Construction of a Machine Learning-Based Risk Prediction Model for Drug-Induced Acute Kidney Injury in Elderly Patients

Abstract

Background

Elderly individuals are particularly vulnerable to drug-induced acute kidney injury (DI-AKI) due to their distinct physiological and pathophysiological traits. DI-AKI’s non-specific symptoms complicate the identification of causative medications, highlighting the urgent need for accurate predictive tools to detect AKI risk early in this group.

Aim

The objective is to construct a machine learning-based predictive model for DI-AKI in elderly patients utilizing real-world data, with the aim of offering a decision-support tool for the early clinical identification of DI-AKI risk.

Methods

The electronic health records of 2,389 patients aged ≥60 years at Shenzhen Luohu People’s Hospital from January 2023 to December 2024 were retrospectively analyzed. Drug-induced AKI was defined by KDIGO creatinine criteria (≥0.3 mg/dL within 48h or ≥1.5×baseline within 7d) plus Naranjo score for drug attribution. Forty optimal features (30 original and 10 interaction terms) were selected, and seven machine learning algorithms were assessed using nested cross-validation with different feature selection and interactive feature construction methods. SHAP values were employed for model interpretability.

Results

 In a study of 2,389 patients, 39.9% (953 individuals) experienced drug-induced AKI. The Random Forest model performed best on the test set, with an Area Under the Curve (AUC) of 0.9209 [95% CI: 0.896–0.946], showing 84.9% sensitivity, 89.1% specificity, and an 85.5% positive predictive value. Key predictors included a history of renal failure (importance score: 0.1703) and drug-disease interactions, such as those involving antihypertensives and renal failure (importance score: 0.1000).

Conclusion

 This machine learning model is capable of aiding in the identification of high-risk elderly patients within electronic medical record systems, with a particular emphasis on drug interactions and renal function as pivotal risk factors. By employing SHAP values for analysis, this study elucidates the contributions of these risk factors and offers support for making personalized medication decisions; However, external validation remains necessary prior to clinical implementation.

Plain Language Summary

Why did we do this study? Older adults often take many medicines at the same time. Sometimes, these medicines can hurt their kidneys. We wanted to see if a computer could learn to predict which patients are at higher risk.

What did we do? We analyzed health records from 2,389 older patients at a hospital in Shenzhen. We used computer programs to learn patterns from these records. We tested seven different computer methods to find the best one.

What did we find? The best method was called “random forest.” It correctly identified about 85 out of 100 patients who would develop kidney injury and about 89 out of 100 who would not. We also found that patients who already had kidney problems and were taking certain blood pressure or blood thinner medicines at the same time had the highest risk.

What do these results mean? This computer tool could help doctors identify high-risk patients and adjust their medicines more carefully. However, this study was only done using data from one hospital and has not been tested in real clinical practice yet. More testing in other hospitals is needed before it can be used regularly.

Introduction

Drug-Induced Kidney Injury (DIKI) refers to abnormalities in renal function or structure following drug exposure, presenting with diverse clinical manifestations. These manifestations include acute kidney injury (AKI), chronic kidney disease (CKD), acute interstitial nephritis (AIN), proteinuria, and nephrotic syndrome due to glomerular damage. Among these conditions, AKI is the most prevalent form of drug-induced renal injury.1 It is particularly common among hospitalized patients, especially those in intensive care units (ICUs), where its incidence can reach up to 60%.2 Nephrotoxic medications are a significant factor in the development of AKI, contributing to approximately 25% of all cases.3

Currently, China is undergoing a rapid transition towards becoming a moderately aging society. By the conclusion of 2024, it is anticipated that 22.0% of the population will be aged 60 years and older, with 15.6% being 65 years and older.4 The incidence of hospital-acquired AKI among community-dwelling elderly increases substantially with age: 26.8 per 1,000 person-years for those aged ≥70 years and 54.6 per 1,000 person-years for those aged ≥90 years.5 Drug-related factors are a leading cause of AKI in elderly hospitalized patients. On one hand, advancing age is associated with a natural decline in renal function, such as decreased Glomerular Filtration Rate (GFR) and reduced renal blood flow, which diminish renal resilience to drug toxicity.6 On the other hand, elderly individuals frequently present with multiple chronic underlying conditions, necessitating prolonged use of various potentially nephrotoxic medications, thereby compounding the risk of renal injury.7,8 Furthermore, the concurrent use of multiple drugs may result in synergistic toxic effects through intricate pharmacokinetic or pharmacodynamic interactions.9

