Machine learning studies of drug-induced nephrotoxicity: a scoping review

Abstract

Background:

Machine learning methods have emerged as a promising approach to prevent drug-induced nephrotoxicity.

Objective:

This review evaluates the quality and highlights recent advances of machine learning algorithms for predicting drug-induced nephrotoxicity.

Eligibility criteria:

Studies on machine learning models to predict drug-induced acute kidney injury, acute kidney disease, or both published between January 2014 and August 2024 were eligible.

Sources of evidence:

A comprehensive search was conducted by using PubMed, Embase, Web of Science, Cochrane Library, and Scopus.

Charting methods:

A standardized charting form was developed based on CHARMS, TRIPOD+AI, and PROBAST tools to assess the quality and risk of bias across studies.

Results:

From the initial 5,179 articles searched, 24 studies were included in this review. All studies achieved good area under the receiver operating characteristic curves (AUROCs) above 0.75, with boosting machines being the most frequently outperforming algorithms (n = 7, 29.17%), and neural networks showed the highest median AUROC of 0.90 (0.86–0.92). Two-thirds of studies (n = 16; 66.67%) predicted acute kidney injury, whereas only 5 (20.83%) focused on acute kidney disease. Estimated glomerular filtration rate, blood urea nitrogen, serum creatinine, hemoglobin, and albumin emerged as the most utilized features by 10 (41.67%), 9 (37.5%), 9 (37.5%), 8 (33.33%), and 8 (33.33%) studies, respectively. Diabetes, heart failure, diuretics, and non-steroidal anti-inflammatory drugs were frequently selected features by 7 (29.17%), 5 (20.83%), 5 (20.83%), and 4 (16.67%) studies, respectively. The 2025 PROBAST+AI risk-of-bias assessment indicated that 7 (29.17%) studies had a low risk of bias. A high risk of bias was observed in 20 (83.33%), 18 (75%), and 17 (70.83%) studies due to insufficient performance evaluation, small sample sizes, and lack of external validation.

Conclusion:

Recent machine learning studies have demonstrated great performance using clinically obtainable features. Incorporating acute kidney injury and disease, methodological enhancement, and guideline adherence can facilitate clinical applicability in preventing drug-induced nephrotoxicity.

Plain language summary

Machine learning studies of drug-induced nephrotoxicity: a scoping review

Why was the study conducted? Drug-induced nephrotoxicity is a significant but unpredictable clinical issue. There are several machine learning (ML) studies published to identify patients at high-risk for drug-induced nephrotoxicity to prevent adverse drug reactions. This research was conducted to provide clinicians with an understanding of the current progress in ML-based studies aimed at predicting drug-induced nephrotoxicity, thereby enhancing clinical practice.

What did the research team do? This review identified research trends and performance and provided recommendations for implementing ML-based prediction models for drug-induced nephrotoxicity. The research team conducted a systematic search across five medical databases for relevant studies published between 2014 and 2024. The ML models used, variables, predicted outcomes, performance metrics, and the detailed methods were compared to assess the quality and achievement of the studies.

What did the research team find? The number of publications has significantly increased over the last two years. Twenty-four studies were included in this review, and all demonstrated good predictive performance (AUROC > 0.75). Neural Networks and Boosting Machine algorithms have demonstrated superior predictive ability and have been widely adopted in recent studies. The models predominantly used eGFR, serum creatinine, BUN, history of diabetes, and history of heart failure. Large sample sizes, external validation, and model evaluation should be used to improve methodological rigor in future studies.

What do the findings imply? This review revealed considerable potential for ML-based preventive tools to help clinicians identify patients at risk of drug-induced nephrotoxicity. Multiple models predicting drug-induced nephrotoxicity have identified overlapping variables, indicating the possibility of constructing an integrated predictive model. Integrating ML-based prediction into the clinical decision support systems provides clinicians with an actionable strategy to implement timely interventions to prevent drug-induced nephrotoxicity.

Keywords

AI drug drug-induced nephrotoxicity machine learning nephrotoxicity prediction risk of bias

Introduction

Utilizing machine learning to predict drug-induced nephrotoxicity is increasingly essential in modern healthcare. Common causal agents include antibiotics, contrast agents, and chemotherapeutic drugs.¹ Strategies such as temporary drug discontinuation,² hydration for cisplatin-induced toxicity,³ and administering sodium bicarbonate for contrast-related nephrotoxicity,⁴ cannot entirely prevent drug-induced nephrotoxicity. Advanced machine learning technologies can identify intricate patterns to predict drug-induced nephrotoxicity with higher accuracy than conventional clinical approaches.⁵ For patients at extremely high risk of drug-induced nephrotoxicity, clinicians can avoid the selection of nephrotoxic drugs at the onset of prescribing. For patients treated with nephrotoxic agents, continuous monitoring by machine learning techniques can predict any impending drug-induced nephrotoxicity, allowing clinicians to intervene before significant kidney damage occurs.⁶ The patient’s clinical outcomes, long-term mortality,⁷ and healthcare costs⁸ can thus be improved by machine learning-based predictions of drug-induced nephrotoxicity.

Understanding the pathophysiology and progression patterns of drug-induced nephrotoxicity is vital for developing and designing machine learning algorithms. Renal damage may present as rapid acute kidney injury (AKI) in the initial days after use of a nephrotoxic agent. Traditionally, clinicians rely on standardized diagnostic frameworks, such as the Kidney Disease: Improving Global Outcomes (KDIGO) or the Risk, Injury, Failure, Loss of kidney function, and End-stage kidney disease (RIFLE) criteria, to stage these conditions. Some patients then progress to acute kidney disease (AKD), a transitional phase toward chronic kidney disease (CKD),⁹ with or without the presence of AKI. Continuous and careful follow-up, even after AKI resolution, can prevent the occurrence of irreversible CKD. Detecting patients at risk of AKD provides an opportunity to utilize the critical window period to rescue kidney function.¹⁰ The machine learning-based application meets the need for proactive approaches before the onset of acute renal damage and moderate ongoing kidney injury, thereby making long-term renal care strategies possible.

Machine learning has rapidly emerged as a transformative force in nephrology, enabling early risk stratification to improve patient outcomes. The mainstream predictive applications of machine learning in nephrology are to identify and mitigate the risk of high-mortality or morbidity conditions, including AKI and CKD, across different patient populations or time periods, such as post-COVID-19.¹¹ Halder et al. developed an intelligent web-based application for CKD prediction utilizing several machine learning algorithms, which demonstrated robust predictive accuracy.¹² The machine learning prediction model can further leverage Electronic Health Records (EHRs) to proactively identify patients at elevated risk prior to clinical presentation.

