming a tightly linked core connected by the strongest positive interaction edge. This configuration reflects the overlapping yet nonredundant contributions of AUC and trough levels to the predicted AKI risk. Discussion In this multicentre study, we comprehensively evaluated multiple machine learning algorithms to compare candidate predictors of vancomycin-associated AKI. Early vancomycin exposure indices, particularly the day 2 AUC and trough levels, were consistently the strongest predictors of AKI across the models, highlighting the critical role of cumulative exposure within the initial dosing period. To the best of our knowledge, this is the first report to construct a predictive model for vancomycin-associated AKI by considering both day 2 AUC and trough levels. Consistent with multivariate logistic regression analysis, the concomitant use of piperacillin/tazobactam, diuretics, or vasopressors was identified as a risk factor in the machine learning models. 25 Notably, the selected factors were similar to those previously reported. 34,35,44 Machine learning models have emerged as valuable tools in several medical fields and industry. 45 The techniques are considered attractive for assisting precision dosing using Bayesian forecasting tools. 5,46 Although several types of machine learning algorithms have become available in recent years, an optimal machine learning model may depend on the population, sample size, and purpose of use. Therefore, we tested 11 machine learning models and compared their performances. The consistent predominance of day 2 and 2-d AUC across the models highlights the pivotal role of early vancomycin exposure in the onset of AKI. These findings suggest that cumulative exposure within the first 48 h is a critical determinant of AKI risk. In addition, although trough levels contributed substantially to the model predictions, their importance was lower than that of the AUC-based metrics. These findings indicate that AUC-guided monitoring provides a physiologically grounded framework for model-based assessment of vancomycin-associated AKI risk in the present analytical setting. Consistent with the leave-one-feature-out analysis, both day 2 AUC and trough levels provided independent and complementary predictive information for AKI risk. The decline in model performance following the exclusion of either variable indicates that these indices capture distinct and complementary aspects of vancomycin exposure dynamics: day 2 AUC reflecting cumulative systemic burden over the initial dosing period and trough concentration representing sustained exposure and dosing-interval–dependent clearance status. Notably, the absence of a performance decline after removing the trough concentration in the full-feature model suggests that other exposure-related variables, such as day 1 AUC and 2-d AUC, sufficiently captured the overall vancomycin burden within this broader feature space. This complementary relationship reinforces the clinical rationale for incorporating early AUC monitoring and trough assessment, particularly during the critical initial dosing period. From a pharmacokinetic–pharmacodynamic perspective, the stronger contribution of AUC-based metrics aligns with the mechanistic evidence linking total vancomycin exposure to AKI risk. Although trough levels remain a practical surrogate for therapeutic monitoring, their predictive capacity is limited when evaluated alone. Therefore, early estimation of the AUC within the first 48 h may enable more precise dose optimization and timely identification of high-risk patients, providing a potentially clinically informative strategy to minimise vancomycin-associated AKI. Model explanations consistently highlight early vancomycin exposure, especially day 2 AUC, as the primary determinant of AKI risk, with trough levels providing complementary information that reflects sustained exposure and dosing interval/clearance. Nonlinear SHAP patterns indicated that AKI risk increased sharply with high vancomycin exposure (day 2 AUC ≥ 550 mg·h/L or trough ≥ 20 µg/mL), supporting early AUC-guided dose adjustment within the first 48 h. The minimal performance impact of removing the trough concentration