The prognosis of drug-induced AKI (DI-AKI) is critically dependent on the early identification and prompt cessation of the implicated drugs, as timely intervention can substantially enhance patient outcomes. However, DI-AKI frequently manifests with nonspecific clinical symptoms, which often lead to oversight or misdiagnosis, thereby complicating the rapid and accurate attribution to specific drugs. Furthermore, the pathogenic mechanisms of drug-induced AKI are highly complex, encompassing direct drug cytotoxicity (eg, aminoglycosides), immune-mediated damage (eg, NSAIDs), intratubular crystallization or aggregate formation (eg, vancomycin), and interindividual metabolic variability10 All of these complexities further increase the difficulty of drug traceability Therefore, in the context of the lack of reliable ultra-early biomarkers, there is an urgent need for precise prediction to achieve early warning of AKI risk.

Regarding research on monitoring and early warning of acute kidney injury (AKI), a UK teaching hospital established an automated real-time alert system in 2014, which improved patient outcomes, reduced mortality, shortened hospital stays, and enhanced AKI management through early detection and intervention.11 Although China started later in this field, it has progressed rapidly; for example, Guo Daihong’s team at the Chinese PLA General Hospital developed the ADE-ASAS and ADE-ASAS-II systems based on HIS, using trigger principles and text recognition for active monitoring of adverse drug events.12 At the methodological level, domestic scholars have introduced machine learning into pharmacovigilance: Lin et al used XGBoost and SHAP to identify associations between immunomodulators, antiviral drugs, and abortion risk using the FAERS database,13 and Hu Yanghui et al built an intelligent query platform for ADR signal mining based on the DeepSeek AI model.14 Traditional signal detection methods like ROR and PRR struggle with multi-source heterogeneous data and nonlinear relationships, whereas machine learning methods have shown great potential in assisting code generation and application development, offering new solutions for the digital transformation of pharmacovigilance.14,15 Meanwhile, the rise of real-world studies (RWS) also provides excellent methodological support for risk warning analysis of adverse drug reactions or drug-induced diseases at home and abroad.

The electronic health records (EHRs) system has accumulated extensive real-world clinical data, providing valuable resources for elucidating complex clinical patterns. By employing big data analytics, particularly machine learning methodologies, it becomes possible to identify multidimensional features derived from EHRs that are significantly associated with drug-induced AKI development and to construct precise risk prediction models.16 Several AKI prediction models have been developed in recent years, ranging from traditional logistic regression to machine learning algorithms such as random forest and XGBoost.17 Many of these models have demonstrated good performance in intensive care or perioperative settings. However, current AKI prediction models have two major gaps. First, most models have been developed in intensive care or perioperative settings, and few are specifically designed for drug-induced AKI in elderly general ward patients.18 Second, existing models rarely incorporate drug-disease interaction terms, despite the fact that such interactions may significantly modify AKI risk in multimorbid elderly patients taking multiple medications.19 Moreover, Observational EHR-based studies face inherent challenges in attributing AKI to specific drugs, including confounding by indication, limitations of the Naranjo score, and an arbitrary time window——all of which were carefully considered in our study design.

To address these gaps, this study seeks to develop a machine learning-based prediction model for drug-induced AKI in elderly patients, utilizing real-world clinical data. Specifically, we will systematically extract and select key predictive features from patients’ clinical profiles, medication histories, comorbidities, and their complex interactions through comprehensive feature engineering. Subsequently, multiple machine learning algorithms will be employed to construct and rigorously evaluate the predictive model. Finally, we will apply interpretable methods, including SHAP (SHapley Additive exPlanations), to conduct an in-depth analysis of the model, identify core predictors, and establish a reliable foundation for early clinical risk assessment and targeted intervention strategies.