Predicting drug-induced nephrotoxicity by machine learning algorithms holds great promise but requires careful optimization for clinical implementation. Limited sample sizes,¹³ a lack of external validation,¹⁴ overfitting, improper missing data management,¹⁵ and a restrained explanation may reduce the utility of such models. Employing quality assessment tools to guide the design and development of machine learning-based models for drug-induced nephrotoxicity facilitates clinical application. Existing tools such as CHARMS, PROBAST, APPRAISE-AI, and TRIPOD+AI are very extensive but general for all kinds of predictive models.^15–18 The unique aspects and current progress of predictive models for drug-induced nephrotoxicity require specific evaluation metrics to expedite their clinical implementation.

This review aims to address the notable progress in recent machine learning studies that specifically focused on drug-induced nephrotoxicity. Available machine learning reviews were conducted for AKI and sepsis-associated AKI,^19–23 and no literature has comprehensively evaluated the machine learning studies for drug-induced nephrotoxicity. This review also seeks to analyze innovative data-mining methods and algorithms currently available in the literature, identify the most frequently selected predictive features, and evaluate model performance, in order to provide constructive guidance to enhance the methodological rigor and clinical relevance of machine learning models for drug-induced nephrotoxicity.

Methods

Study design and search strategy

This review analyzed studies that used machine learning models to predict drug-induced nephrotoxicity, published in English, between January 1, 2014 and August 31, 2024. It was conducted in accordance with the PRISMA-ScR guidelines for scoping reviews²⁴ (Supplemental Table 1). This review considered a range of machine learning algorithms, including neural networks (NNs), boosting machines (BMs), random forests (RFs), decision trees (DTs), support vector machines (SVMs), naïve Bayes, and logistic regressions. Original studies of machine learning models to predict drug-induced AKI, AKD, or both were eligible. The nephrotoxicity definitions used in each study varied and are summarized in Supplemental Table 2. Exclusion criteria applied to review articles, conference papers, case reports, non-journal sources, animal or in vitro research, studies predicting multiple adverse drug reactions (ADRs) as a composite outcome, without separating nephrotoxicity and other ADRs, those not using creatinine or the estimated glomerular filtration rate (eGFR) to define drug-induced nephrotoxicity, and model updating studies. The predefined protocol for paper selection is outlined in Supplemental Figure 1.

A thorough search was performed across PubMed, Cochrane Library, Embase, Scopus, and Web of Science. To focus on recent advances in the clinical field, the IEEE Xplore was not used. The initial search strategy was adapted from Li et al. and further refined.²¹ Key terms included “artificial intelligence,” “machine learning,” and “deep learning,” as well as kidney injury-related terms such as “acute kidney injury,” “acute kidney disease,” and “nephrotoxicity.” Additional targeted keywords included “prediction,” “prognosis,” “diagnosis,” and “risk assessment.” The finalized search strategies for each database are detailed in Supplemental Table 3. All results were exported to EndNote, with duplicates removed both automatically and manually.

Two reviewers initially screened a subset of publications, refined the data analysis process to ensure consistency, and independently assessed the titles, abstracts, and full texts of all identified studies. The items collected for each study are listed in Supplemental Figure 2. Baseline conditions (e.g., age, comorbidities/medical history) were collected as features used in each study, as shown in Supplemental Table 4. Any disagreements during the study selection, data charting, and quality assessment were resolved through consensus with a third reviewer.

Analysis of study characteristics and features

Studies were first grouped by drug and year. A timeline plot was used to visualize the trend of the machine learning studies published during the study period. Data were presented using both absolute values and proportions to represent the core characteristics of the studies. Frequency of study types, validation methods, sample sizes, and machine learning algorithm were calculated to reveal the current development of machine learning studies for drug-induced nephrotoxicity. Key features in establishing machine learning models were grouped by drugs, and their frequencies of selection are illustrated with a bar plot.

Comparison of model performances

Model performance metrics, including the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score, were summarized to reveal the performance of the published machine learning models in predicting drug-induced nephrotoxicity. The performance of these metrics was compared by drug type and algorithm, using box plots to demonstrate the model’s predictive ability. The frequencies of each metric employed in the studies were calculated to reveal their utility. Supplemental Table 5 showed the definitions of the metrics. An independent t test or Mann–Whitney U test was used to perform a comparative analysis, depending on the data distribution (Table 2). Statistical significance was determined using a two-tailed test with p < 0.05. The statistical test and data visualization in this review were performed using Python v3.12.12 (Python Software Foundation, Beaverton, OR, USA).

Quality assessment of the machine learning studies

Two tools were employed to assess the quality of current machine learning research for drug-induced nephrotoxicity. The 2025 PROBAST+AI tool was first utilized, and results are summarized in a stacked bar plot.²⁵ Four subdomains—analysis, outcome, predictor, and participants—were assessed alongside the overall assessment. The study also developed a standardized charting tool by adapting important quality criteria supported by CHARMS, PROBAST, APPRAISE-AI, and TRIPOD+AI^15–18 (Supplemental Table 6). The author team discussed, selected, and rephrased 12 criteria to form a quality assessment tool to evaluate studies on drug-induced nephrotoxicity. Each criterion was rated using scores of +0, +1, or +2. The quality rate of each study was defined as the sum of scores divided by the maximum score of 48 points. The rate was, respectively, classified into low, moderate, and high quality as <40%, 40%–59%, and ⩾60%, as adapted from APPRAISE-AI.

Results

Characteristics of drug-induced nephrotoxicity in machine learning studies

In total, 3,710 unique citations were identified through database searches, and 2,695 studies not meeting specific inclusion criteria were excluded (Supplemental Figure 1). There were 982 reviews/meta-analyses/non-journal articles, six animal/in vitro studies, one study examining multiple ADRs, and two studies that did not define nephrotoxicity by serum creatinine or eGFR, which were excluded. The main study design and methods of the 24 studies are summarized in Supplemental Table 2.^26–49 Most studies utilized a short prediction window (⩽7 days), specifically within 7 days (n = 8, 33.33%), 72 h (n = 5, 20.83%), and 48 h (n = 3, 12.50%), while the remaining studies either employed longer timeframes or did not specify a prediction window (Supplemental Table 2). Figure 1 illustrates evolving trends in the included studies. The number of publications significantly increased in 2023 and 2024 (n = 14, 58.33%), marking a notable rise compared to earlier years (2017–2022). RFs led the way (n = 3, 12.5%) before 2022. BMs emerged as the most frequent model (n = 6, 25%) in 2023 and 2024.

Figure 1.

Timeline plot of drug-induced nephrotoxicity machine learning studies. A chronological overview of studies on drug-induced nephrotoxicity is illustrated in a timeline plot. Each study is organized by year and its clinical focus (type of nephrotoxicity). Colors represent the machine learning algorithm employed.

