This is a retrospective observational study conducted at the First Affiliated Hospital of Shandong First Medical University. The Institutional Review Board committees at the First Affiliated Hospital of Shandong First Medical University approved the study (No. YXLLKY-2022-024). The requirement for informed consent was granted a waiver due to the retrospective nature and minimal risk of this study.

We enrolled patients who received etimicin or amikacin treatment at the center between 1 January 2018 and 31 December 2020. Patients were excluded if they met the following criteria: (1) aged <18 years; (2) hospital stay <48 h; (3) AKI was diagnosed on admission; end-stage kidney disease, or dialysis; (4) Serum creatinine (SCr) < 40 μmol/L during hospitalization, which are not considered clinically plausible and may distort the analysis; (5) less than two SCr test results during hospitalization; (6) incomplete medical records. Initial administration date of the first aminoglycoside administered, chosen as the exposure date, was determined from electronic medical record system. Records of each hospitalization were regarded as an independent case for patients admitted to hospital more than once during the study period. The study cohort selection process is shown in Figure 1.

The data for this study was collected from Shandong Provincial Qianfoshan Hospital Healthcare Big Data Platform. The platform integrates multi-source data from hospital information system, electronic medical records, laboratory information management system, picture archiving and communication system, nursing information system. The encrypted personal identification number was used as a unique identifier to interlink each person’s data information in the above-mentioned database. Based on a literature review and expert opinions, we identified fifty-eight candidate predictor variables (Supplementary Material).

For each patient, only clinical and laboratory variables measured prior to the onset of AKI (or the last SCr test in non-AKI patients) were included as candidate predictors. For AKI cases, we used the most recent measurement prior to the first creatinine elevation that met the KDIGO criteria. For non-AKI patients, we selected values closest to the time of discharge or the last aminoglycoside dose. This ensured that all predictor variables temporally preceded the outcome.

In electronic medical record systems, missing data is a common occurrence. Less than 10% of the missing values were found in all variables. For missing variables, we did not perform imputation and opted for complete-case analysis. Although tree-based algorithms such as gradient boosting can internally handle missing values, we excluded records with missing data to ensure consistency in preprocessing across all model types and simplify interpretation. Furthermore, none of the dichotomous variables used in our study were missing.

The diagnosis of AKI was based on SCr changes in accordance with the 2012 Kidney Disease: Improving Global Outcomes (KDIGO) Clinical Practice Guideline for AKI (Khwaja, 2012). AKI would be diagnosed if met one of the following criteria: (1) within 48 h, the absolute value of SCr increased ≥0.3 mg/dL (26.5 μmol/L); (2) known or speculated increase of SCr within 7days ≥1.5 times of baseline value; (3) urine volume ≤0.5 mL/kg/h for more than 6 h. Due to the influence of multiple factors on urine output, it is difficult to record in retrospective studies. In the present study, we defined AKI only according to the change in SCr. The SCr baseline was defined as the last laboratory measurement 7 days prior to receiving amikacin or etimicin treatment.

Given the retrospective nature of this study, it was not possible to definitively establish a causal relationship between aminoglycoside exposure and AKI. The temporal sequence-based approach used to evaluate the potential linkage between AKI and aminoglycoside administration was defined as follows: For patients who developed AKI, drug exposure was considered to be within 72 h prior to the first observed SCr elevation meeting the KDIGO AKI criteria. For non-AKI patients, we selected variables measured prior to the last serum creatinine test performed during the hospital stay. In cases where the last SCr test was unavailable, we used the most recent measurements before the last aminoglycoside dose or discharge (whichever occurred earlier). This approach ensured that all predictor variables were measured prior to the onset of AKI (or last exposure to aminoglycosides in non-AKI patients). Patients with AKI due to other identifiable etiologies (e.g., sepsis, acute obstructive uropathy, contrast-induced nephropathy, reperfusion injury) were excluded based on explicit documentation in clinical course records. All AKI events were temporally linked to aminoglycoside exposure, and medication administration records were reviewed to confirm that drug initiation preceded AKI onset.

R software (version 3.6.3) was used to statistically analyze the data obtained in the study. The significance threshold for all statistical tests was 0.05, and two-sided tests were conducted for all statistical tests. The Kolmogorov-Smirnov test was initially employed to assess the normal distribution of continuous variables. Variables conforming to a normal distribution were presented as mean ± standard deviation, and group comparisons were conducted using the t-test. For variables not following a normal distribution, they were expressed as median (interquartile range, IQR), and the Wilcoxon rank sum test was employed for intergroup comparisons. Categorical variables, represented as counts (n) or percentages (n%), underwent group comparisons using the chi-square test.

