Abstract
BACKGROUND: Acute kidney injury (AKI) in patients with liver cirrhosis represents a significant clinical challenge with high mortality rates. This study aimed to develop and validate a machine learning-based prediction model for 28-day mortality in AKI patients with liver cirrhosis using the MIMIC-IV database. METHODS: This retrospective study analyzed data from 4,168 AKI patients, including 601 with concurrent liver cirrhosis, from the MIMIC-IV database. Patient selection followed strict inclusion and exclusion criteria. The study implemented comprehensive data preprocessing, including feature normalization and selection through Recursive Feature Elimination. Multiple machine learning algorithms were evaluated, with model performance assessed through ROC curves, calibration curves, and precision-recall analysis. SHAP analysis was conducted to interpret feature contributions to mortality prediction. RESULTS: The liver cirrhosis group demonstrated distinct clinical characteristics, including significantly lower age (median 60 vs 70 years, p < 0.001) and higher disease severity scores (SOFA 11 vs 8 points) compared to non-cirrhotic patients. Survival analysis confirmed significantly lower 28-day survival probability in the cirrhosis group (Log-rank test, χ2 = 46.5, p < 0.001). The Random Forest model achieved optimal performance with an AUC of 0.85 and precision-recall area of 0.81. SHAP analysis identified pH, anion gap, and total CO2 as the most significant predictive factors, with notable interaction effects among these indicators. CONCLUSION: This study successfully developed a machine learning model for predicting 28-day mortality in AKI patients with liver cirrhosis. The model demonstrated superior clinical decision-making value compared to traditional scoring systems, particularly in moderate-risk threshold intervals. The findings emphasize the crucial role of acid-base balance indicators in mortality risk assessment, providing valuable insights for clinical intervention strategies.