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

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.

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.

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.

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).

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.