The patients included in the study were randomly divided into a training set and a test set in a ratio of 8:2. The training set data were only used for training and parameter tuning of the AKI risk prediction model, and the test set data were used to evaluate the model performance obtained from the training set.

In order to avoid overfitting of the model caused by too many variables, data dimension reduction and feature screening are carried out step by step. Firstly, univariate analysis was used to initially screen out information on variables that were significantly correlated (P < 0.05) with AKI. Then, the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm further performed data dimensionality reduction, and the optimal penalty coefficient (λ) was determined by validation with the ten-fold crossover method to screen out the characteristic variables with predictive value. Based on the screened variables, we used five different machine learning methods to construct prediction models, including logistic regression (LR) (Jiang et al., 2020), random forest (RF) (Breiman, 2001), gradient boosting machine (GBM) (Friedman, 2001), eXtreme gradient boosting (XGBoost) (Ester et al., 2022), and light gradient boosting machine (LightGBM) (Ke et al., 2017). We also ranked the variable importance by incorporating the feature variables into the RF model. All statistical analyses and model development were performed using R software (version 3.6.3). The following R packages were utilized in this study: glmnet (version 4.1-1) for LASSO regression, randomForest (version 4.6-14), xgboost (version 1.0.0.2), gbm (version 2.1.5), and lightgbm (version 3.2.1 for machine learning model construction and evaluation, pROC for ROC curve plotting and AUC calculation, and caret for model training workflow and parameter tuning. For each machine learning model, hyperparameter tuning was performed using a grid search strategy combined with five-fold cross-validation within the training set. For the XGBoost model, we optimized max_depth, eta (learning rate), gamma, subsample, and colsample_bytree. For the Random Forest model, we tuned the number of trees (ntree) and the maximum tree depth. For the GBM model, we adjusted n.trees, interaction.depth, shrinkage, and n.minobsinnode. For the LightGBM model, hyperparameters including num_leaves, learning_rate, and min_data_in_leaf were tuned. The optimal combination of hyperparameters was selected based on the cross-validation folds in the training dataset. Model calibration was assessed by plotting calibration curves using the calibration_curve function from scikit-learn. Predicted probabilities were grouped into deciles, and the observed AKI incidence within each group was plotted against the mean predicted risk. Calibration was visually inspected for the best performed models.

The interactive online prediction tools were developed using the shiny (version 1.4.0.2) and shinydashboard (version 0.7.1) packages in R, and deployed through the ShinyApps.io platform.

We used the area under the subject operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN) and specificity (SPE) to assess models’ performance. The confusion matrix divides patients into four categories, including true positive (TP), true negative (TN), false positive (FP) and false negative (FN). The number of cases in which a positive sample was correctly predicted as positive is represented by TP. Similarly, TN denotes the count of cases where a negative sample was accurately predicted as negative. The FP stands for the number of cases where a negative sample was erroneously predicted as positive, while FN indicates the instances where a positive sample was mistakenly predicted as negative. Accuracy measures the overall correctness of the model’s predictions. Sensitivity assesses the accuracy of the model in predicting positive samples. The optimal cutoff threshold was determined based on the maximum Youden index, and calculated the corresponding sensitivity and specificity (Fluss et al., 2005). Specificity assesses the accuracy of the model in predicting negative samples. The Equations 1–4 for these metrics are as follows. Youden  index = S P E + S E N − 1 (1) ACC = T P + T N T P + T N + F P + F N (2) SEN = T P T P + F N (3) S P E = T N T N + F P (4)

To address the class imbalance, we employed algorithms tolerant to skewed distributions and used discrimination-focused metrics (AUC, sensitivity, Youden index) rather than accuracy alone.

Among 307 inpatients treated with amikacin, 47 (15.3%) were diagnosed with AKI based on the KDIGO criteria. The baseline characteristics of patients with and without AKI after receiving amikacin are summarized in Table 1. The age, gender, height, smoking history and the length of stay were found to have no significant difference between patients with or without AKI. While, patients in the AKI group had significantly higher hospital costs than those in the non-AKI group (91,289.6 yuan vs 100,058.8 yuan, P = 0.001).

Based on univariate logistic regression analyses, 15 features were reduced to four potential predictors (Supplementary Figure S1). These four features included Cys C, total bilirubin, tetracycline, and acidosis (Figure 2). In addition, the feature importance plots were created by the RF algorithm to rank the importance levels. The results showed that Cys C had the greatest impact on the prediction results of amikacin-related AKI (Figure 2).