in the 23-variable model suggested an overlap among the exposure features (day 1 AUC, day 2 AUC, and 2-d AUC). By contrast, the larger decline in the 5-variable model implies that the trough remains informative when the feature space is reduced, complementing the AUC to capture exposure persistence. VIVI networks visualise how vancomycin exposure metrics serve as central hubs linking pharmacokinetic intensity (AUC and trough levels) and renal vulnerability. In the full model, dense connections among exposure and renal indices indicated that the model integrated both the exposure burden and renal sensitivity to infer the probability of AKI. In contrast, the simplified 5-variable model preserved the core day 2 AUC–trough axis while revealing clinically interpretable modifiers such as piperacillin/tazobactam, diuretics, and vasopressors. A simplified VIVI network based on LASSO selection demonstrated that a compact feature set effectively captured key nonlinear and interactive factors associated with vancomycin-induced AKI. Six years have passed since the current clinical guidelines on AUC-guided vancomycin dosing were published. 1,5,26,27 The current TDM guidelines state that trough-only monitoring targeting between 15 and 20 μg/mL is no longer recommended. 1,5 To date, machine learning models for predicting vancomycin-associated AKI have been proposed based on only using trough levels or AUC ss . 36,47 In this study, we demonstrated that the model performances of AUC-based models were superior to those based on trough levels. The model performance improved when both the day 2 AUC and trough levels were considered. Vancomycin trough concentrations correlate poorly with clinical and microbiological efficacy outcomes, whereas a systematic review and meta-analysis have demonstrated a strong association with nephrotoxicity. 3,13,48 Collectively, constructing a risk prediction model for vancomycin-associated AKI using both the AUC and trough levels is a good option. This study had certain limitations. First, it was conducted in a single-country cohort with a relatively homogeneous ethnicity, and the findings may not be directly applicable to multi-ethnic or international populations. Vancomycin dosing strategies, renal function assessment, and AKI risk management may differ across ethnic groups and healthcare systems. Therefore, external validation in more diverse populations is needed to confirm the generalizability of our model and assess potential race- or ethnicity-related biases that were not captured in the present analysis. Second, our feature-selection approach depends on LASSO regression, which is an embedded method. Different feature selection strategies, such as filter or wrapper methods, may yield slightly different sets of predictors. Moreover, the variables excluded by LASSO in this dataset could still be relevant to other populations or under different modelling approaches. Therefore, the generalizability of the selected predictors should be interpreted with caution, and future studies should confirm whether these predictors remain stable across diverse patient populations and modelling strategies. Third, the individual AUCs on day 1 may not have been correctly estimated because most blood samples were drawn between days 2 and 3. In conclusion, we developed a machine learning-based risk prediction model for vancomycin-associated AKI during the early phases of therapy. Our findings highlight the critical role of both the day 2 AUC and trough levels as complementary predictors of nephrotoxicity risk. These results provide a methodological basis for future studies integrating machine learning with pharmacokinetic modelling to advance the precise dosing of vancomycin. Acknowledgments We thank the study participants and supportive medical doctors of the Division of Infection Prevention and Control at Nagoya City University Hospital, who helped conduct the current study. In addition, we thank the clinical pharmacists for their assistance with data collection and processing. ChatGPT (GPT-5, OpenAI) was used for language editing and refinement. The final manuscript has been reviewed and approved by the authors. Author contributions