Most studies were conducted in East Asia, particularly in China (n = 10, 41.7%), Japan (n = 6, 25.0%), Taiwan (n = 3, 12.5%), and South Korea (n = 2, 8.3%). Three studies were conducted in North America/Europe, including in the United States (n = 1, 4.17%), Denmark (n = 1, 4.17%), and Turkey (n = 1, 4.17%; Table 1). All studies employed retrospective designs, with two studies also adding prospective validation. Twenty-two (91.67%) studies conducted internal validation with 7 (29.17%), 11 (45.83%), 3 (12.50%), and 1 (4.17%) using cross, both cross and split, split, and unspecified validation, respectively. External validation was performed in only 7 (29.17%) studies. Eleven (45.83%) studies included patient numbers ranging from 100 to 1,000, while only 4 (16.67%) studies exceeded 10,000 participants. The number of studies that focused on contrast-, platinum-, and vancomycin-induced nephrotoxicity was 10 (41.76%), 4 (16.67%), and 3 (12.5%), respectively. Two-thirds (n = 16, 66.67%) of the studies predicted AKI, with 5 (20.83%) focusing on AKD.

Table 1.

Summary of included studies.

Characteristics	Overall (N = 24), n (%)	Contrast (N = 10), n (%)	Platinum (N = 4), n (%)	Vancomycin (N = 3), n (%)	Other (N = 7), n (%)
Data source, n (%)
Singe center	16 (66.67)	6 (60.00)	3 (75.00)	3 (100.00)	4 (57.14)
Multi center	6 (25.00)	3 (30.00)	0 (0.00)	0 (0.00)	3 (42.86)
Database	3 (12.50)	2 (20.00)	1 (25.00)	0 (0.00)	0 (0.00)
Country of origin, n (%)
China	10 (41.67)	6 (60.00)	1 (25.00)	1 (33.33)	3 (42.86)
Japan	6 (25.00)	1 (10.00)	1 (25.00)	1 (33.33)	2 (28.57)
Taiwan	3 (12.50)	1 (10.00)	1 (25.00)	0 (0.00)	1 (14.29)
South Korea	2 (8.33)	1 (10.00)	0 (0.00)	1 (3.33)	0 (0.00)
United States	1 (4.17)	1 (10.00)	1 (25.00)	0 (0.00)	0 (0.00)
Denmark	1 (4.17)	0 (0.00)	0 (0.00)	0 (0.00)	0 (0.00)
Turkey	1 (4.17)	0 (0.00)	0 (0.00)	0 (0.00)	1 (14.29)
Time setting, n (%)
Retrospective	24 (100.00)	10 (100.00)	4 (100.00)	3 (100.00)	7 (100.00)
Prospective	2 (8.33)	1 (10.00)	1 (25.00)	0 (0.00)	0 (0.00)
Validation method, n (%)
Internal and external validation	5 (20.83)	3 (30.00)	1 (25.00)	0 (0.00)	1 (14.29)
External validation	2 (8.33)	1 (10.00)	1 (25.00)	0 (0.00)	0 (0.00)
Internal validation	17 (70.83)	6 (60.00)	2 (50.00)	3 (100.00)	6 (85.71)
Internal validation method, n (%)
Cross-validation	7 (29.17)	0 (0.00)	2 (50.00)	3 (100.00)	2 (28.57)
Cross- and split validation	11 (45.83)	6 (60.00)	1 (25.00)	0 (0.00)	4 (57.14)
Split validation	3 (12.50)	2 (20.00)	0 (0.00)	0 (0.00)	1 (14.29)
Unspecified validation	1 (4.17)	1 (10.00)	0 (0.00	0 (0.00	0 (0.00
No internal validation	2 (8.33)	1 (10.00)	1 (25.00)	0 (0.00)	0 (0.00)
External validation method, n (%)
External set validation	2 (8.33)	2 (20.00)	0 (0.00)	0 (0.00)	0 (0.00)
External set and prospective validation	1 (4.17)	1 (10.00)	0 (0.00)	0 (0.00)	0 (0.00)
Prospective validation	1 (4.17)	0 (0.00)	1 (25.00)	0 (0.00)	0 (0.00)
Temporal validation	3 (12.50)	1 (10.00)	1 (25.00)	0 (0.00)	1 (14.29)
No external validation	17 (70.83)	6 (60.00)	2 (50.00)	3 (100.00)	6 (85.71)
Drug type, n (%)
Contrast	10 (41.67)	10 (100.00)	0 (0.00)	0 (0.00)	0 (0.00)
Platinum	4 (16.67)	0 (0.00)	4 (100.00)	0 (0.00)	0 (0.00)
Vancomycin	3 (12.50)	0 (0.00)	0 (0.00)	3 (100.00)	0 (0.00)
Other	7 (29.17)	0 (0.00)	0 (0.00)	0 (0.00)	7 (100.00)
Training sample size, n (%)
100–1000	11 (45.83)	1 (10.00)	3 (75.00)	2 (66.67)	5 (71.43)
1001–10,000	9 (37.50)	6 (60.00)	1 (25.00)	1 (33.33)	1 (14.29)
>10,000	4 (16.67)	3 (30.00)	0 (0.00)	0 (0.00)	1 (14.29)
Nephrotoxicity type, n (%)
AKI	16 (66.67)	9 (90.00)	0 (0.00)	3 (100.00)	4 (57.14)
AKD	5 (20.83)	1 (10.00)	1 (25.00)	0 (0.00)	3 (42.86)
Not specified	3 (12.50)	0 (0.00)	3 (75.00)	0 (0.00)	0 (0.00)
Best algorithm, n (%)
Neural network	3 (12.50)	1 (10.00)	1 (25.00)	1 (33.33)	0 (0.00)
Boosting machine	7 (29.17)	2 (20.00)	0 (0.00)	1 (33.33)	4 (57.14)
Random forest	5 (20.83)	3 (30.00)	1 (25.00)	0 (0.00)	1 (14.29)
SVM	2 (8.33)	1 (10.00)	0 (0.00)	0 (0.00)	1 (14.29)
Logistic regression	4 (16.67)	2 (20.00)	1 (25.00)	1 (33.33)	0 (0.00)
Other	3 (12.50)	1 (10.00)	1 (25.00)	0 (0.00)	1 (14.29)
Reported metric, n (%)
AUROC	24 (100.00)	10 (100.00)	4 (100.00)	3 (100.00)	7 (100.00)
AUPRC	6 (25.00)	2 (20.00)	0 (0.00)	1 (33.33)	3 (42.86)
Accuracy	19 (79.17)	8 (80.00)	4 (100.00)	2 (66.67)	5 (71.43)
Sensitivity (recall)	20 (83.33)	9 (90.00)	4 (100.00)	0 (0.00)	7 (100.00)
Specificity	18 (75.00)	8 (80.00)	4 (100.00)	0 (0.00)	6 (85.71)
PPV (precision)	16 (66.67)	6 (60.00)	4 (100.00)	0 (0.00)	6 (85.71)
NPV	13 (54.17)	5 (50.00)	4 (100.00)	0 (0.00)	4 (57.14)
F1-score	16 (66.67)	6 (60.00)	4 (100.00)	0 (0.00)	6 (85.71)
Model explanation, n (%)
SHAP	9 (37.5)	4 (40.00)	0 (0.0)	1 (33.33)	4 (57.14)
Attention visualization	0 (0.00)	0 (0.00)	0 (0.00)	0 (0.00)	0 (0.00)

AKD, acute kidney disease; AKI, acute kidney injury; AUPRC, area under the precision-recall curve; AUROC, area under the receiver operating characteristic curve; NPV, negative predictive value; PPV, positive predictive value; SHAP, SHapley Additive exPlanations; SVM, support vector machine.

