대한신장학회

논문분류	춘계학술대회 초록집
제목	Machine Learning Prediction of Drug-Induced Acute Kidney Injury Across Inpatient and Outpatient Settings Using Real-World Electronic Health Records
저자	Sang Hun Eum
출판정보	2026; 2026(1):
키워드	acute kidney injury, drug-induced nephrotoxicity, electronic health records, machine learning, risk prediction
초록	Objectives: Drug-induced acute kidney injury (DI-AKI) is a common and potentially preventable cause of acute kidney injury (AKI). Existing predictive models frequently focus on specific inpatient populations, leaving a gap in predicting DI-AKI across generalized inpatient and outpatient settings where polypharmacy is prevalent. We aimed to develop and validate prediction models using real-world electronic health record (EHR) data to preemptively stratify DI-AKI risk associated with nephrotoxic medications across both settings. Methods: We conducted a retrospective cohort study using the multicenter EHR database (2 academic hospitals in South Korea, 2006–2022). The cohort included adults receiving nephrotoxic medications including antibiotics, nonsteroidal anti-inflammatory drugs, diuretics, renin-angiotensin system inhibitors, and sodium glucose cotransporter-2 inhibitors. We evaluated four machine learning algorithms (Histogram-based Gradient Boosting [HGB], XGBoost, Random Forest, and Logistic Regression) for risk prediction. Furthermore, a generalized linear mixed-effects logistic regression was utilized to identify independent clinical predictors while accounting for within-patient clustering. Results: The analytic cohort comprised 82,284 patients and 325,602 medication administration events, with an overall AKI incidence of 14.1%. Gradient-boosting models yielded the highest discriminative performance, with HGB achieving an area under the receiver operating characteristic curve of 0.827 and an area under the precision-recall curve of 0.477. Mixed-effects analysis identified medication exposure duration and baseline kidney function as primary independent predictors of DI-AKI. Applications of predefined alert threshold in the test set achieved 90.1% sensitivity. A three-tier traffic-light risk stratification system successfully categorized 49.3% of all AKI events in the high-risk (Red) tier (46.6% observed incidence), compared to a 2.9% incidence in the low-risk (Green) tier. Conclusion: This EHR-based DI-AKI prediction model can provide clinically actionable risk stratification in both inpatient and outpatient settings. By identifying modifiable risk factors and translating probabilistic risk into an interpretable traffic-light framework, this approach can facilitate targeted monitoring and optimize nephrotoxic medication stewardship.
원문(PDF)	PDF 원문보기

간행물 검색

대한신장학회 회원관리 규정