ATBiGRU: an interpretable temporal model for in-hospital mortality prediction in acute kidney injury

Acute kidney injury (AKI) is associated with high mortality and healthcare burden, particularly in developing countries. Existing machine learning models mainly rely on static data and cannot effectively capture dynamic clinical features. We propose ATBiGRU, an explainable deep learning model integrating attention mechanisms, Time Convolutional Networks, and Bidirectional Gated Recurrent Units to predict in-hospital mortality risk in AKI patients using the MIMIC III dataset. SHAP was applied for feature interpretation. ATBiGRU achieved an AUROC of 95.65%, outperforming XGBoost (90.37%) and LSTM (90.16%). SHAP analysis identified advanced age and low serum creatinine as major mortality predictors.