The predictors associated with AKI screened by LASSO regression were incorporated into the machine learning model and the receiver operating characteristic curve (ROC) was plotted (Figure 3).

As shown in Table 2, the XGBoost model achieved the best ACC value and has a better balance of sensitivity and specificity. Calibration plot for the model is shown in Supplementary Figure S2. The plot displays the predicted probabilities on the x-axis and the actual observed AKI incidence on the y-axis. The calibration curve indicates good alignment between predicted probabilities and observed outcomes, suggesting that the model is well-calibrated for predicting AKI occurrence in the amikacin cohort. Therefore, we developed the web-based calculator using the XGBoost model for risk prediction of amikacin-associated AKI (Supplementary Figure S3). This online risk calculator for amikacin-associated AKI was made freely available (https://akigbm.shinyapps.io/Amikacin/). Users can enter relevant data before and predict the risk of amikacin-associated AKI on the website.

Among 6,901 patients treated with etimicin, 319 (4.6%) developed AKI. The patient characteristics for those treated with etimicin are presented in Table 3. Our analysis revealed that patients in the AKI group were older and also incurred higher costs and longer hospital stays compared to those in the non-AKI group. However, there were no significant differences in gender or smoking history between patients in the AKI and non-AKI groups.

Based on univariate logistic regression analysis, the 47 characteristics were reduced to 13 potential predictors (Supplementary Figure S4). These 13 features included shock, acidosis, chronic obstructive pulmonary disease, chronic kidney disease, contrast examination, admission to ICU, mechanical ventilation, combined use of diuretics or NSAIDs, RBC, SCr, β2-MG, and Cys C (Figure 4). The importance of the study variables was ranked by the RF algorithm. The results showed that Cys C, β2-MG, and SCr were the top three variables affecting the predicted outcome of etimicin-associated AKI (Figure 4).

The predictors associated with AKI screened by LASSO regression were incorporated into the machine learning model and the ROC was plotted (Figure 5).

As shown in Table 4, the GBM model preformed the best on test set. Calibration plot for the model is shown in Supplementary Figure S6. The plot illustrates the relationship between predicted probabilities and observed outcomes, with the x-axis representing predicted probabilities and the y-axis showing the actual observed AKI incidence. The calibration curve demonstrates reasonable agreement between predicted and observed values, indicating that the model is well-calibrated for predicting AKI occurrence in the etimicin cohort. We made this model available online as a web-based calculator https://akigbm.shinyapps.io/EAPP/) to predict the risk of developing etimicin-associated AKI (Supplementary Figure S7). The user enters the relevant variables on the website and clicks the Predict button, and the results are displayed in the form of high and low risks. When the results suggest that the patient has a high risk of developing AKI after receiving etimicin, it suggests that clinicians need to strengthen the monitoring of the patient’s renal function and adjust the treatment regimen in a timely manner.

In our study, the incidence of AKI in inpatients treated with amikacin was 15.3% (47/307), while the incidence of AKI in inpatients treated with etimicin was 4.6% (319/6,901). The incidence of AKI in hospitalized patients taking aminoglycosides had been shown in previous studies to range from 4.3% to 26.7%, depending on the definition and study population (Selby et al., 2009; Ong et al., 2016; Ergun et al., 2022). Clinical studies have shown that etimicin has lower minimum inhibitory concentration and minimum bactericidal concentration than amikacin. Meanwhile, etimicin exhibited longer-lasting bactericidal activity based on the time kill-curve (Chaudhary et al., 2012). Yao et al. compared the nephrotoxicity of three different aminoglycosides at the same dose (Yao et al., 2020). It was shown that etimicin had lower nephrotoxicity compared to amikacin and gentamicin.

AKI patients incurred significantly higher costs and longer hospital stays than non-AKI patients. In addition, the probability of patients in AKI group to undergo continuous renal replacement therapy (CRRT) was higher, which increased the cost of hospitalization to a certain extent. A case-control study in the United States shows that the severityof acute kidney injury is significantly associated with the risk of death (Shusterman et al., 1987). Although the overall mortality rate for patients with AKI is 21% globally, the mortality rate for patients with severe AKI is as high as 42% (Mehta et al., 2015b). Oliveira et al. found that the mortality rate among ICU patients treated with aminoglycosides was significantly higher in the AKI group (44.5%) than in the non-AKI group (29.1%) (Oliveira et al., 2009). Due to the high prevalence and mortality of in-hospital AKI, early recognition and prevention are critical to patient outcomes.