statement All listed authors meet the International Committee of Medical Journal Editors (ICMJE) criteria for authorship. Y. Horita, M.I., Y. Furukawa‐Hibi, and A.N. wrote the manuscript; Y. Horita, M.I., M.N., Y. Hisada, N.M., and M. Kawahara designed the study; K.T., S.M., K.O., Y.N., T.M., T.H., M.A., and H.A. performed the data collection and research; Y. Horita, M.I., Y.M., and T.M. performed the data analyses; and C.W. and M. Kudo provided information about the machine learning technique; Y. Hotta, M. Kondo, and N.M. drafted and critically revised the manuscript for important intellectual content. All the authors have read and approved the final version of this manuscript. Disclosure Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher The preliminary results were presented orally at the 40th Annual Meeting of the Japanese Society of Therapeutic Drug Monitoring (abstract number 11), the 23rd Congress of the International Association of Therapeutic Drug Monitoring and Clinical Toxicology (abstract number 212), and IDWeek 2025 (abstract P-1250). Funding This work was supported by the Grants-in-Aid for Research provided by JSPS KAKENHI (grant numbers: JP 23K11922, JP 25H00279, and JP 23K06238). Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher Transparency declarations We would like to thank Editage (www.editage.jp) for English language editing. Y.FH. received a speaking fee at a meeting organised on behalf of Daiichi Sankyo Co., Ltd. A.N. received speaking fees at meetings organised on behalf of Meiji Seika Pharma Co., Ltd. and Shionogi & Co., Ltd. Hybrid Field Variables C. R. Gimarelli (December 25, 2025) \affiliation Independent Researcher Conflicts of interests The authors declare no conflicts of interest in association with the present study. References 1. Rybak, MJ, Le J, Lodise TP, et al . Therapeutic monitoring of vancomycin for serious methicillin-resistant Staphylococcus aureus Infections: a revised consensus guideline and review by the American Society of Health-system Pharmacists, the Infectious Diseases Society of America, the Pediatric Infectious Diseases Society, and the Society of Infectious Diseases Pharmacists. Clin. Infect. Dis. 2020; 71 :1361–1364. 2. Tsutsuura, M, Moriyama H, Kojima N, et al . The monitoring of vancomycin: a systematic review and meta-analyses of area under the concentration-time curve-guided dosing and trough-guided dosing. BMC Infect. Dis. 2021; 21 :153. 3. Prybylski, J.P. Vancomycin trough concentration as a predictor of clinical outcomes in patients with Staphylococcus aureus bacteremia: a meta-analysis of observational studies. Pharmacotherapy . 2015; 35 :889–898. 4. Tongsai S & Koomanachai P. The safety and efficacy of high versus low vancomycin trough levels in the treatment of patients with infections caused by methicillin-resistant Staphylococcus aureus : a meta-analysis. BMC Res. Notes . 2016; 9 :455. 5. Matsumoto K, Oda K, Shoji K, et al . Clinical practice guidelines for therapeutic drug monitoring of vancomycin in the framework of model-informed precision dosing: a consensus review by the Japanese Society of Chemotherapy and the Japanese Society of Therapeutic Drug Monitoring. Pharmaceutics . 2022; 14 :489. 6. Pitayakittiwong C, Sermsappasuk P, Meesing A, & Jaisue, S. Association of vancomycin AUC/MIC and trough concentration with early clinical response in Enterococcus or coagulase-negative Staphylococcus infection: A prospective study. J. Clin. Pharmacol. 2024; 64 :1006–1014. 7. Katip W, Lee SWH, Kasatpibal N. & Rayanakorn A. Systematic review and meta-analysis of vancomycin therapeutic level for treatment of vancomycin-sensitive enterococcal infections. Br. J. Clin. Pharmacol. 2025;91:1250–1262. 8. Holmes NE, Turnidge JD, Munckhof WJ, et al. Vancomycin AUC/MIC ratio and 30-day mortality in patients with Staphylococcus aureus bacteremia. Antimicrob. Agents Chemother. 2013; 57 :1654–1663. 9. Ueda T, Takesue Y, Nakajima K, et al . Validation of vancomycin area under the concentration-time curve estimation by the Bayesian approach using one-point samples for predicting clinical outcomes in patients with methicillin-resistant Staphylococcus aureus infections. Antibiotics (Basel) . 