An XGBoost model using seven predictors with SHAP interpretation accurately predicted in-hospital mortality in ICU patients with CHF and T2DM (AUC 0.917), offering an explainable tool for early risk stratification.
A retrospective cohort of 573 ICU patients with congestive heart failure and type 2 diabetes mellitus from three hospitals in Henan, China was used to build the mortality prediction model.
Stepwise logistic regression identified seven predictors of in-hospital mortality for model construction.
XGBoost performed best among the five machine learning algorithms compared.
SHAP analysis identified mild liver disease, LODS score, BMI, WBC count, and SOFA score as the top contributors to XGBoost predictions.
Myocardial infarction, liver disease, elevated WBC count, and higher SOFA/LODS scores were associated with increased mortality risk.
Higher BMI and higher serum creatinine were associated with lower predicted mortality risk.
The authors conclude that the XGBoost-SHAP model provides accurate, explainable mortality prediction that could aid early risk stratification and clinical decision-making.
This study built a computer model to predict which ICU patients with both heart failure and type 2 diabetes are most likely to die in the hospital. Researchers used records from 573 patients at three hospitals in Henan, China, and chose seven routinely available factors: body mass index, prior heart attack, mild liver disease, two severity-of-illness scores (SOFA and LODS), white blood cell count, and serum creatinine. They compared five machine learning methods, and a method called XGBoost performed best, with an AUC of 0.917, indicating strong ability to separate patients who died from those who survived. The team also used a technique called SHAP to show how each factor influenced the predictions. Mild liver disease, LODS score, BMI, white blood cell count, and SOFA score mattered most. Heart attack history, liver disease, high white blood cell count, and higher illness-severity scores pushed predicted risk up, while higher BMI and higher creatinine pushed it down. The authors interpret this as consistent with the 'obesity paradox' and with muscle-loss (sarcopenia) ideas. This research suggests that explainable machine learning could help clinicians identify high-risk patients earlier. However, the study was retrospective, relatively small, and drawn from one region, and the reported accuracy (74.8%) was notably lower than the AUC, sensitivity, and specificity. Further validation in other settings would be needed before such a tool could be relied on in practice.
Hu J, Su S, Liu D. (2026). A machine learning-driven and explainable prediction model for in-hospital mortality in patients with diabetes mellitus and congestive heart failure: integrating XGBoost and SHAP approaches.. Frontiers in public health. https://doi.org/10.3389/fpubh.2026.1951635