Analysis of feature selection

Figure 2 and Supplemental Table 4 show distributions of the most frequently selected features across studies in predicting drug-induced nephrotoxicity. Laboratory tests, including eGFR, serum creatinine, (SCr), Blood Urea Nitrogen (BUN), hemoglobin, and albumin, were selected by 10 (41.67%), 9 (37.5%), 9 (37.5%), 8 (33.33%), and 8 (33.33%) studies in all drug categories, respectively. Diabetes and heart failure were the most common comorbidities selected by 7 (29.17%) and 5 (20.83%) studies, respectively. The concurrent use of diuretics and NSAIDs was noted by 5 (20.83%) and 4 (16.67%) studies, respectively. The most frequently selected features for contrast agent-induced nephrotoxicity included eGFR, age, BUN, serum creatinine, hemoglobin, and diabetes. At least 50% of studies that focused on platinum agents selected eGFR, age, albumin, and sex as features. Notably, all vancomycin studies selected the drug’s blood concentration, and two of them (66.67%) selected creatinine clearance as a feature.

Figure 2.

The most selected features of drug-induced nephrotoxicity machine learning studies. The proportional use of various clinical predictors in drug-induced nephrotoxicity studies is illustrated in bar plots. Columns represent causative drug classes, the y-axis represents features, the x-axis represents proportion, and colors represent categories.

Comparison of model performances

Table 1 shows the frequency of each metric employed by the 24 studies. All studies (n = 24, 100%) reported model performance using the AUROC, while accuracy, sensitivity, and specificity were also widely reported by 19 (79.17%), 20 (83.33%), and 18 (75%) studies, respectively. The metric for imbalanced data, AUPRC, was reported in only 6 (25%) studies.

Figure 3(a) presents the model performance for different algorithms. All studies reported AUROC values exceeding 0.75, with the NN algorithm achieving the highest median AUROC of 0.90. Sensitivity values of 0.84, 0.75, and 0.67 were recorded for NNs, BMs, and RFs, respectively, while specificity values were 0.89, 0.77, and 0.80 for those corresponding models. F1-scores, which balance precision and recall, were 0.67 for NNs, 0.32 for BMs, and 0.45 for RFs. NNs demonstrated the narrowest variability across all evaluated metrics. BMs exhibited wide variations in performance across several metrics, particularly sensitivity, PPV, and F1-score. RFs also demonstrated pronounced variations in the PPV, NPV, and F1-score.

Figure 3.

Model performances of drug-induced nephrotoxicity machine learning studies across different metrics. (a) The box plot compares performance distributions of various machine learning algorithms. The x-axis represents key evaluation metrics, the y-axis represents the corresponding scores, and colors represent the machine learning algorithms employed. (b) The box plot compares the performance score distributions for predictive models of different drug-induced nephrotoxicity. Colors represent inducing drug agents.

Figure 3(b) shows the performance metrics of studies across different drug categories. AUROC scores ranged 0.75–0.85. Notably, median AUROC scores were 0.82 and 0.85 for vancomycin- and other drug-induced nephrotoxicity models, respectively, while they were 0.81 and 0.76 for contrast- and platinum-induced models, respectively. Median accuracy ranged 0.74–0.89, with 0.89, 0.79, 0.76, and 0.74 for vancomycin-, other drug-, contrast-, and platinum-induced nephrotoxicity, respectively. Supplemental Table 7 shows the performance metrics of each study.

Table 2 presents selected deeper comparative analyses for a few specific comparisons. Models with and without age showed no statistically significant difference in AUROCs (mean AUROC ± SD: 0.811 ± 0.092 vs 0.840 ± 0.058). Excluding diabetes resulted in a statistically insignificant increase in AUROC, from 0.810 ± 0.095 (diabetes) to 0.834 ± 0.067 (no diabetes). Including eGFR yielded an AUROC that was not statistically different from that when it was excluded (0.815 ± 0.065 vs 0.844 ± 0.088). A statistically insignificant difference in AUROC was observed between with and without diuretics (0.878 ± 0.054 vs 0.813 ± 0.075, p = 0.087). Models that underwent external validation reported an AUROC that was statistically insignificant from those that did not (median AUROC (IQR): 0.817 (0.063) vs 0.827 (0.096), p = 0.105). Studies employing appropriate missing data handling achieved a significantly higher AUROC than those using inappropriate handling (0.825 (0.069) vs 0.729 (0.074), p < 0.001). Only 9 (37.5%) studies used SHapley Additive exPlanations (SHAP) as the explainability approach, while none used attention visualization (Table 1).

Table 2.

Comparative analysis of model performance for selected features or study methods.

Aspect	Category	AUROC		p Value
Aspect	Category	With	Without	p Value
Inclusion of a particular feature	Age (mean ± SD)	0.811 ± 0.092	0.840 ± 0.058	0.365^*,$
	Diabetes mellitus (mean ± SD)	0.810 ± 0.095	0.834 ± 0.067	0.480^*,$
	eGFR (mean ± SD)	0.844 ± 0.088	0.815 ± 0.065	0.369^*,$
	Diuretics (mean ± SD)	0.878 ± 0.054	0.813 ± 0.075	0.087^*,$
Study method	External validation (median (IQR))	0.817 (0.063)	0.827 (0.096)	0.105^‡,$
Engineering strategy	Appropriate missing data handling (median (IQR))^a	0.825 (0.069)	0.729 (0.074)	<0.001^‡,***

AUROC, area under receiver operating characteristic; eGFR, estimated glomerular filtration rate; IQR, interquartile range; SD, standard deviation.

Refer to PROBAST+AI domain 4.3

Independent t test.

p > 0.05.

‡

Mann–Whitney U test.

***

p < 0.001.