Data and Methods

Data Sources

This retrospective cohort study analyzed medical records of elderly patients (aged ≥60 years) who received care at Luohu District People’s Hospital or its affiliated community health centers from January 2023 to December 2024. The investigation employed extensive clinical data, encompassing demographic details, medical histories, medication records, laboratory results, and diagnostic codes obtained from EHRs to investigate drug-induced AKI risk factors in this population.

Study Population Inclusion

The diagnosis of AKI was primarily guided by the 2012 Kidney Disease Improving Global Outcomes (KDIGO) guidelines, with a study-specific adaptation using an elevation in serum creatinine of ≥1.5 times the baseline as the diagnostic threshold. This modification was due to data availability constraints at Luohu District People’s Hospital. Patients were excluded if their acute renal insufficiency was attributable to non-pharmacologic factors such as hypoperfusion, trauma, hepatorenal syndrome, or if they had a prior history of renal transplantation or resection, or were undergoing dialysis/resuscitation. The assessment of DI-AKI relevance considered the temporal association between renal injury and suspected drug exposure (occurring within two weeks prior to renal impairment), the history of common nephrotoxic drug use, and the Naranjo score.

Data Extraction

For patients who met the inclusion criteria, we systematically collected clinical data encompassing demographic characteristics (including gender and age), underlying comorbidities, medication history, and laboratory test results, with a particular focus on serum creatinine values at the onset of AKI.

Data Preprocessing

Following the extraction of raw data, preliminary processing was conducted using R, including data cleaning and calibration to remove missing and erroneous entries. Subsequently, the disease names were standardized and mapped to the ICD-11 classification system through vector search, keyword filtering, and large language model (LLM)-assisted semantic reasoning, supplemented by manual calibration. Similarly, drug data underwent standardization and classification. This process began according to the hospital’s internal system, and subsequently cross-referenced with the national health insurance catalog for ambiguous cases. The final drug category was then used as the model’s input variable.

After standardization, the variables were transformed as follows: gender was converted into a binary categorical variable, while underlying disease and drug category were binarized and one-hot encoded; age was retained as a continuous variable. The study’s outcome, abnormal creatinine values, was defined as a binary categorical variable.

Feature Screening and Model Construction

In this study, all statistical analyses and model development were performed using R and Python (version 3.11), leveraging core libraries including scikit-learn, XGBoost, LightGBM, imbalanced-learn, NumPy, and Pandas.

After the raw data were preprocessed, the data were divided into training and testing subsets in a 7:3 ratio using stratified sampling. The RobustScaler was employed to standardize continuous variables, specifically age. To identify the features most pertinent to the outcome variables, univariate ANOVA F-tests and mutual information metrics were applied to evaluate the relevance of each feature to the target variable. Additionally, absolute coefficients derived from Lasso regression, feature importance rankings from Random Forest and XGBoost, and the Recursive Feature Elimination (RFE) method using logistic regression as a baseline estimator were employed for comprehensive feature screening. Concurrently, a chi-square test was conducted on the drug and disease variable groups to select significant variables with a frequency of ≥20 and a P-value of <0.05, and both the risk ratio and odds ratio were calculated. Based on this approach, three categories of interaction terms——drug-disease, drug-age, and drug-sex——were developed and subsequently evaluated using both mutual information and F-test methodologies, with a significance threshold set at P < 0.01, limited to the top 10 interactions. The purpose was to identify and retain significant interaction features. Ultimately, the original features that ranked within the top 30 according to each method were combined with the significant interaction features to form the feature set employed for subsequent modeling.

