Photo Credit: Marco Marca

The following is a summary of “Construction of a machine learning-based interpretable prediction model for acute kidney injury in hospitalized patients,” published in the March 2025 issue of Scientific Reports by Yu et al. 

Researchers conducted a retrospective study using data from 59,936 hospitalized adults to develop a predictive model. 

They constructed models using 53 variables, all achieving acceptable performance in the validation cohort, with extreme gradient boosting (XGBoost) showing the best efficacy and stability (area under the curve (AUC), 0.9301). They also built simpler models with 39 significant variables screened by random forest recursive feature elimination, where XGBoost again performed best (AUC, 0.9357). 

The results showed significant net returns for all models, with XGBoost achieving optimal results. Shapley additive explanation (SHAP) importance matrices identified uric acid, colloidal solution, first creatinine on admission, pulse, and albumin as the top 5 variables. With the external validation cohort (4,022 patients), performance declined. Support vector machine (SVM) showed the best efficacy (AUC, 0.8230 and 0.8329), followed by XGBoost (0.8124 and 0.8316). 

Investigators predicted the risk of acute kidney injury (AKI) up to 48 hours in advance, allowing clinicians to assess risk more accurately and develop necessary management strategies. 

Source: nature.com/articles/s41598-025-90459-5