Quality assessment of the machine learning studies

Figure 4 displays the 2025 PROBAST+AI risk-of-bias results, indicating that 11 (45.83%) studies had a high risk of bias in the overall assessment. The predictor and outcome domains had the lowest risk of bias, with 19 (79.17%) and 22 (91.67%) studies, respectively, rated as low risk. The participant domain had 9 (37.5%) studies with unclear risk and 4 (16.67%) with high risk. The highest proportion of studies at high risk of bias was observed in the analysis domain (n = 9, 37.5%; Supplemental Table 8). Figure 5 further showed that, within the analysis domain, the subdomains of performance evaluation and sample size had relatively high proportions of studies rated as high risk, with 20 (83.33%) and 18 (75.00%), respectively. In contrast, all studies were rated as having a low risk of bias in the three subdomains of data source, uniform definition/assessment for predictors, and outcomes (Supplemental Table 8).

Figure 4.

Proportions of drug-induced nephrotoxicity machine learning studies rated using 2025 PROBAST+AI. The risk-of-bias assessment is summarized in the stacked bar plot. Each horizontal bar corresponds to a study domain, and its colored segments illustrate the proportion of studies rated as having high, unclear, or low risk. The risk of bias of each domain was determined by answers to signaling questions, where a low-risk rating was assigned if all responses were affirmative. An unclear risk rating was assigned due to insufficient information, while any negative response required reviewers to apply their judgment to determine a final rating of low, high, or unclear risk.

Figure 5.

Frequencies of aspects with high and unclear risk of bias in drug-induced nephrotoxicity machine learning studies. Details of the risk-of-bias assessment are illustrated in the stacked bar plot, similar to the overall risk-of-bias assessment. Each horizontal bar represents a specific methodological subdomain.

Figure 6 illustrates the overall quality of the studies evaluated using 12 criteria specifically designed to assess machine learning studies of drug-induced nephrotoxicity prediction. Five (20.83%) studies had high overall quality, while 14 (58.33%) and 5 (20.83%) had moderate and low overall quality, respectively. Criteria achieving a quality rate of over 60% were predictive time for nephrotoxicity, handling of continuous features, feature definition, feature selection, and reporting of final features, with respective rates of 83.33%, 83.33%, 75%, 79.17%, and 87.5%. In contrast, criteria with low quality rates included reporting the number of participants with missing data, performing external validation, and using multicenter data sources, with respective quality rates of only 18.75%, 29.17%, and 33.33% (Supplemental Table 9).

Figure 6.

Quality assessment of drug-induced machine learning studies. The quality assessment of each study is summarized in the summary plot. The x-axis on the dot matrix on the left represents individual criterion scores, the y-axis represents the quality of each study, and the x-axis on the bar chart on the right represents aggregated quality scores. Colors represent either drug classes or quality points.

Discussion

This review revealed several key findings on the current progress of machine learning development in predicting drug-induced nephrotoxicity. The number of publications has significantly increased over the last 2 years, underscoring the importance of applying innovative prediction methods to prevent drug-induced nephrotoxicity. A clear evolution in selecting algorithms was also observed. A notable overlap in predictive features was also identified across different types of drug-induced nephrotoxicity. NNs and BMs consistently demonstrated superior predictive performances in predicting drug-induced nephrotoxicity. The present study applied integrated tools to assess the included studies for providing a balanced clinical appraisal. The frontline innovations of current machine learning models for predicting drug-induced nephrotoxicity, with further methodological improvements, hold high potential for clinical application in the near future.

The risk-of-bias analysis using integrated tools in the current study highlighted the methodological strengths and weaknesses of current studies for predicting drug-induced nephrotoxicity. The comparative analyses in Table 2 showed that appropriately handling missing data substantially enhanced model performance. Most studies carefully performed the process of defining and selecting features, handling continuous features, and determining final features. However, very few studies calibrated models to assess the accuracy and reliability of predicted probabilities.^15,50 Only a limited number of studies incorporated model explanation techniques, such as SHAP or attention visualization, to elucidate how predictions were derived. This lack of transparency can hinder clinicians’ trust and understanding of the models, potentially limiting their adoption in clinical decision-making and compromising the interpretability needed for safe and effective integration into practice. The issue of small sample sizes observed in several studies can lead to model overfitting and performance overestimation.^51,52 External validation and multicenter data sources to increase sample size and generality will be necessary to strengthen current drug-induced nephrotoxicity models. Utilizing an international database or a shared database through collaborative research can be useful to enhance the robustness of the models.

The study showed that drug-related nephrotoxicity was highly associated with the patient’s baseline condition. Age was one of the most commonly selected features, and its associated declines in glomerular filtration rate and renal reserve were supported by epidemiological and meta-analytic studies.⁵³ Diabetes mellitus, the primary comorbidities identified, reflect their established contributions to preexisting renal impairment and heightened susceptibility to drug toxicity. On the other hand, the comparative analyses in Table 2 revealed that including age, diabetes, eGFR, or diuretics does not significantly affect model performance. The results indicated that the feature selection methods employed in each study helped build robust models to reduce the complex confounding effects and causal relationships among factors.

Additional frequently selected dynamic features included SCr, BUN, hemoglobin, and albumin. The inherent antioxidant properties of albumin may help attenuate contrast-induced nephrotoxicity, a condition partially mediated by oxidative stress.^54,55 Low hemoglobin levels can precipitate AKI, presumably by causing renal hypoxia, thereby compromising essential oxygen delivery to the kidneys.⁵⁶ A high trough concentration of vancomycin was a consistently strong predictor of AKI.⁵⁷ Diuretic use was a commonly selected feature, which agrees with pathophysiologic and pharmacologic mechanisms. These rapidly changing laboratory data or modifiable medication use provided an opportunity for actionable strategies to interrupt the AKI progression.

The differences in the features selected by individual models can be explained by the pharmacologic mechanisms underlying those models. Nephrotoxicity associated with vancomycin, cisplatin, and contrast media involves distinct but some overlapping pathophysiological pathways. While all three primarily target the renal tubules, their mechanisms differ in how they initiate injury. The primary mechanism of vancomycin-induced nephrotoxicity is excessive oxidative stress that targets tubular cells.⁵⁸ Cisplatin-associated nephrotoxicity is highly related to DNA damage and also related to the organic cation transporters (OCT2) in the kidney.⁵⁹ However, contrast media induced nephrotoxicity by direct contact with the high-osmolality, high-viscosity molecules.⁶⁰ These differences in pharmacological mechanisms led to distinct features in each model.

The drug-induced nephrotoxicity models included in this review mostly predicted AKI, which can be extended to long-term outcomes for leading comprehensive renal care. The continuum of renal impairment is initiated at the occurrence of AKI. Should renal function not fully recover during this initial period, the patient then enters the AKD phase.⁶¹ Preventing AKD is of paramount importance as this critical window determines whether the kidney enters into either full recovery or permanent CKD. Most studies lack explicit modeling of temporal dynamics, with risk predictions largely confined to predose assessments. Consequently, these models offer limited insight into temporal risk evolution or the optimal timing for postdose interventions. However, a shorter prediction window (e.g., within 7 days) remains crucial for identifying patients at high risk of immediate progression and guiding prompt intervention or change in drug therapy. A longer prediction window (e.g., up to 90 days) provides a comprehensive risk assessment for the development of sustained AKD or subsequent CKD. The machine learning model design for drug-induced nephrotoxicity should accommodate the concepts of AKI and AKD to reinforce long-term renal care strategies and provide a comprehensive framework for prevention.