In this study, seven machine learning algorithms—logistic regression, support vector machine, random forest, gradient boosting machine, XGBoost, LightGBM, and multilayer perceptron—were employed for model construction. A predefined hyperparameter search space was established for each model, encompassing regularization strength, tree depth, learning rate, and network structure. The nested cross-validation was implemented, with an outer 5-fold Stratified K-Fold for performance evaluation and an inner 3-fold cross-validation for hyperparameter optimization using either GridSearchCV or RandomizedSearchCV (n_iter=20). The scoring metric was standardized to the Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). The model performance was assessed using a comprehensive set of metrics, including ROC-AUC, Precision-Recall Area Under the Curve (PR-AUC), average precision, accuracy, sensitivity, specificity, precision, negative predictive value (NPV), and F1 score. The optimal classification thresholds were determined using the Youden index. Ultimately, the candidate model was retrained on the complete training dataset using the optimal parameters identified from the performance metrics.

Continuous variables were imputed using random forest-based multiple imputation (missForest) and categorical variables by mode, with all imputations performed within crossvalidation folds.

Results

Patient Characteristics

The study examined 2,389 records of elderly patients (≥60 years), including 953 (39.9%) in the drug‑induced AKI group and 1,436 (60.1%) in the control group. According to Table 1, the gender distribution was similar between the groups (P=0.216), with males comprising 49.2% of the AKI group and 51.8% of controls. A significantly higher percentage of octogenarians was observed in the AKI group (29.1%) compared to control group (13.6%, P<0.001). The average baseline creatinine level in the AKI group was 484.2 μmol/L.

Table 1.

Baseline Data of the Study Population

Variable	Drug-Induced AKI Group (n=953)	Control Group (n=1436)	P-value
Gender
Male	469	745	0.2168
Female	484	691	0.2168
Age Group
60-70 years	371	781	0.0000
71-80 years	304	459	1.0000
>80 years	278	196	0.0000
Age (years, mean±SD)	74.6±9.3	70.6±7.9	0.0000
Baseline creatinine (μmol/L, mean)	484.2	——	——

Notes: P-value≤0.05 was considered statistically significant.

Abbreviation: AKI, acute kidney injury;

The AKI cohort demonstrated a distinct clustering of kidney-related comorbidities (76.8%), with a significantly higher prevalence of kidney failure compared to the control group. Other prevalent conditions within this cohort included fluid, electrolyte, and acid-base disorders (22.5%), anemia and erythrocyte disorders (28.0%), and heart failure (27.1%). Conversely, the control group primarily exhibited conditions commonly associated with general hospitalizations, such as hepatic diseases, malignancies, gastric disorders, and thyroid dysfunction, all of which have lower nephrotoxic potential (Figure 1). An analysis of medication usage revealed a marked prevalence of nephrotoxic drug administration in the AKI group, particularly diuretics and dehydrating agents (41.5%), hematinics (41.2%), insulin analogs (38.8%), and acid-base regulators (37.9%). In contrast, the control group predominantly received conventional therapies with lower nephrotoxic risk, such as oral hypoglycemics, herbal medicines, and local anesthetics (Figure 2).

Figure 1.

Top 10 underlying diseases in drug-induced AKI group versus control group.

Figure 2.

Top 10 drug classes of use in the drug-induced AKI and control group.

Subsequently, we investigated the risk ratio (RR) between pharmacological agents and diseases. The analysis of RR identified kidney failure as the most prominent underlying disease risk factor for AKI (RR=6.56). Additional notable disease associations included parathyroid and hormone-related disorders (RR=2.28), kidney tubulo-interstitial diseases (RR=2.03), and anemia or erythrocyte disorders (RR=1.99). Regarding medications, treatments for hyperkalemia and hyperphosphatemia presented the highest risk for AKI (RR=2.66), with an incidence rate of 95.15% among users compared to 35.79% among non-users. Other high-risk drug categories comprised blood cell-raising agents (RR=2.38), diuretics and dehydrating drugs (RR=2.31), antigout medications (RR=2.20), and urinary-specific drugs (RR=2.05) (Figure 3).

Figure 3.

Disease variables and drug variables with top 10 risk ratios (RR). Due to automatic label placement, slight overlap persists. All data are legible at 1200 dpi. Risk ratio >1: elevated risk; risk ratio <1: reduced risk.

