Abstract: FR-PO0051
Enhanced Artificial Intelligence-Based Risk Score Using XGBoost for Predicting In-Hospital AKI in a Bolivian Cohort
Session Information
- AKI: Epidemiology, Risk Factors, and Prevention
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Acute Kidney Injury
- 101 AKI: Epidemiology, Risk Factors, and Prevention
Authors
- Claure-Del Granado, Rolando, Division of Nephrology, Hospital Obrero No 2 - Caja Nacional de Salud, Cochabamba, Bolivia, Plurinational State of
- Moya-Mamani, Juan C., Universidad Mayor de San Simon Facultad de Medicina, Cochabamba, Bolivia, Plurinational State of
- Malhotra, Rakesh, University of California San Diego Department of Medicine, La Jolla, California, United States
Background
Artificial intelligence (AI) is increasingly used to predict hospital-acquired acute kidney injury (AKI). We evaluated and optimized an existing AI-generated AKI risk score (J Am Soc Nephrol 35, 2024: FR-PO036) using XGBoost, a non-linear machine learning method, in a prospective cohort.
Methods
A prospective cohort of 350 adults hospitalized without AKI was enrolled at Hospital Obrero No 2 in Bolivia. The original AI risk score, previously validated in the same setting, was based on 13 clinical variables and developed using logistic regression with expert-defined weights, producing a cumulative score from 0 to 14.
For optimization, the same 13 predictors were used to train a non-linear machine learning model using XGBoost (Figure 1) to estimate the probability of developing AKI within 7 days, as defined by KDIGO criteria. Model performance was evaluated using area under the ROC curve (AUC-ROC), sensitivity, specificity, positive and negative predictive values, and odds ratio
Results
A total of 350 adult patients hospitalized without AKI were prospectively enrolled. Mean age was 67.7 ± 17.5 years, with 53% female. AKI incidence was 33% (n=117). KRT was initiated in 2%, and in-hospital mortality was 0.5%. The original AI-based score had an AUC of 0.74 (95% CI: 0.64–0.77), with 90% sensitivity and 35% specificity. The optimized XGBoost model achieved a significantly higher AUC of 0.99 (95% CI: 0.982–0.996). At a cutoff of ≥0.68, sensitivity was 92% and specificity 96%. The odds ratio associated with the score was 729.3, indicating strong discriminative power (Table). Key predictors included age group, eGFR, hypotension, and cancer.
Conclusion
XGBoost-based modeling greatly enhanced the predictive accuracy of the AKI risk score. This optimized score could support early identification and intervention for at-risk hospitalized patients. Further external validation is warranted.
Components of the XGBoost AKI Risk Score
Performance of Original vs. XGBoost Models