MACHINE learning was used to predict moderate to severe acute kidney injury (AKI) in patients with acute pancreatitis admitted to intensive care, with an XGBoost model achieving area under the curve values of 0.850 in internal validation and 0.702 in external validation.
The study developed and externally validated an interpretable machine learning model to support early risk stratification for moderate to severe AKI among patients with acute pancreatitis in the ICU. The model used clinical information available within 24 hours of ICU admission, aiming to identify patients at increased risk during their ICU stay.
Data from 801 patients with acute pancreatitis in MIMIC-IV were used for model development and internal validation. A further 368 patients from eICU formed the external validation cohort. Ten routinely available variables were included, covering vital signs, comorbidities and early diagnostic conditions, laboratory findings, and early treatment requirements.
External Validation Shows Lower Performance
Across the algorithms assessed, differences in performance were modest, and no single approach demonstrated consistent superiority. XGBoost was selected as the primary model for detailed interpretation.
The model achieved an area under the curve of 0.850 in the internal validation cohort and 0.702 in the external validation cohort. The lower performance observed during external validation highlighted the need to assess how well the model generalises beyond the data used for development.
The researchers also used Shapley additive explanations to provide transparent, individualised estimates of acute kidney injury risk. This approach was intended to make the model’s risk assessments <a href="https://myappsplus.com/the-fix-for-rogue-ai-agents-could-be-more-ai/” title=”The fix for rogue AI agents could be more AI”>more interpretable for clinical use.
Prospective Validation Remains Necessary
The findings suggest that machine learning could provide an interpretable summary of early ICU acute kidney injury risk using routinely available clinical information. Early identification of patients at risk of moderate to severe AKI may help inform clinical attention, although the study did not establish that using the model would improve patient outcomes.
The researchers emphasised that the model’s temporal predictive value and clinical actionability still require prospective validation. Further assessment will therefore be needed before machine learning-based acute kidney injury risk estimates can be considered for routine clinical decision-making.
Reference
An R et al. Machine learning to predict moderate-to-severe AKI in ICU acute pancreatitis patients. Sci Rep. 2026;DOI:10.1038/s41598-026-67437-6.
Featured Image: Dragan on Adobe Stock