Feature Screening Results

The initial dataset consisted of one continuous variable (age) and 510 categorical variables, which included 192 drug-related and 318 disease-related features. The training and test sets comprised 1,672 and 717 samples, respectively, both maintaining a balanced proportion of 39.89% for AKI-positive cases. Figure 4 illustrates the heatmap of the top 30 features identified through univariate screening. In the process of constructing interaction features, an initial screening identified 65 high-frequency drug features and 47 significant disease features, which were subsequently used to generate 3,055 drug-disease, 65 drug-age, and 65 drug-gender interaction terms. The final dual screening process, employing mutual information and F-test methods, identified the top 30 interaction features ranked by their Mutual Information (MI) scores, as depicted in Figure 5. Ultimately, the final feature selection process yielded 40 optimal features, comprising 30 original features and 10 interaction features, for model construction.

Figure 4.

Heat map of the top 30 important features of the univariate screening composite.

Figure 5.

Top 30 interaction terms ranked by mutual information scores.

Model Construction and Evaluation

We evaluated seven machine learning models for predicting the risk of AKI in elderly patients. The models assessed included logistic regression, support vector machine (SVM), random forest, gradient boosting machine (GBM), XGBoost, LightGBM, and multilayer perceptron (MLP). As demonstrated in Table 2 and Figure 6, all models exhibited exceptional predictive performance, with AUC values on the training set ranging from 0.9425 to 0.9895, thereby indicating robust model fitting capabilities.

Table 2.

Comparison of Performance Metrics of Each Machine Learning Model

Model	Training Set AUC (±SD)	Validation Set AUC (±SD)	Overfitting Degree	Validation Set Sensitivity (±SD)	Validation Set Specificity (±SD)	Validation Set Precision (±SD)	Validation Set F1 (±SD)
Logistic Regression	0.9425 ± 0.0026	0.9383 ± 0.0086	0.0042	0.8305 ± 0.0244	0.9055 ± 0.0176	0.8892 ± 0.0168	0.8585 ± 0.0125
SVM	0.9573 ± 0.0161	0.9341 ± 0.0098	0.0232	0.8586 ± 0.0236	0.8936 ± 0.0206	0.8805 ± 0.0191	0.8691 ± 0.0132
Random Forest	0.9895 ± 0.0038	0.9489 ± 0.0124	0.0406	0.8706 ± 0.0242	0.9001 ± 0.0148	0.8881 ± 0.0145	0.8790 ± 0.0151
GBM	0.9766 ± 0.0052	0.9444 ± 0.0076	0.0322	0.8601 ± 0.0213	0.8933 ± 0.0168	0.8803 ± 0.0149	0.8698 ± 0.0104
XGBoost	0.9654 ± 0.0137	0.9402 ± 0.0103	0.0253	0.8803 ± 0.0290	0.8574 ± 0.0603	0.8528 ± 0.0477	0.8647 ± 0.0152
LightGBM	0.9750 ± 0.0063	0.9429 ± 0.0069	0.0321	0.8661 ± 0.0231	0.8906 ± 0.0171	0.8784 ± 0.0142	0.8719 ± 0.0093
MLP	0.9517 ± 0.0057	0.9398 ± 0.0082	0.0119	0.8422 ± 0.0144	0.8933 ± 0.0214	0.8783 ± 0.0200	0.8596 ± 0.0085

Abbreviations: AUC, Area Under the Curve; SVM, support vector machine; GBM, gradient boosting machine; MLP, multilayer perceptron.

Figure 6.

PR curves of the seven models. Due to automatic label placement, slight overlap persists. All data are legible at 1200 dpi.