2022; 11 :96. 10. Kullar R, Davis SL, Levine DP & Rybak MJ. Impact of vancomycin exposure on outcomes in patients with methicillin-resistant Staphylococcus aureus bacteremia: support for consensus guidelines suggested targets. Clin. Infect. Dis. 2011; 52 :975–981. 11. Tochikura N, et al. Pharmacokinetic/pharmacodynamic analysis of vancomycin in patients with Enterococcus faecium bacteraemia: a retrospective cohort study. Eur. J. Hosp. Pharm. 2024; 31 :440–446. 12. Niinuma Y, Kagami K, Sugawara M, & Takekuma Y. The crucial relationship between vancomycin minimum inhibitory concentration and therapeutic efficacy against methicillin-resistant coagulase-negative Staphylococci . J. Chemother. 2025;37:326–333. 13. van Hal SJ, Paterson DL, & Lodise TP. Systematic review and meta-analysis of vancomycin-induced nephrotoxicity associated with dosing schedules that maintain troughs between 15 and 20 milligrams per liter. Antimicrob. Agents Chemother. 2013; 57 :734–744. 14. Case J, Khan S, Khalid R, & Khan A. Epidemiology of acute kidney injury in the intensive care unit. Crit. Care Res. Pract. 2013 ;479730. 15. Barlam TF, Cosgrove SE, Abbo LM, et al . Implementing an antibiotic stewardship program: guidelines by the Infectious Diseases Society of America and the Society for Healthcare Epidemiology of America. Clin. Infect. Dis. 2016; 62 :e51–e77. 16. Fernández de Gatta MD, Calvo MV, Hernández JM, Caballero D, San Miguel JF, & Domínguez-Gil A. Cost-effectiveness analysis of serum vancomycin concentration monitoring in patients with hematologic malignancies. Clin. Pharmacol. Ther. 1996; 60 :332–340. 17. Chuma M, Hamano H, Bando T, et al . Non-recovery of vancomycin-associated nephrotoxicity is related to worsening survival outcomes: combined retrospective analyses of two real-world databases. Basic Clin. Pharmacol. Toxicol. 2022; 131 :525–535. 18. Hashimoto N, Kimura T, Hamada Y, et al . Candidates for area under the concentration-time curve (AUC)-guided dosing and risk reduction based on analyses of risk factors associated with nephrotoxicity in vancomycin-treated patients. J. Glob. Antimicrob. Resist. 2021; 27 :12–19. 19. Zasowski EJ, Murray KP, Trinh TD, et al . Identification of vancomycin exposure-toxicity thresholds in hospitalized patients receiving intravenous vancomycin. Antimicrob. Agents Chemother. 2018; 62 :e01684–e01617. 20. Moise-Broder PA, Forrest A, Birmingham MC, & Schentag JJ. Pharmacodynamics of vancomycin and other antimicrobials in patients with Staphylococcus aureus lower respiratory tract infections. Clin. Pharmacokinet. 2004; 43 :925–942. 21. Brown J, Brown K, & Forrest A. Vancomycin AUC 24 /MIC ratio in patients with complicated bacteremia and infective endocarditis due to methicillin-resistant Staphylococcus aureus and its association with attributable mortality during hospitalization. Antimicrob. Agents Chemother. 2012; 56 :634–638. 22. Gawronski KM, Goff DA, Brown J, Khadem TM, & Bauer KA. A stewardship program’s retrospective evaluation of vancomycin AUC 24 /MIC and time to microbiological clearance in patients with methicillin-resistant Staphylococcus aureus bacteremia and osteomyelitis. Clin. Ther. 2013; 35 :772–779. 23. Endo. A, Hanawa K, Asakawa D, et al . Potential risk factors for early acute kidney injury in patients treated with vancomycin. J. Infect. Chemother. 2024; 30 :989–994. 24. Lodise, TP, Rosenkranz SL, Finnemeyer M, et al . The Emperor’s new clothes: prospective observational evaluation of the association between initial vancomycin exposure and failure rates among adult hospitalized patients with methicillin-resistant Staphylococcus aureus bloodstream infections (PROVIDE). Clin. Infect. Dis. 2020; 70 :1536–1545. 25. Iida M, Horita Y, Asaoka M, et al . Evaluation of target area under the concentration-time curve of vancomycin in an initial dosing design: a retrospective cohort study. J. Antimicrob. Chemother. 2024; 79 :2518–2527. 26. He N, Su S, Ye Z, et al . Evidence-based guideline for therapeutic drug monitoring of vancomycin: 2020 update by the Division of Therapeutic Drug Monitoring, Chinese Pharmacological Society. Clin. Infect. Dis. 2020; 71 (Supplement 4):S363–S371. 