Previous studies have identified several independent predictors associated with AKI, including shock, COPD, CHF, CKD, chronic liver disease, nephrotoxic drugs, SCr, β2-MG, ICU admission, and mechanical ventilation (Liu et al., 2019; Yue et al., 2022; Feng et al., 2023), which were confirmed in our study as well. In addition, we observed some new risk factors associated with AKI, such as acidosis, tetracyclines, red blood cell count, TBiL, and Cys C. Studies have shown that Cys C increases earlier than serum creatinine in the early stages of kidney injury and is not affected by age, sex, or race (Dharnidharka et al., 2002; Herget-Rosenthal et al., 2004). A Meta-analysis that included 3,336 patients noted that early serum Cys C could be used to predict AKI with higher predictive power than urinary Cys C (Zhang et al., 2011). The predictive ability of serum Cys C in assessing the occurrence of AKI has been demonstrated in various patients including those with traumatic brain injury, acute aortic coarctation, and those undergoing cardiac surgery (Wang et al., 2020; Wang et al., 2021; Wang et al., 2022). TBiL is an important indicator of the liver function of the body. It has been shown that TBiL concentration is closely related to aminoglycoside nephrotoxicity (Desai and Tsang, 1988). Bikrant et al. reported that when the bilirubin content was greater than 17.7 mg/dL, the risk of AKI in patients with chronic acute liver failure was 6.17 times higher than that without chronic acute liver failure (OR: 6.17, P = 0.011) (Lal et al., 2018). The red blood cells play an important role in maintaining normal life activities in the body as carriers of oxygen transport. In our study, patients in the AKI group had a lower number of red blood cells in their bodies prior to drug administration compared to patients in the non-AKI group. A study of liver transplant patients showed that lower red blood cell counts were associated with an increased risk of AKI (Zeng et al., 2023). This was consistent with the results obtained in this study. Many studies have found an association between red blood cell distribution width and AKI (Ramires et al., 2022; Zhu et al., 2022). As the main organ regulating systemic HCO3-concentration, the kidney plays a vital role in maintaining acid-base balance in the body. Magalhães et al. reported that acidosis promotes a decrease in glomerular filtration rate and tubular function, increases nuclear factor κB and heme oxygenase one levels, and aggravates the degree of kidney injury (Magalhaes et al., 2016). A study based on the Food and Drug Administration Adverse Event Reporting System revealed that the risk of AKI was 1.73 times higher in patients using tetracyclines than in those not using tetracyclines (Patek et al., 2020). In another study, the combined use of tetracycline and aminoglycosides triggered an inflammatory response by releasing large amounts of inflammatory factors (IL-6). Meanwhile, the combination of these two drugs also caused a significant decrease in glutathione levels and catalase activity in kidney tissue (Elgazzar et al., 2022). With patients who have multiple potential risk factors, we recommend that clinicians intensify their attention to patients in order to make timely adjustments to their treatment prescriptions.

Furthermore, our study shows that there was no significant difference in gender between the AKI and non-AKI groups. The KDIGO guideline (2012) indicates that women are more prone to developing AKI (Khwaja, 2012). However, there is still some controversy regarding the association between gender and AKI. Loutradis et al. reported that the incidence of AKI was significantly higher in male patients than in females (11.3% vs. 7.1%, P < 0.001). The association between men and AKI persisted after adjusting for confounders such as age, smoking history, and alcohol consumption (P = 0.001) (Loutradis et al., 2021). Neugarten et al. conducted a systematic review and meta-analysis of studies on aminoglycoside-related nephrotoxicity published between 1978 and 2015 (Neugarten and Golestaneh, 2016). Their study showed that male patients have a significantly higher risk of developing AKI than female patients. The differences may due to the study population.

