[Artificial intelligence and acute kidney injury].

Area of Science:

• Artificial intelligence (AI) applications in critical care medicine.
• Focus on predictive and classification models for acute kidney injury (AKI).

Context:

• Digitalization is transforming intensive care units (ICUs).
• AI is being explored for predicting and phenotyping AKI in critically ill patients.
• Current AI models primarily use serum creatinine and urinary output, with known limitations.

Purpose:

• To evaluate the current state and challenges of AI in AKI prediction and classification.
• To assess the performance of AI models using metrics like AUROC.
• To identify AI-specific shortcomings hindering clinical integration.

Summary:

• AI models for AKI prediction show variable performance (AUROC 0.650-0.900), influenced by prediction time and AKI criteria (KDIGO vs. AKIN).
• Phenotyping by AI aids risk stratification for mortality and RRT but lacks etiological and therapeutic insights.
• Limitations include inability to incorporate recent therapeutic changes/biomarkers and lack of model transparency, hindering physician trust.

Impact:

• Successful AI integration in ICUs hinges on overcoming data limitations and enhancing model interpretability.
• Physician trust is crucial for adopting AI-driven alerts for AKI.
• Clinicians remain essential for patient management, integrating AI insights with clinical judgment.