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

Delayed identification of acute kidney injury (AKI) limits timely intervention, as diagnosis often depends on serum creatinine elevation after renal damage has occurred. This study aims to develop a machine learning model capable of identifying early AKI risk at the time of hospital admission using routinely collected biochemical and physiological parameters and their prediagnostic trends.

Clinical, biochemical, and physiological parameters were retrospectively extracted from the hospital database and discharge summaries of patients. For each patient's lab values, rate-of-change (RoC) values were calculated to capture dynamic trends preceding AKI onset. XGBoost models were trained using admission features and in combination with RoC values. Model performance was evaluated using accuracy, precision, recall, F1 score, and area under the curve (AUC), while Shapley Additive Explanations (SHAP) provided interpretable insights into feature importance.

The XGBoost model achieved robust performance (AUC = 0.83; recall = 0.75) in predicting AKI risk using data available at admission. Incorporating RoC features further improved discrimination (AUC = 0.90), indicating that subtle biochemical trends enhance early detection. Key predictors included values at admission and RoC values of blood urea nitrogen, calcium, albumin-globulin ratio, lactate, alanine aminotransferase, and pulse/min, reflecting early renal, metabolic, and hemodynamic alterations preceding AKI onset.

Machine learning models leveraging admission and temporal features can predict AKI risk with high accuracy using routinely collected clinical data. Future integration into electronic health records could enable real-time risk stratification and timely clinical intervention to improve patient outcomes.