The current study found that more complex methods, such as BMs and NNs, performed better than algorithms created by RFs, DTs, and SVMs. Despite that, tree-based BMs possess inherent capabilities for variable selection, causal effect estimation, and sophisticated handling of missing data.⁶² This review showed that NNs, by capturing complex, non-linear relationships, had better and more balanced positive and negative predictions in all of the performance metrics.^63,64 The most significant capacity of these advanced algorithms is to model time-series data. The simultaneous prediction of kinetic changes can timely reflect the progress of disease which is more practical for clinical use. However, high computational requirements may limit clinical deployment in resource-limited settings. A feasibility analysis should be carefully conducted to determine whether the clinical settings have adequate technological support to implement models with high computational requirements.

Selecting machine learning algorithms to predict and support clinical practice highly depends on data type and the needs of clinical specialties. Convolutional neural networks excel in oncology and pathology by analyzing complex medical images to aid in diagnosis and determine cancer stage.⁶⁵ On the other hand, long short-term memory networks can be used in intensive care medicine or in medical conditions that require analyzing sequential physiological data to predict outcomes or identify risk.⁶⁶ Other algorithms, such as NNs and BMs, are common in nephrology in utilizing EHR data to predict acute disease diagnoses or chronic disease progression and to stratify risk. Recent studies reported that gradient BMs and extreme gradient BMs achieved good predictive performance for vancomycin and amikacin, respectively.^67,68

Incorporating machine learning models into clinical decision support systems (CDSSs)⁶⁹ is highly feasible as current models use clinically obtainable features but it requires overcoming barriers. The machine learning-based CDSSs⁷⁰ can provide clinicians with simultaneous patient-specific recommendations, serving as a tangible strategy for preventing drug-induced nephrotoxicity. However, potential barriers to clinical practical applications include the computational capacity of the EHR system, regulatory oversight of AI/software-assisted medical devices (SaMD), and model interpretation. Training on machine learning or AI-assisted CDSSs is essential before clinical implementation to address prediction errors, the circumstances under which a model might fail, or the criteria for determining when model outputs are reliable.⁷¹

Limitations

This review has several limitations. Literature searching was limited to published articles and did not encompass potential unpublished studies. This approach introduces the possibility of publication bias, as studies with null or negative findings may have been underrepresented. The preponderance of studies conducted in East Asian and North American/European populations limits the generalizability of the findings, particularly concerning African populations, which had higher CKD prevalence. The variability in diagnostic criteria for nephrotoxicity and the aggregation of data by drug class in some studies, rather than individual agents, constrain the interpretation of these results and limit the specificity of the findings. While most research has focused on contrast-induced nephrotoxicity, interpreting results from different drug-induced nephrotoxicity models should be done carefully. A significant limitation was that many articles included in this review did not undergo external validation. Internal validation, regardless of the method used, can be very similar to the training set and generate overly optimistic predictions. The actual performance of these algorithms on new, independent patient data remains mainly unknown and unverified.

Future studies

Future studies should prioritize the cohorts with a high prevalence of CKD, such as in African countries, given the disproportionate burden of kidney disease.⁷² Standardized nephrotoxicity criteria and drug-specific analyses should be applied to yield more robust and reliable results. Given the scarcity of studies benchmarking machine learning models against established clinical guidelines—such as KDIGO or RIFLE—future prospective trials are warranted to evaluate the comparative effectiveness of machine learning-integrated care versus routine clinical practice. Lastly, reasonable clinical actions to avoid drug-induced side effects should be taken while balancing the pharmacologic effects. Since dose attenuation may compromise the cytotoxicity of antineoplastic agents or the bactericidal activity of antibiotics such as vancomycin, the risk of treatment failure remains a significant concern. Consequently, future research should prioritize the development of clinical actionability frameworks that include extensive hydration protocols or switching therapeutic agents, effectively balancing nephrotoxicity mitigation with optimal therapeutic outcomes.

Conclusion

Recent advancements in machine learning algorithms for predicting drug-induced nephrotoxicity indicate the high potential for clinical implementation in the near future. The feasibility analysis of the computational resources required to analyze time-series data and the optimal selection of algorithms should be carefully assessed, depending on the specific data types and the actual needs of clinical settings. These models of drug-induced nephrotoxicity utilizing clinically obtainable features have achieved excellent AUROCs and revealed that drug-related nephrotoxicity is profoundly associated with a patient’s baseline condition. To ensure a comprehensive scope of prevention, future study designs and algorithms should accommodate the concepts of both AKI and AKD, thereby reinforcing long-term renal care strategies. Methodological rigor, including performance evaluations, sufficient sample sizes, and external validation, must be thoroughly considered and applied to ensure the validity and reliability of the results. Embracing multicenter or multicountry datasets and adhering to standardized guidelines, such as PROBAST+AI and TRIPOD+AI, can further enhance the validity and generalizability of predictive models for drug-induced nephrotoxicity. Future studies must prioritize clinically-actionable frameworks in the prediction model that balance nephrotoxicity mitigation with therapeutic outcomes. Integrating these models into CDSS is highly feasible; however, successful deployment requires overcoming implementation barriers to provide timely estimates, enabling the proactive prevention of drug-induced nephrotoxicity.

Supplemental Material

sj-docx-1-taw-10.1177_20420986261430234 – Supplemental material for Machine learning studies of drug-induced nephrotoxicity: a scoping review

Supplemental material, sj-docx-1-taw-10.1177_20420986261430234 for Machine learning studies of drug-induced nephrotoxicity: a scoping review by Mawardi Ihsan, Shu-Ting Chang, Wei-Kai Chan and Hsiang-Yin Chen in Therapeutic Advances in Drug Safety

Footnotes

Acknowledgements

The authors thank Taipei Medical University for providing the library resources and the National Science and Technology Council for supporting this work through research grants.

Declarations

ORCID iDs

Mawardi Ihsan

Shu-Ting Chang

Hsiang-Yin Chen

Supplemental material

Supplemental material for this article is available online.

References

Mody

Ramakrishnan

Chaar

, et al. A review on drug-induced nephrotoxicity: pathophysiological mechanisms, drug classes, clinical management, and recent advances in mathematical modeling and simulation approaches. Clin Pharmacol Drug Dev 2020; 9(8): 896–909.

