Prospective and external evaluation of an AI model for continuous and early prediction of moderate and severe AKI in critically ill patients

Background: Acute kidney injury (AKI) is a major complication in critically ill patients, burdening both patients and healthcare systems. We previously introduced an AI-based model for early and continuous prediction of ICU-acquired AKI (ICU-A-AKI-2/3). In this study, we enhanced the model to better handle missing data, a common challenge in clinical settings. The upgraded model was validated in both retrospective and prospective cohorts, demonstrating improved robustness and predictive performance. Methods: The model was validated in retrospective cohorts from three countries (US, Netherlands, Italy; N = 70,107 ICU admissions) across 176 ICUs. It was then prospectively validated in three European hospitals (Italy, Spain; N = 329) from May to October 2023. Using an XGBoost classifier, the model analyzes clinical data from ICU patients to predict hourly risk probabilities for AKI stages 2 and 3, as defined by KDIGO. Results: In retrospective cohorts, the AI model achieved an auROC greater than 0.89 for early detection of ICU-A-AKI-2/3. Prospective validation showed auROCs between 0.82 (95% CI 0.73–0.92) and 0.96 (95% CI 0.92–0.99) across hospitals, with a mean lead time of approximately 14 h. Conclusions: This enhanced AI model offers timely prediction of ICU-A-AKI-2/3 episodes, as demonstrated across diverse cohorts. Its high predictive performance represents a significant advancement in integrating AI into clinical workflows, enhancing AKI management and improving clinical outcomes in ICU settings.