27. Reuter SE, Stocker SL, Alffenaar JC, et al . Optimal practice for vancomycin therapeutic drug monitoring: position statement from the Anti-Infectives Committee of the International Association of Therapeutic Drug Monitoring and Clinical Toxicology. Ther. Drug Monit. 2022; 44 :121–132. 28. Yu Z, Liu J, Yu H, et al . Should the trough concentration of vancomycin be abandoned in therapeutic drug monitoring? A multicentre, retrospective study of critically ill patients without any form of dialysis. Int. J. Antimicrob. Agents . 2023; 61 :106812. 29. Suzuki A, Hamada Y, Ikeda H, et al . Comparison of trough concentration and area under the curve of vancomycin associated with the incidence of nephrotoxicity and predictors of a high trough level. J. Infect. Chemother. 2021; 27 :455–460. 30. Yang W, Zhang K, Chen Y, Fan Y, & Zhang J. Is it still beneficial to monitor the trough concentration of vancomycin? A quantitative meta-analysis of nephrotoxicity and efficacy. Antibiotics (Basel) . 2024; 13 :497. 31. Christensen A, Ryberg E, Nelson Z, et al . Better together? Reducing vancomycin use and acute kidney injury with a blended AUC and trough-based dosing guideline. Pharmacotherapy 2025;45:273–281. 32. Badillo S, Banfai B, Birzele F, et al . An introduction to machine learning. Clin. Pharmacol. Ther. 2020; 107 :871–885. 33. Corrigan BW. Artificial intelligence and machine learning: will clinical pharmacologists be needed in the next decade? The John Henry question. Clin. Pharmacol. Ther. 2020; 107 :697–699. 34. Miyai T, Imai S, Kashiwagi H, et al. A risk prediction flowchart of vancomycin-induced acute kidney injury to use when starting vancomycin administration: A multicenter retrospective study. Antibiotics (Basel) . 2020; 9 :920. 35. Imai S, Yamada T, Kasashi K, Niinuma Y, Kobayashi M, & Iseki K. Construction of a risk prediction model of vancomycin-associated nephrotoxicity to be used at the time of initial therapeutic drug monitoring: A data mining analysis using a decision tree model. J. Eval. Clin. Pract. 2019; 25 :163–170. 36. Mu F, Cui C, Tang M, et al . Analysis of a machine learning-based risk stratification scheme for acute kidney injury in vancomycin. Front. Pharmacol. 2022; 13 :1027230. 37. KDIGO. KDIGO clinical practice guideline for acute kidney injury. Kidney Int. Suppl. 2012; 2 :1–138. 38. Filippone EJ, Kraft WK, & Farber JL. The nephrotoxicity of vancomycin. Clin. Pharmacol. Ther. 2017; 102 :459–469. 39. Horita Y, Asaoka M, Iida M, et al . Development and evaluation of a novel software program, SAKURA-TDM, for area under the concentration-time curve-guided vancomycin dosing: A short communication. Ther. Drug Monit. 2023; 45 :245–250. 40. Yasuhara M, Iga T, Zenda H, et al. Population pharmacokinetics of vancomycin in Japanese adult patients. Ther Drug Monit . 1998;20:139–148. 41. Chawla NV, Bowyer KW, Hall LO, & Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J. Artif. Intell. Res. 2002; 16 :321–357. 42. Bakirarar B & Elhan AH. Class weighting technique to deal with imbalanced class problem in machine learning: methodological research. Turkiye Klinikleri J. Biostat. 2023; 15 :19–29. 43. Febriantono MA, Pramono SH, Rahmadwati R, & Naghdy G. Classification of multiclass imbalanced data using cost-sensitive decision tree C5.0. IAES Int. J. Artif. Intell . 2020;9:65–72. 44. Imai S, Yamada T, Kasashi K, Kobayashi M, & Iseki K. Usefulness of a decision tree model for the analysis of adverse drug reactions: evaluation of a risk prediction model of vancomycin-associated nephrotoxicity constructed using a data mining procedure. J. Eval. Clin. Pract. 2017; 23 :1240–1246. 45. Zhang A, Xing L, Zou J, & Wu JC. Shifting machine learning for healthcare from development to deployment and from models to data. Nat. Biomed. Eng. 2022; 6 :1330–1345. 46. Toni E, Ayatollahi H, Abbaszadeh R, & Fotuhi Siahpirani A. Machine learning techniques for predicting drug-related side effects: A scoping review. Pharmaceuticals (Basel) 2024; 17 :795. 47. Mizuno S, Noda T, Mogushi K, et al . Prediction of vancomycin-associated nephrotoxicity based on the area under the concentration-time curve of vancomycin: A machine learning analysis. Biol. Pharm. Bull. 2024; 47 :1946–1952. 