Whiting

Morden

Tomlinson

, et al. What are the risks and benefits of temporarily discontinuing medications to prevent acute kidney injury? A systematic review and meta-analysis. BMJ Open 2017; 7(4): e012674.

Crona

Faso

Nishijima

, et al. A systematic review of strategies to prevent cisplatin-induced nephrotoxicity. Oncologist 2017; 22(5): 609–619.

Brar

Hiremath

Dangas

, et al. Sodium bicarbonate for the prevention of contrast induced-acute kidney injury: a systematic review and meta-analysis. Clin J Am Soc Nephrol 2009; 4(10): 1584.

Doorn

WPTM

van Stassen

Borggreve

, et al. A comparison of machine learning models versus clinical evaluation for mortality prediction in patients with sepsis. PLoS One 2021; 16(1): e0245157.

Heo

Kang

, et al. Time series AI model for acute kidney injury detection based on a multicenter distributed research network: development and verification study. JMIR Med Inform 2024; 12(1): e47693.

Lafrance

Miller

. Acute kidney injury associates with increased long-term mortality. J Am Soc Nephrol 2010; 21(2): 345.

Stottlemyer

Tran

Suh

, et al. A systematic review of the costs of drug-associated acute kidney injury and potential cost savings with nephrotoxin stewardship prevention strategies. Clin Pharmacol Ther 2025; 117(4): 989–1004.

Kung

Chou

. Acute kidney disease: an overview of the epidemiology, pathophysiology, and management. Kidney Res Clin Pract 2023; 42(6): 686–699.

10.

Sawhney

Ball

Bell

, et al. Recovery of kidney function after acute kidney disease: a multi-cohort analysis. Nephrol Dial Transplant 2024; 39(3): 426–435.

11.

Zhang

Ghahramani

, et al. Prediction of acute and chronic kidney diseases during the post-covid-19 pandemic with machine learning models: utilizing national electronic health records in the US. eBioMedicine 2025; 115: 105726.

12.

Halder

Uddin

, et al. ML-CKDP: machine learning-based chronic kidney disease prediction with smart web application. J Pathol Inform 2024; 15: 100371.

13.

Dhiman

, et al. Sample size requirements are not being considered in studies developing prediction models for binary outcomes: a systematic review. BMC Med Res Methodol 2023; 23(1): 188.

14.

Collins

de Groot

Dutton

, et al. External validation of multivariable prediction models: a systematic review of methodological conduct and reporting. BMC Med Res Methodol 2014; 14(1): 40.

15.

Moons

KGM

Wolff

Riley

, et al. PROBAST: a tool to assess risk of bias and applicability of prediction model studies: explanation and elaboration. Ann Intern Med 2019; 170(1): W1–W33.

16.

Moons

KGM

de Groot

JAH

Bouwmeester

, et al. Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist. PLoS Med 2014; 11(10): e1001744.

17.

Kwong

JCC

Khondker

Lajkosz

, et al. APPRAISE-AI tool for quantitative evaluation of AI studies for clinical decision support. JAMA Netw Open 2023; 6(9): e2335377.

18.

Collins

Moons

KGM

Dhiman

, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024; 385: e078378.

19.

Guan

, et al. Predictive value of machine learning for the risk of acute kidney injury (AKI) in hospital intensive care units (ICU) patients: a systematic review and meta-analysis. PeerJ 2023; 11: e16405.

20.

Cama-Olivares

Braun

Takeuchi

, et al. Systematic review and meta-analysis of machine learning models for acute kidney injury risk classification. J Am Soc Nephrol 2025; 36(10): 1969–1983

21.

Zhu

Yan

. Predictive models of sepsis-associated acute kidney injury based on machine learning: a scoping review. Ren Fail 2024; 46(2): 2380748.

22.

Poly

Weng

, et al. Machine learning models for predicting mortality in critically ill patients with sepsis-associated acute kidney injury: a systematic review. Diagnostics 2024; 14(15): 1594.

23.

Liu

Chen

, et al. Machine learning for the prediction of mortality in patients with sepsis-associated acute kidney injury: a systematic review and meta-analysis. BMC Infect Dis 2024; 24(1): 1454.

24.

Tricco

Lillie

Zarin

, et al. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med 2018; 169(7): 467–473.

25.

Moons

KGM

Damen

JAA

Kaul

, et al. PROBAST+AI: an updated quality, risk of bias, and applicability assessment tool for prediction models using regression or artificial intelligence methods. BMJ 2025; 388: e082505.

26.

Yin

Guan

, et al. Preprocedural prediction model for contrast-induced nephropathy patients. J Am Heart Assoc 2017; 6(2): e004498.

27.

Garcia

Lauritsen

Zhang

, et al. Prediction of nephrotoxicity associated with cisplatin-based chemotherapy in testicular cancer patients. JNCI Cancer Spectr 2020; 4(3): pkaa032.

28.

Imai

Takekuma

Kashiwagi

, et al. Validation of the usefulness of artificial neural networks for risk prediction of adverse drug reactions used for individual patients in clinical practice. PLoS One 2020; 15(7): e0236789.

29.

Sun

Zhu

Chen

, et al. Machine learning to predict contrast-induced acute kidney injury in patients with acute myocardial infarction. Front Med 2020; 7: 592007.

30.

Chen

, et al. A predictive model based on a new CI-AKI definition to predict contrast induced nephropathy in patients with coronary artery disease with relatively normal renal function. Front Cardiovasc Med 2021; 8: 762576

31.

Huang

Chu

Hsu

, et al. How platinum-induced nephrotoxicity occurs? Machine learning prediction in non-small cell lung cancer patients. Comput Methods Programs Biomed 2022; 221: 106839.

32.

Kim

Yee

, et al. Risk scoring system for vancomycin-associated acute kidney injury. Front Pharmacol 2022; 13: 815188

33.

Cui

Tang

, et al. Analysis of a machine learning-based risk stratification scheme for acute kidney injury in vancomycin. Front Pharmacol 2022; 13: 1027230.

34.

Okawa

Mizuno

Hanabusa

, et al. Prediction model of acute kidney injury induced by cisplatin in older adults using a machine learning algorithm. PLoS One 2022; 17(1): e0262021.

35.

, et al. Identifying patients at risk of acute kidney injury among patients receiving immune checkpoint inhibitors: a machine learning approach. Diagnostics 2022; 12(12): 3157.

36.

Akimoto

Hayakawa

Nagashima

, et al. Detection of potential drug-drug interactions for risk of acute kidney injury: a population-based case-control study using interpretable machine-learning models. Front Pharmacol 2023; 14: 1176096.

37.

Chen

Liu

Shen

, et al. Development of real-time individualized risk prediction models for contrast associated acute kidney injury and 30-day dialysis after contrast enhanced computed tomography. Eur J Radiol 2023; 167: 111034.

38.

