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

Acute kidney injury (AKI) in Intensive Care Units (ICU) poses a significant threat to patient health, with potential lasting damage and high mortality risks associated with delayed diagnosis. Serum creatinine (SCr), a key indicator of renal function, is a crucial tool for timely prognosis and treatment. However, SCr values are influenced by numerous complex biological variables that are measured at irregular sparse intervals due to the nature of ICU patient monitoring. This makes algorithmic SCr forecasting particularly challenging.In this paper, we present an embedding-based interfaced Transformer neural network that effectively handles the complexities of clinical multivariate irregular time series data found in ICU. Our approach demonstrates exceptional performance on SCr forecasting against nine competing models, on two extensive datasets, and across four feature selection scenarios, achieving a median absolute error of 0.043 mg/dL in the most common setting as the first ever model to predict raw SCr values.To adequately support clinical decision-making, we also emphasize prediction explainability through token-level and feature-level contributions. By providing accurate SCr forecasts, our research has the potential to improve patient outcomes and reduce the risk of AKI-related mortality in ICU settings.Clinical relevance- This provides an accurate prediction of serum creatinine levels which is crucial in ICU where AKI can significantly impact patient outcomes.