After comprehensive evaluation of various machine learning models, the Random Forest model was identified as the most suitable for predicting AKI in elderly patients. Through hyperparameter tuning via cross-validation, the optimal parameters were established as follows: {“n_estimators”: 300, “min_samples_split”: 2, “min_samples_leaf”: 1, “max_features”: “log2”, “max_depth”: 15, “class_weight”: “balanced_subsample”}. The model exhibited exceptional performance on the independent test set, achieving an AUC of 0.9209 and a PR-AUC of 0.9103, significantly surpassing random chance levels. At the optimal threshold of 0.3400, as determined by Youden’s index, the model demonstrated balanced performance, with a sensitivity of 0.8494, specificity of 0.8912, and an F1-score of 0.8555 (Figure 7). Analysis of the confusion matrix (Figure 8) confirmed a high overall accuracy, while the calibration curve (Figure 9) indicated excellent calibration in low-probability ranges, although predictions were slightly conservative in medium-to-high probability ranges.

Figure 7.

ROC curve for Random Forest model.

Figure 8.

Confusion matrix diagram for the Random Forest model.

Figure 9.

Calibration curves for the Random Forest model.

Model Interpretability Analysis

As depicted in Figure 10, feature importance analysis of the Random Forest model revealed “kidney failure” as the most influential single variable, occupying a central predictive role (importance score=0.1703). Notably, interactions between various drug classes and kidney failure exhibited significant predictive value: the “antihypertensives_kidney failure” interaction ranked second (importance score=0.1000), followed by “anticoagulants, antiplatelet agents, thrombolytics_kidney failure” (importance score=0.0810) and “proprietary medicines_kidney failure” (importance score=0.0725). These findings highlight that the combined effect of antihypertensive medications and pre-existing kidney dysfunction serves as a particularly strong predictor. Age also demonstrated moderate predictive importance with importance score of 0.0699. Collectively, these results demonstrate that baseline renal function status and its complex interactions with specific medication classes - particularly antihypertensives, anticoagulant and antiplatelet agents - constitute critical determinants of AKI risk in elderly populations.

Figure 10.

Top 20 important features of the Random Forest model.

To elucidate the predictive mechanisms of the model and quantify the contributions of various features, we performed an analysis using SHapley Additive exPlanations (SHAP) values, as illustrated in Figure 11. This analysis revealed that critical features, such as kidney failure, consistently exhibited high positive SHAP values at elevated feature levels, indicating a strong positive association with the predicted risk of AKI in elderly patients. Significant interaction terms, such as antihypertensives_kidney failure, demonstrated similar patterns of risk enhancement. In contrast, features like arterial or arteriolar disease displayed SHAP values that ranged from positive to negative, suggesting context-dependent effects that vary according to sample characteristics and feature values. Additionally, variables including general anesthetic for intravenous use and urological specialty drugs exhibited SHAP values clustered around zero, indicating their minimal impact on the model’s predictions.

Figure 11.

SHAP summary plot for the Random Forest mode.

The SHAP dependence plot for the interaction between anti-hypertensives and kidney failure is presented in Figure 12. Specifically, lower feature values are associated with predominantly negative SHAP values, indicating an inverse association with AKI risk. In contrast, higher feature values correspond to increasingly positive SHAP values, suggesting a positive association with AKI risk. Furthermore, the coloration of the points represents the use of anti-gout medications. In the higher value range, red points (indicating anti-gout use) and blue points (indicating no anti-gout use) exhibit distinct SHAP value distributions. This underscores the importance of conducting a careful risk assessment when these drug classes are used in combination.

Figure 12.

SHAP dependence plot for the antihypertensives_kidney_failure interaction term in the Random Forest model.

Discussion

This study utilized EHR data from 2,389 elderly patients to systematically implemented data governance, feature engineering, and machine learning modeling. The research culminated in the development of seven distinct machine learning models, each incorporating 40 feature variables, to evaluate the risk of drug-induced AKI. Among these models, the random forest algorithm emerged as the most effective predictor, as evidenced by a comprehensive evaluation of its sensitivity (84.9%), specificity (89.1%), and F1 score, achieving an AUC of 0.9209 on independent testing. This performance underscores the model’s superior capability in identifying DI-AKI cases. Its strength is attributed to an exceptional ability to manage high-dimensional data while capturing complex nonlinear relationships and feature interactions, which is particularly advantageous for elderly patients with multimorbidity and polypharmacy profiles.20,21 This robust performance aligns with established literature; for instance, Chiofolo implemented random forest classification in ICU populations, achieving a ROC AUC of 0.88 with 92% sensitivity.22 Their model successfully predicted 30% of AKI cases at least six hours in advance and effectively stratified AKI stages 2–3, demonstrating similar utility for clinical risk stratification and the targeting of preventive interventions. Additionally, our SHAP value analysis improved the interpretability of the model by quantitatively elucidating the contributions of specific risk factors and their interactions to AKI prediction.