Cox

Panagides

Di Capua

, et al. An interpretable machine learning model for the prevention of contrast-induced nephropathy in patients undergoing lower extremity endovascular interventions for peripheral arterial disease. Clin Imaging 2023; 101: 1–7.

39.

Güven

Özdede

Şener

, et al. Evaluation of machine learning algorithms for renin-angiotensin-aldosterone system inhibitors associated renal adverse event prediction. Eur J Intern Med 2023; 114: 74–83.

40.

Wan

Chen

, et al. A novel explainable online calculator for contrast-induced AKI in diabetics: a multi-centre validation and prospective evaluation study. J Transl Med 2023; 21(1): 517.

41.

, et al. Prediction of the development of contrast‑induced nephropathy following percutaneous coronary artery intervention by machine learning. Acta Cardiol 2023; 78(8): 912–921.

42.

Yan

Duan

Luo

, et al. Development and validation of a deep neural network–based model to predict acute kidney injury following intravenous administration of iodinated contrast media in hospitalized patients with chronic kidney disease: a multicohort analysis. Nephrol Dial Transplant 2023; 38(2): 352–361.

43.

Zhou

, et al. Correlation between neutrophil-to-lymphocyte ratio and contrast-induced acute kidney injury and the establishment of machine-learning-based predictive models. Ren Fail 2023; 45(2): 2258983.

44.

Chiu

Chan

, et al. Machine learning algorithms to predict colistin-induced nephrotoxicity from electronic health records in patients with multidrug-resistant Gram-negative infection. Int J Antimicrob Agents 2024; 64(1): 107175.

45.

Choi

Han

, et al. Applicable machine learning model for predicting contrast-induced nephropathy based on pre-catheterization variables. Intern Med 2024; 63(6): 773–780.

46.

Noda

Mizuno

Mogushi

, et al. Development of a predictive model for nephrotoxicity during tacrolimus treatment using machine learning methods. Br J Clin Pharmacol 2024; 90(3): 675–683.

47.

Sakuragi

Uchino

Sato

, et al. Interpretable machine learning-based individual analysis of acute kidney injury in immune checkpoint inhibitor therapy. PLoS One 2024; 19(3): e0298673.

48.

Zhang

Luo

Fan

, et al. Development and validation of a LASSO prediction model for cisplatin induced nephrotoxicity: a case-control study in China. BMC Nephrol 2024; 25(1): 194.

49.

Zhang

Lao

Chen

, et al. Development and validation of a machine learning algorithm‑based risk prediction model of esomeprazole‑associated acute kidney injury [in Chinese]. Adverse Drug React J 2024; 26(7): 405–411.

50.

Moons

KGM

Altman

Reitsma

, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 2015; 162(1): W1–W73.

51.

Vergouwe

Steyerberg

Eijkemans

MJC

, et al. Substantial effective sample sizes were required for external validation studies of predictive logistic regression models. J Clin Epidemiol 2005; 58(5): 475–483.

52.

Riley

Snell

Ensor

, et al. Minimum sample size for developing a multivariable prediction model: PART II—binary and time-to-event outcomes. Stat Med 2019; 38(7): 1276–1296.

53.

Grams

Sang

Ballew

, et al. A meta-analysis of the association of estimated GFR, albuminuria, age, race, and sex with acute kidney injury. Am J Kidney Dis 2015; 66(4): 591–601.

54.

Heyman

Rosen

Khamaisi

, et al. Reactive oxygen species and the pathogenesis of radiocontrast-induced nephropathy. Invest Radiol 2010; 45(4): 188.

55.

Roche

Rondeau

Singh

, et al. The antioxidant properties of serum albumin. FEBS Lett 2008; 582(13): 1783–1787.

56.

Rosenberger

Rosen

Heyman

. Renal parenchymal oxygenation and hypoxia adaptation in acute kidney injury. Clin Exp Pharmacol Physiol 2006; 33(10): 980–988.

57.

Katip

Okonogi

Oberdorfer

. The thirty-day mortality rate and nephrotoxicity associated with trough serum vancomycin concentrations during treatment of enterococcal infections: a propensity score matching analysis. Front Pharmacol 2022; 12: 773994.

58.

Tabarzad

Torshabi

Heidari

, et al. Vancomycin insights: an update on mechanism, activity, toxicity, resistance, and novel drug delivery systems. Iran J Pharm Res 2025; 24(1): e160885.

59.

Tang

Livingston

Safirstein

, et al. Cisplatin nephrotoxicity: new insights and therapeutic implications. Nat Rev Nephrol 2023; 19(1): 53–72.

60.

Andreucci

Faga

Pisani

, et al. Prevention of contrast-induced nephropathy through a knowledge of its pathogenesis and risk factors. Sci World J 2014; 2014(1): 823169.

61.

Chawla

Bellomo

Bihorac

, et al. Acute kidney disease and renal recovery: consensus report of the Acute Disease Quality Initiative (ADQI) 16 Workgroup. Nat Rev Nephrol 2017; 13(4): 241–257.

62.

. Using tree-based machine learning for health studies: literature review and case series. Int J Environ Res Public Health 2022; 19(23): 16080.

63.

Lisboa

PJG

. A review of evidence of health benefit from artificial neural networks in medical intervention. Neural Netw 2002; 15(1): 11–39.

64.

Schmidhuber

. Deep learning in neural networks: an overview. Neural Netw 2015; 61: 85–117.

65.

Esteva

Kuprel

Novoa

, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 2017; 542(7639): 115–118.

66.

Scherpf

Gräßer

Malberg

, et al. Predicting sepsis with a recurrent neural network using the MIMIC III database. Comput Biol Med 2019; 113: 103395.

67.

Yang

Yan

Wang

, et al. Clinical risk assessment of serum creatinine abnormalities during vancomycin therapy: a retrospective study using machine learning models. Int J Clin Pharm 2025; 47: 1830–1840.

68.

Zhang

Chen

Lao

, et al. Machine learning modeling for the risk of acute kidney injury in inpatients receiving amikacin and etimicin. Front Pharmacol 2025; 16: 1538074.

69.

Sutton

Pincock

Baumgart

, et al. An overview of clinical decision support systems: benefits, risks, and strategies for success. NPJ Digit Med 2020; 3(1): 17.

70.

Peiffer-Smadja

Rawson

Ahmad

, et al. Machine learning for clinical decision support in infectious diseases: a narrative review of current applications. Clin Microbiol Infect 2020; 26(5): 584–595.

71.

Rasheed

Qayyum

Ghaly

, et al. Explainable, trustworthy, and ethical machine learning for healthcare: a survey. Comput Biol Med 2022; 149: 106043.

72.

Bikbov

Purcell

Levey

, et al. Global, regional, and national burden of chronic kidney disease, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet 2020; 395(10225): 709–733.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.27 MB