Interpretable machine learning model for early prediction of 30-day mortality in ICU patients with coexisting hypertension and atrial fibrillation: A retrospective cohort study

Hypertension and atrial fibrillation (AF) frequently coexist in critically ill patients and jointly increase the risk of adverse cardiovascular outcomes. Despite their clinical importance, short-term mortality prediction for ICU patients with concurrent hypertension and AF remains underexplored. Accurate early identification of high-risk individuals could support timely intervention and optimize critical care resource allocation. We performed a retrospective analysis of 1301 adult ICU patients with both hypertension and AF from the MIMIC-IV database. Structured data from the first 24 h after ICU admission–including chart events, laboratory tests, procedures, medications, demographics, and comorbidities–were extracted and preprocessed. After data cleaning, imputation, and expert-guided feature selection, 17 clinically interpretable variables were retained. The cohort was stratified into training (70%) and test (30%) sets. Outcome-weighted training and stratified five-fold cross-validation were applied to address class imbalance and optimize generalization. CatBoost and five baseline machine learning models (LightGBM, XGBoost, logistic regression, Naive Bayes, and neural networks) were benchmarked using AUROC as the primary performance metric. Model interpretability was assessed through SHapley Additive exPlanations (SHAP), Accumulated Local Effects (ALE), and DREAM analyses. CatBoost achieved the highest predictive performance with an AUROC of 0.889 (95% CI: 0.840–0.924), accuracy of 0.831, F1-score of 0.522, sensitivity of 0.837, specificity of 0.830, PPV of 0.379, and NPV of 0.976. Global and local interpretability analyses consistently identified Richmond-RAS Scale, pO 2 , CefePIME administration, and Invasive Ventilation as key predictors, reflecting clinically plausible risk factors for mortality. This study presents the first interpretable machine learning framework for early prediction of 30-day mortality in ICU patients with coexisting hypertension and AF. The model demonstrates robust discriminative ability and clinical interpretability, providing a practical tool for precision decision support in critical care. Future work will focus on external validation and longitudinal extension across multi-center ICU datasets. • First ML model for short-term ICU mortality in hypertension and AF. • Interpretable framework built with 17 expert-selected variables. • CatBoost outperformed 5 models (AUROC = 0.889) in mortality prediction. • SHAP, ALE, and DREAM revealed key clinically relevant predictors.