Development of an Interpretable Early AKI Prediction Model in ICU Using the LightGBM

The study predicts acute kidney injury (AKI) episodes using clinical data from the first day of ICU admission in the MIMIC-IV intensive care medical information database. Eight machine learning models—Random Forest, Logistic Regression, LightGBM, XGBoost, Decision Tree, Gaussian Naïve Bayes, K-Nearest Neighbors (KNN), and Support Vector Machine (SVM)—were used for cross-valuation using eight different feature datasets; the occurrence of AKI serves as the outcome variable. The LightGBM model in conjunction with the XGBset dataset shows the best performance according to the outcomes. Incorporating 19 clinical characteristics, the LightGBM model was finally chosen to provide an early risk prediction model for AKI occurrences. With an AUC of 0.97 and an accuracy of 0.88 the model performed. Moreover, SHapley Additive Explanations (SHAP) were proposed to improve the interpretability of the machine learning model, so increasing the transparency and clinical relevance of the prediction outcomes. Finally, this work shows that a predictive model with just 19 characteristics can still show good predictive accuracy, therefore offering vital support for individualized intervention and early AKI prediction. The application of interpretability methods provides insightful analysis for clinical decision-making, therefore guiding medical practitioners in their informed decisions.