Leveraging an Explainable Machine Learning Model for Early Identification of Acute Kidney Injury: A Retrospective...

Area of Science:

• Biomedical informatics
• Machine learning in healthcare
• Renal medicine

Background:

• Acute kidney injury (AKI) diagnosis is often delayed, relying on creatinine increases after damage.
• Timely intervention for AKI is crucial but hindered by late identification.
• Current diagnostic methods for AKI lack early predictive capabilities.

Purpose of the Study:

• To develop a machine learning model for early AKI risk prediction at hospital admission.
• To utilize routinely collected biochemical and physiological parameters and their trends.
• To enable timely intervention by identifying high-risk patients upon entry.

Main Methods:

• Retrospective extraction of clinical, biochemical, and physiological data.
• Calculation of rate-of-change (RoC) values for lab parameters to capture trends.
• XGBoost model training with admission features and RoC values, evaluated using AUC, accuracy, precision, recall, and F1 score.

Main Results:

• XGBoost model achieved AUC of 0.83 using admission data.
• Incorporating RoC features improved predictive performance to AUC 0.90.
• Key predictors included admission values and RoC of blood urea nitrogen, calcium, albumin-globulin ratio, lactate, alanine aminotransferase, and pulse.

Conclusions:

• Machine learning models effectively predict AKI risk using admission and temporal data.
• Routine clinical data combined with temporal trends enhance predictive accuracy.
• Integration into EHRs can facilitate real-time risk stratification and improve patient outcomes.