Interpretable AI Model Predicts 28-Day Mortality in Sepsis-Linked Kidney Injury

Researchers have developed and validated an interpretable machine learning model that predicts 28-day mortality in critically ill patients with sepsis-associated acute kidney injury (SA-AKI), reporting stronger discrimination than eight conventional clinical scoring indices. The work, published in BMC Medical Informatics and Decision Making, addresses a recognized gap in prognostic tools for a complication the authors describe as carrying a high 28-day mortality rate, particularly in its early or mild stages.

The study was a retrospective cohort analysis of 1,823 adult SA-AKI patients treated in a single intensive care unit. The team built a baseline model from 40 features spanning demographics, comorbidities, vital signs, laboratory indices and organ support status. Five machine learning algorithms were trained and then evaluated on an independent test set, with performance measured by the area under the receiver operating characteristic curve (AUC).

An XGBoost model emerged as the top performer for predicting 28-day mortality, achieving an AUC of 0.881 (95% CI: 0.840–0.916) in the test cohort. According to the reported results, this significantly outperformed all traditional clinical indices against which it was compared. The model also maintained robust predictive accuracy across key clinical subgroups, with particularly strong performance among patients with milder disease, a group where existing prognostic tools are described as lacking accuracy.

To make the model’s internal logic transparent, the researchers applied SHAP (SHapley Additive exPlanations) analysis. This identified lactic acid, FiO2, the Sequential Organ Failure Assessment (SOFA) score and the Glasgow Coma Scale (GCS) as the most influential predictors of short-term death.

The authors conclude that the XGBoost model offers a highly accurate and interpretable tool for early prediction of 28-day mortality in SA-AKI patients, outperforming conventional scores. They suggest that its ability to flag high-risk patients even in milder cases could support timely interventions and improve resource allocation.

Subject of Research: Machine learning prediction of short-term mortality in sepsis-associated acute kidney injury

Article Title: Development and validation of an interpretable machine learning model for predicting short-term mortality in critically ill patients with sepsis-associated acute kidney injury

Article References: Cai, S., Lin, J., Liu, X., & Chen, X. (2026). Development and validation of an interpretable machine learning model for predicting short-term mortality in critically ill patients with sepsis-associated acute kidney injury. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03879-2

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03879-2

Keywords: sepsis-associated acute kidney injury, machine learning, XGBoost, 28-day mortality, SHAP, intensive care, prediction model, SOFA score, Glasgow Coma Scale, lactic acid, retrospective cohort, clinical decision support