48. Liang X, Fan Y, Yang M, et al. A prospective multicenter clinical observational study on vancomycin efficiency and safety with therapeutic drug monitoring. Clin. Infect. Dis. 2018; 67 (Supplement 2):S249–S255. Table 1. Patient characteristics during vancomycin dosing in the model development cohorts Total AKI group Non-AKI group P value ( n = 347) ( n = 18) (n = 329) Age (years) 72 (60–79) 71 (64–76) 72 (60–79) 0.787 ≥65 years, n (%) 237 (68.3%) 13 (72.2%) 224 (68.1%) 0.713 Sex, number of male (%) 235 (67.7%) 12 (66.7%) 223 (67.8%) 0.922 Height (cm) 162.0 (155.0–167.7) 162.3 (159.6–166.0) 162.0 (155.0–167.9) 0.662 Body weight (kg) 52.3 (44.2–61.2) 55.2 (42.8–61.9) 52.1 (44.2–61.1) 0.748 Body surface area (m 2 ) 1.54 (1.40–1.67) 1.54 (1.41–1.66) 1.54 (1.40–1.68) 0.739 BMI (kg/m 2 ) 20.1 (17.5–23.0) 20.6 (16.5–24.5) 20.1 (17.5–23.0) 0.926 BMI group 0.433 BMI < 18.5 kg/m 2 , n (%) 115 (33.1%) 7 (38.9%) 108 (32.8%) BMI 18.5–25 kg/m 2 , n (%) 180 (51.9%) 7 (38.9%) 173 (52.6%) BMI ≥25 kg/m 2 , n (%) 49 (14.1%) 4 (22.2%) 45 (13.7%) White blood cells (10 3 /μL) 8.8 (5.9–12.7) 9.7 (6.4–13.3) 8.8 (5.9–12.6) 0.495 Haemoglobin (g/dL) 10.3 (9.1–11.3) 9.6 (8.6–10.5) 10.3 (9.2–11.4) 0.142 Platelets (10 3 /μL) 216.0 (107.5–326.0) 136.5 (79.8–294.0) 219.0 (109.0–329.0) 0.213 Serum albumin (g/dL) 2.6 (2.2–3.0) 2.3 (2.1–2.8) 2.6 (2.2–3.0) 0.137 Serum creatinine (mg/dL) 0.64 (0.49–0.82) 0.52 (0.42–0.66) 0.64 (0.50–0.82) 0.019 Creatinine clearance (mL/min) 76.7 (55.9–107.2) 90.4 (66.8–115.4) 74.6 (55.2–105.8) 0.079 eGFR (mL/min/1.73 m 2 ) 86.7 (66.4–115.7) 108.9 (80.4–129.8) 85.2 (64.8–113.6) 0.026 C-reactive protein (mg/dL) 8.2 (4.0–16.2) 13.5 (5.3–21.8) 8.1 (3.7–15.9) 0.169 Blood urea nitrogen (mg/dL) 14.6 (10.9–21.4) 19.9 (11.4–26.1) 14.4 (10.8–20.9) 0.212 Blood urea nitrogen/serum creatinine 23.2 (16.7–32.8) 29.2 (19.2–53.3) 22.8 (16.7–32.4) 0.030 Dosing duration until the first TDM (h) 40.0 (36.0–60.0) 37.5 (36.0–60.0) 40.5 (36.0–60.0) 0.717 Type of infectious diseases Bacteremia/sepsis, n (%) 110 (31.7%) 6 (33.3%) 104 (31.6%) 0.878 Pneumonia, n (%) 111 (32.0%) 4 (22.2%) 107 (32.5%) 0.362 Skin and soft tissue infection, n (%) 26 (7.5%) 1 (5.6%) 25 (7.6%) 0.749 Abdominal infection, n (%) 19 (5.5%) 2 (11.1%) 17 (5.2%) 0.280 Meningitis, n (%) 16 (4.6%) 2 (11.1%) 14 (4.3%) 0.177 Urinary tract infection, n (%) 16 (4.6%) 1 (5.6%) 15 (4.6%) 0.844 Comorbidities ICU stay at the first sampling point, n (%) 21 (6.1%) 2 (11.1%) 19 (5.8%) 0.355 Type 2 diabetes mellitus, n (%) 96 (27.7%) 5 (27.8%) 91 (27.7%) 0.991 Cancer, n (%) 188 (54.2%) 9 (50.0%) 179 (54.4%) 0.715 Heart failure, n (%) 103 (29.7%) 5 (27.8%) 98 (29.8%) 0.856 Pharmacokinetic parameters Population parameters AUC 0–24 h (mg·h/L) 375.9 (297.8–447.0) 378.8 (298.9–465.4) 375.9 (299.0–445.9) 0.572 AUC 24–48 h (mg·h/L) 425.5 (346.3–469.2) 416.5 (367.3–477.4) 425.5 (343.3–469.2) 0.562 AUC 0–48 h (mg·h/L) 808.2 (644.6–908.7) 827.3 (651.5–941.2) 808.2 (647.8–903.9) 0.477 Predicted trough level (μg/mL) 10.6 (8.8–12.1) 11.2 (9.5–12.5) 10.6 (8.8–12.0) 0.357 Individual parameters a AUC 0–24 h (mg·h/L) 377.7 (298.7–462.8) 506.5 (367.2–550.5) 376.7 (295.5–455.8) 0.005 AUC 24–48 h (mg·h/L) 424.3 (346.5–501.9) 627.1 (475.9–746.3) 421.1 (341.1–490.2) <0.001 AUC 0–48 h (mg·h/L) 819.3 (651.6–956.3) 1157.4 (870.0–1276.9) 806.6 (643.7–943.1) <0.001 Measured trough level (μg/mL) 11.0 (8.0–14.0) 18.0 (11.8–25.8) 10.0 (8.0–14.0) <0.001 a The individual parameters were estimated using measured serum concentrations at the first sampling points. AKI, acute kidney injury; BMI, body mass index; eGFR, estimated glomerular filtration rate; TDM, therapeutic drug monitoring; ICU, intensive care unit. Table 2. Comparative performance of 11 machine learning models based on nested cross-validation Full model CART 0.60 (0.15) 0.24 (0.25) C5.0 0.80 (0.15) 0.31 (0.26) CTree 0.66 (0.16) 0.26 (0.24) NB 0.72 (0.18) 0.41 (0.34) NN 0.80 (0.10) 0.27 (0.28) SVM-LN 0.57 (0.04) 0.04 (0.00) SVM-RBF 0.57 (0.02) 0.06 (0.01) RF 0.89 (0.09) 0.47 (0.27) XGBoost 0.76 (0.16) 0.31 (0.11) LightGBM 0.86 (0.11) 0.44 (0.25) CatBoost 0.87 (0.08) 0.47 (0.19) LASSO-based model CART 0.78 (0.14) 0.35 (0.34) C5.0 0.82 (0.09) 0.38 (0.31) CTree 0.78 (0.14) 0.34 (0.21) NB 0.88 (0.11) 0.50 (0.30) NN 0.69 (0.12) 0.30 (0.14) SVM-LN 0.62 (0.05) 0.06 (0.04) SVM-RBF 0.58 (0.04) 0.10 (0.07) RF 0.87 (0.11) 0.53 (0.17) XGBoost 0.86 (0.10) 0.48 (0.30) LightGBM 0.69 (0.12) 0.30 (0.14) CatBoost 0.84 (0.13) 0.38 (0.21) Models were trained using inner three-fold and evaluated on five outer folds. Reported values indicate the mean and standard deviation of AUROC and AUPRC across outer folds. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision–recall curve. Table 3. Model performance on training and validation datasets following nested cross-validation Full model Training data AUROC 0.600 0.780 0.660 0.710 0.790 0.570 0.570 0.850 0.730 0.820 0.830 AUPRC 0.130 0.220 0.240 0.280 0.210 0.040 0.060 0.330 0.260 0.300 0.360 Accuracy (95% CI) 88.2 (84.2–91.4) 92.9 (89.6–95.4) 88.5 (84.6–91.7) 89.6 (85.9–92.7) 90.2 (86.6–93.2) 79.2 (77.9–80.4) 66.5 (65.1–68.0) 95.3 (92.4–97.3) 94.1 (91.0–96.3) 95.0 (92.1–97.0) 95.6 (92.8–97.5) F1 score 20.0 25.0 29.1 36.4 26.7 1.9 9.2 33.3 41.2 26.1 40.0 Validation data AUROC 0.450 0.860 0.740 0.940 0.930 0.890 0.750 0.810 0.790 0.790 0.890 AUPRC 0.060 0.620 0.380 0.690 0.630 0.610 0.320 0.640 0.600 0.650 0.660 Accuracy (95% CI) 84.2 (75.3–90.9) 91.6 (84.1–96.3) 92.6 (85.4–97.0) 95.8 (89.6–98.8) 92.6 (85.4–97.0) 75.8 (65.9–84.0) 92.6 (85.4–97.0) 94.7 (88.1–98.3) 95.8 (89.6–98.8) 92.6 (85.4–97.0) 95.8 (89.6–98.8) F1 score NA 50 NA 60 53.3 34.3 NA 61.5 66.7 NA 66.7 LASSO-based model Training data AUROC 0.790 0.810 0.800 0.830 0.700 0.630 0.580 0.830 0.830 0.700 0.720 AUPRC 0.210 0.370 0.320 0.370 0.210 0.040 0.060 0.390 0.470 0.210 0.280 Accuracy (95% CI) 89.3 (85.6–92.4) 94.2 (91.2–96.4) 93.4 (90.2–95.8) 78.4 (73.7–82.6) 91.6 (88.2–94.3) 70.8 (69.4–72.2) 80.6 (79.4–81.8) 95.4 (92.6–97.3) 95.4 (92.6–97.3) 91.6 (88.2–94.3) 94.2 (91.2–96.4) F1 score 39.3 44.4 41 27.2 29.3 2.6 9.8 42.9 52.9 29.3 41.2 Validation data AUROC 0.770 0.620 0.720 0.890 0.590 0.890 0.850 0.750 0.710 0.750 0.730 AUPRC 0.360 0.410 0.500 0.670 0.330 0.520 0.550 0.610 0.610 0.540 0.500 Accuracy (95% CI) 92.6 (85.4–97.0) 90.5 (82.8–95.6) 95.8 (89.6–98.8) 94.7 (88.1–98.3) 90.5 (82.8–95.6) 7.4 (3.0–14.6) 85.3 (76.5–91.7) 96.8 (91.0–99.3) 95.8 (89.6–98.8) 91.6 (84.1–96.3) 84.2 (75.3–90.9) F1 score 53.3 47.1 66.7 44.4 30.8 13.7 36.4 72.7 66.7 50 34.8 Values represent the mean of performance metrics. Accuracy is shown with 95% confidence intervals. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision–recall curve. Full model LASSO-based model Original Excluding day 2 AUC Excluding trough Original Excluding day 2 AUC Excluding trough Training data AUROC 0.85 0.77 0.82 0.83 0.83 0.84 AUPRC 0.33 0.22 0.3 0.39 0.29 0.34 Accuracy (95% CI) 95.3 (92.4–97.3) 95.0 (92.1–97.0) 95.0 (92.1–97.0) 95.4 (92.6–97.3) 94.5 (91.6–96.7) 94.8 (91.9–96.9) F1 score 33.3 19 26.1 42.9 17.4 35.7 Validation data AUROC 0.81 0.81 0.8 0.75 0.76 0.71 AUPRC 0.64 0.62 0.65 0.61 0.47 0.54 Accuracy (95% CI) 94.7 (88.1–98.3) 92.6 (85.4–97.0) 94.7 (88.1–98.3) 96.8 (91.0–99.3) 93.7 (86.8–97.6) 92.6 (85.4–97.0) F1 score 61.5 53.3 61.5 72.7 57.1 53.3 Values represent the mean of performance metrics. Accuracy is shown with 95% confidence intervals. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision–recall curve. Figure Legends Figure 1. Flowchart of patient inclusion and exclusion for the study cohort. AKI, acute kidney injury; TDM, therapeutic drug monitoring; EMC, East Medical Center; WMC, West Medical Center. Figure 2. SHAP summary plot for the random forest model using 23 variables. The SHAP summary plot displays the contribution of each feature to the model’s prediction of acute kidney injury (AKI). Features are ranked from top to bottom in order of decreasing importance, based on the mean absolute SHAP value. Each dot represents one patient, where the horizontal axis indicates the SHAP value. A positive SHAP value reflects an increased probability of AKI, whereas a negative value suggests a decreased risk. The colour scale represents the actual value of each feature: yellow indicates a higher feature value, and purple indicates a lower one. P/T, piperacillin/tazobactam; BMI, body mass index; CRP, C-reactive protein; BW, body weight; CrCl, creatinine clearance; BUN, blood urea