The principal strength of the Random Forest model resides in its proficiency in managing high-dimensional data, discerning intricate nonlinear relationships, and capturing interactions among features. This capability is of paramount importance for accurately modeling the risk of DI-AKI in elderly patients, who frequently present with complex clinical profiles characterized by comorbidities and polypharmacy. Nevertheless, the AUC of the training set (0.9895) in this study markedly surpasses that of the validation set (0.9489), suggesting a certain degree of overfitting. This phenomenon may be attributed to the original high-dimensional feature space (comprising 510 features) and the supplementary features generated subsequent to the construction of interaction terms. Despite the implementation of measures such as nested cross-validation, RFE for multi-faceted feature selection, and hyperparameter optimization, the risk of overfitting has not been entirely eradicated. In the future, this issue can be further mitigated by augmenting the sample size, incorporating more stringent regularization techniques (eg, increasing min_samples_leaf), or employing an ensemble pruning strategy Furthermore, this study solely utilized a single-center internal test set for validation purposes, without conducting external validation Consequently, the generalization capability of the model across diverse regions, hospitals of varying tiers, or among different ethnic populations remains uncertain. Prior to clinical deployment, it is imperative to undertake prospective, multi-center external validation.

Recent research has identified several critical risk factors for Drug-Induced Acute Kidney Injury in elderly populations, with pre-existing kidney failure emerging as the most significant predictor, evidenced by a feature importance score of 0.1703. This finding highlights the crucial role of baseline renal function in assessing the risk of AKI. Other major risk factors include: (1) specific drug class interactions with kidney failure, such as antihypertensives, anticoagulants/antiplatelet agents, and proprietary medications; (2) advanced age; (3) the use of antigout medications, diuretics, dehydrating agents, and hemoconcentrating drugs; (4) comorbidities such as heart failure and hypertension; and (5) the administration of particular medications, including insulin and its analogues, local anesthetics, and treatments for hyperkalemia and hyperphosphatemia.

This study offers novel contributions to the field of AKI risk prediction in elderly patients by systematically integrating and quantifying drug-disease interaction, a critical aspect that has been largely neglected in previous models. Notably, interaction terms accounted for 50% of the top 20 predictive features, underscoring their clinical significance.23 The interaction between antihypertensives and kidney failure serves as a prime example of this phenomenon. This study identifies a critical predictive factor for drug-induced AKI as the interaction between antihypertensive medications and pre-existing renal insufficiency, referred to as the “antihypertensive drugs_renal failure” interaction. Previous research has demonstrated that certain antihypertensive agents, particularly angiotensin-converting enzyme (ACE) inhibitors and angiotensin receptor blockers (ARBs), may be associated with an elevated risk of AKI in patients with compromised renal function by modifying renal hemodynamics. The SHAP plot confirms that a high interaction term is associated with a higher predicted AKI risk, therefore, clinicians should be cautious with such drugs in elderly renal patients, monitoring creatinine every 48 hours when adjusting treatment. Similarly, the interaction between anticoagulant and antiplatelet_kidney failure exhibited significant predictive capability. This illustrates that impaired renal clearance, as observed with agents such as warfarin or direct oral anticoagulants, increases the risks of both bleeding and nephrotoxicity.24 These findings highlight the critical importance of assessing medication risks in elderly patients through a dual perspective—taking into account not only the intrinsic nephrotoxicity of the drugs but also their dynamic interactions with the patient’s baseline renal function status. If avoidance is not feasible, consideration should be given to reducing the dosage or increasing the frequency of monitoring. For patients with markedly positive SHAP values (eg, SHAP > 005), the system can automatically identify principal risk factors (eg, renal failure_antihypertensives_anticoagulants) to support rapid clinical intervention.