In recent years, machine learning techniques have become increasingly prevalent in addressing medical and clinical challenges. Several studies have developed risk prediction models for drug-related AKIs, such as vancomycin and diuretics (Pan et al., 2020; Zhang et al., 2022). Pan et al. retrospectively analyzed elderly patients treated with vancomycin between January 2016 and June 2018 at a hospital in China. Univariate analysis and multivariable logistic regression analysis revealed that vancomycin trough concentration ≥20 mg/L, surgery, the Charlson Comorbidities Index ≥4 points, concomitant use of cardiotonic drug, plasma volume expander, and piperacillin/tazobactam were risk factors for vancomycin-associated AKI in elderly patients (Pan et al., 2020). The machine learning models constructed based on vancomycin-associated AKI risk factors had AUCs of 0.828 (95% CI 0.758-0.898) and 0.736 (95% CI 0.581-0.890) in the test and validation sets, respectively. In a prior study, we constructed a risk prediction model for AKI in hospitalized patients treated with diuretics (Zhang et al., 2022). There was an AUC of 0.74 (95% CI, 0.72-0.76) for the torasemide-associated AKI prediction model, and an AUC of 0.79 (95% CI, 0.77-0.80) for the furosemide-associated AKI prediction model. In the present study, the AUC values for the AKI-prediction models associated with amikacin and etimicin were 0.841 (95% CI, 0.692-0.990) and 0.900 (95% CI, 0.847-0.953), respectively. Compared to previous studies, our model has better recognition ability. Due to the varying clinical profiles associated with different drugs, aggregating them for an overall analysis may introduce bias and potentially reduce accuracy. Therefore, it is important to select the appropriate prediction model tailored to patients receiving specific medications. These prediction models can serve as valuable tools for identifying patients at high risk of AKI. While, we acknowledge the necessity of clinical judgment in determining the optimal course of action for such patients. Aligned with the principles of personalized medicine and prudent clinical practice, clinicians may consider exploring alternative treatment options or adjusting doses for patients identified as high risk by our calculator. This approach aims to maximize clinical benefits while minimizing harm to patients, thus optimizing therapeutic outcomes.

Compared with previously published machine learning models predicting AKI risk associated with other antibiotics such as vancomycin, β-lactams, or contrast agents, our models for aminoglycoside-related AKI demonstrate several unique clinical insights. First, markers such as cystatin C and total bilirubin (TBIL), which are not commonly highlighted in AKI risk prediction models for other nephrotoxic agents, were among the top-ranked predictors in our study. This may reflect the specific pharmacokinetic characteristics and renal handling mechanisms of aminoglycosides. Second, we observed that relatively moderate baseline kidney dysfunction, rather than overt chronic kidney disease, was strongly associated with D-AKI in aminoglycoside users, suggesting a unique susceptibility window. These findings underline the importance of early biomarker monitoring and individualized dose adjustment during aminoglycoside therapy. To our knowledge, this is one of the few large-scale ML-based studies focused specifically on aminoglycoside-induced AKI risk stratification, and our results may complement existing models for other antibiotic classes in guiding safer antimicrobial therapy.

Our study has several strengths. First, we identified risk factors for AKI associated with etimicin and amikacin, respectively. Second, we ranked the importance of the risk factors obtained from the screening, which helps to understand the impact of single features on the prediction model. Third, we provide a reliable prediction model for assessing the risk of AKI in patients treated with amikacin or etimicin. This will help clinical doctors identify high-risk patients with D-AKI and promptly develop the best treatment strategy.

There are also limitations in our study. First, the data was retrospectively collected from the electronic medical record systems. All patients included in the study were sourced solely one medical center, which may introduce a degree of bias. The models lacked validation against other datasets. Therefore, the applicability of our prediction models needs to be further validated in larger datasets in the future. Second, the variables included in our study included laboratory indicators. If patient test data were missing, model accuracy may be reduced. A number of patients were excluded from the final cohort, primarily due to insufficient serum creatinine measurements that precluded accurate AKI classification per KDIGO criteria, or missing values in key predictor variables. It is possible that patients with sparse creatinine or cystatin C testing were healthier and less closely monitored, thus more likely to belong to the non-AKI group. This pattern suggests a potential missing-not-at-random (MNAR) mechanism, which may introduce selection bias and limit the generalizability of our findings. Future studies with prospective designs and standardized data collection will be needed to address this limitation. Third, the calibration plots revealed slight underestimation of AKI risk at higher predicted probabilities, particularly in the amikacin model. This may be due to a limited number of high-risk cases in our dataset. In future studies, calibration can be further improved using oversampling strategies or post hoc recalibration techniques (e.g., Platt scaling or isotonic regression) to refine risk probability estimates before clinical deployment. Finally, overfitting is a concern in several of our models, despite our efforts to mitigate it through parameter tuning. Despite the relatively small sample size in the amikacin group (n = 307, with 47 AKI cases), we developed and internally validated the model using cross-validation and independent test sets. While the number of positive cases in the validation subset was indeed limited, the model still demonstrated reasonable performance (AUC = 0.841). This model provides a preliminary, data-driven tool for identifying patients at increased risk of AKI when treated with amikacin, and should be interpreted in conjunction with clinical judgment. While this issue may not significantly impact the usability of the models or the validity of our conclusions, it warrants attention in future research. In the future, we will address overfitting by increasing sample size and conducting substantive external validation.