nitrogen; WBC, white blood cells; Alb, serum albumin; BSA, body surface area; Scr, serum creatinine; ICU, intensive care unit. Figure 3. SHAP dependence plots for the six most influential features in the 23-variable RF model. Each plot illustrates the relationship between a feature and its SHAP value, showing how changes in that variable influence the model’s predicted probability of acute kidney injury (AKI). The six features with the highest overall importance are displayed as follows: (A) day 2 AUC, (B) 2-day AUC, (C) trough levels, (D) P/T co-administration, (E) BUN/Scr ratio, and (F) day 1 AUC. P/T, piperacillin/tazobactam; BUN/Scr, blood urea nitrogen/serum creatinine. List of Supplementary Material Table S1. List of 23 variables assessed using the logistic LASSO regression analysis Table S2. Patient characteristics during vancomycin dosing in model validation cohorts Table S3. Additional patient characteristics during vancomycin dosing Table S4. Optimised hyperparameters of the 11 machine learning (ML) algorithms Table S5. Feature importance of 23 variables across machine learning models Table S6. Feature importance of the five LASSO-selected variables across machine learning models Figure S1. Logistic least absolute shrinkage and selection operator (LASSO) regression analysis for variable selection. Figure S2. SHAP summary plot for the random forest model using five predictors. Figure S3. SHAP dependence plots for the five influential features in the LASSO-based RF model. Figure S4. Variable importance and variable interaction (VIVI) networks based on conditional SHAP interaction values. Information & Authors Information Version history V1 Version 1 19 January 2026 Copyright This work is licensed under a Non Exclusive No Reuse License. Authors Affiliations Moeko Iida 0009-0001-5814-6825 Nagoya Shiritsu Daigaku View all articles by this author Yasuhiro Horita 0000-0003-1191-8792 Nagoya Shiritsu Daigaku View all articles by this author Masato NODA Nagoya City University East Medical Center View all articles by this author Yoshinori HISADA Nagoya Shiritsu Daigaku Byoin View all articles by this author Masaya NAGAMIZU Nagoya Shiritsu Daigaku Byoin View all articles by this author Kaori TSUZUKI Nagoya Shiritsu Daigaku Byoin View all articles by this author Sakurako MURAMATSU Aichi Gakuin Daigaku View all articles by this author Kazuki OHASHI 0000-0002-2606-6468 Nagoya Shiritsu Daigaku View all articles by this author Yuki NOMURA Nagoya Shiritsu Daigaku View all articles by this author Minami ASAOKA Nagoya Shiritsu Daigaku View all articles by this author Tomoyo MIZUNO Nagoya Shiritsu Daigaku Byoin View all articles by this author Hiroki ASAKURA Nagoya Shiritsu Daigaku Byoin View all articles by this author Tomoaki HAYAKAWA Nagoya Shiritsu Daigaku View all articles by this author Chiharu WACHINO Nagoya Shiritsu Daigaku View all articles by this author Masaharu KUDO Nagoya Shiritsu Midori Shimin Byoin View all articles by this author Yoshihisa MIMURA Nagoya Shiritsu Daigaku View all articles by this author Taketo MIYAMOTO Nagoya Shiritsu Daigaku View all articles by this author Masami KAWAHARA Aichi Gakuin Daigaku View all articles by this author Nobuyuki MORISHITA Nagoya Shiritsu Daigaku Byoin View all articles by this author Masahiro Kondo Nagoya Shiritsu Daigaku View all articles by this author Yuji HOTTA Nagoya Shiritsu Daigaku View all articles by this author Atsushi NAKAMURA Nagoya Shiritsu Daigaku Byoin View all articles by this author YOKO FURUKAWA-HIBI 0000-0002-2126-5652 [email protected] Nagoya Shiritsu Daigaku View all articles by this author Metrics & Citations Metrics Article Usage 143 views 98 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Moeko Iida, Yasuhiro Horita, Masato NODA, et al. Comparative evaluation of 11 machine learning-based risk prediction models for vancomycin-associated acute kidney injury during initial dosing. Authorea . 19 January 2026. DOI: https://doi.org/10.22541/au.176878282.28318775/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176878282.28318775/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fe0e1a87e5fdf94',t:'MTc3OTE3MDQyOA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();