Multivariable Serum Creatinine Forecasting for Acute Kidney Injury Detection Using an Explainable Transformer-based...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Critical Care Nephrology

Background:

• Acute kidney injury (AKI) in Intensive Care Units (ICUs) presents significant mortality risks, often exacerbated by delayed diagnosis.
• Serum creatinine (SCr) is a vital biomarker for renal function, but its prediction is challenging due to irregular, sparse clinical data.
• Effective AKI management relies on timely and accurate SCr level monitoring.

Purpose of the Study:

• To develop an advanced AI model for accurate serum creatinine (SCr) forecasting in ICU patients.
• To address the challenges of irregular and multivariate time series data in clinical settings.
• To enhance clinical decision-making for acute kidney injury (AKI) management.

Main Methods:

• An embedding-based interfaced Transformer neural network was designed to process complex clinical multivariate irregular time series data.
• The model was evaluated on two extensive datasets across four feature selection scenarios.
• Prediction explainability was incorporated using token-level and feature-level contributions.

Main Results:

• The proposed Transformer network achieved superior performance in SCr forecasting compared to nine other models.
• The model demonstrated a median absolute error of 0.043 mg/dL in the most common setting.
• This represents the first model capable of predicting raw SCr values directly.

Conclusions:

• The developed AI model offers accurate SCr predictions, crucial for timely AKI diagnosis and management in ICUs.
• Improved patient outcomes and reduced AKI-related mortality are potential benefits of this predictive approach.
• Explainable AI predictions support clinical decision-making in critical care settings.