The pathogenesis of DI-AKI in elderly patients transcends mere drug toxicity, involving intricate interactions between polypharmacy and multimorbidity that cannot be entirely replicated in vitro or through animal models.25 Our EHR-based predictive framework addresses this complexity by integrating nonlinear drug-disease and drug-drug interactions. It builds upon established methodologies, such as the recurrent neural network (RNN)-based model developed by Tomašev (achieving a ROC AUC of 92.1% and a sensitivity of 55.8%), while optimizing for routine clinical application.26 Our model is designed to align more closely with the complex pathways of clinical decision-making. In contrast to intensive care unit-specific models that necessitate extensive monitoring, such as the Mayo Clinic’s real-time system, our approach utilizes standard EHR data to identify high-risk medication patterns, with a particular focus on hazardous drug-disease interactions, thereby ensuring broad applicability across diverse clinical environments. The prediction results of the model also indirectly confirmed the core position of serum creatinine levels in the clinical assessment of kidney injury. And in the future, such models could be incorporated into hospital information systems to identify high-frequency yet high-risk medication patterns, particularly in recognizing complex drug-disease and drug-drug interaction risks. This integration would facilitate the automatic generation of medication reassessment alerts for high-risk patients, thereby supporting data-driven clinical adjustments to medication regimens.

This study has several limitations. First, the outcome was defined solely based on serum creatinine, which is a delayed marker of renal impairment. Future studies should incorporate earlier biomarkers, such as urine output, neutrophil gelatinase-associated lipocalin (NGAL), or kidney injury molecule-1 (KIM-1).27 Second, the model exhibits a certain risk of overfitting and lacks external validation, as it was developed and tested only on a single-center internal dataset. Third, the retrospective design may introduce selection bias and residual confounding, despite our efforts to adjust for relevant covariates. Fourth, detailed medication information—including drug dosage, treatment duration, and specific drug–drug interactions—was not available in the EHR data and therefore could not be included in the analysis.

Furthermore, our observational design cannot fully exclude confounding by indication. For example, patients receiving antihypertensives typically have hypertension, which is itself a well-established risk factor for AKI. Consequently, the observed associations may partly reflect the influence of underlying diseases rather than independent drug effects. In a related vein, “kidney failure” emerged as the strongest predictor in our model (feature importance 0.1703), suggesting that the model may largely identify patients with pre-existing renal impairment rather than specifically predicting drug-induced AKI. Given these considerations, our model should be regarded as a risk identification tool—intended to flag high-risk patients for closer monitoring—rather than a causal attribution tool for determining whether AKI is caused by a specific drug. Future studies employing stricter designs, such as self-controlled case series or instrumental variable analysis, are needed to further disentangle drug effects from disease effects and to validate the clinical utility of the model.

To address these limitations and facilitate clinical translation: the model can be integrated into EHR systems via HL7/FHIR interfaces, deployed as a RESTful API for real-time risk scoring, with automated CDS alerts for high-risk patients (probability > 0.34) and a closed-loop feedback mechanism for periodic updating. Future prospective intervention trials based on model alerts are needed to validate its effectiveness in reducing DIKI incidence and improving patient outcomes.

Funding Statement

This work was supported by Sanming Project of Medicine in Shenzhen (SZSM202301035) and Open Projects Fund of NMPA Key Laboratory for Technology Research and Evaluation of Pharmacovigilance (NO.2025YYJJKF03).

Ethics Approval and Informed Consent

In accordance with the ethical guidelines and standards outlined in the Declaration of Helsinki, we hereby confirm that our study fully complies with these principles. This study has been approved by the Ethics Review Committee of Luohu District People’s Hospital, with the approval number 2025-YKYSC-040. All participants were informed and signed the consent form.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.