Machine Learning-Based Mortality Prediction in Critically Ill Patients with Hypertension: Comparative Analysis, Fairness, and Interpretability

Shenghan Zhang¹Sirui Ding²Zidu Xu³Jiancheng Ye^4,5*

• ¹Harvard University, Cambridge, United States
• ²University of California San Francisco, San Francisco, United States
• ³Columbia University, New York, United States
• ⁴Northwestern University, Evanston, United States
• ⁵Weill Cornell Medicine, New York, United States

ABSTRACT Background: Hypertension is a leading global health concern, significantly contributing to cardiovascular, cerebrovascular, and renal diseases. In critically ill patients, hypertension poses increased risks of complications and mortality. Early and accurate mortality prediction in this population is essential for timely intervention and improved outcomes. Machine learning (ML) and deep learning (DL) approaches offer promising solutions by leveraging high-dimensional electronic health record (EHR) data. Objective: To develop and evaluate ML and DL models for predicting in-hospital mortality in hypertensive patients using the MIMIC-IV critical care dataset, and to assess the fairness and interpretability of the models. Methods: We developed four ML models—gradient boosting machine (GBM), logistic regression, support vector machine (SVM), and random forest—and two DL models— multilayer perceptron (MLP) and long short-term memory (LSTM). A comprehensive set of features, including demographics, lab values, vital signs, comorbidities, and ICU-specific variables, were extracted or engineered. Models were trained using 5-fold cross-validation and evaluated on a separate test set. Feature importance was analyzed using SHapley Additive exPlanations (SHAP) values, and fairness was assessed using demographic parity difference (DPD) and equalized odds difference (EOD), with and without the application of debiasing techniques. Results: The GBM model outperformed all other models, with an AUC-ROC score of 96.3%, accuracy of 89.4%, sensitivity of 87.8%, specificity of 90.7%, and F1 score of 89.2%. Key features contributing to mortality prediction included Glasgow Coma Scale (GCS) scores, Braden Scale scores, blood urea nitrogen, age, red cell distribution width (RDW), bicarbonate, and lactate levels. Fairness analysis revealed that models trained on the top 30 most important features demonstrated lower DPD and EOD, suggesting reduced bias. Debiasing methods improved fairness in models trained with all features but had limited effects on models using the top 30 features. Conclusions: ML models show strong potential for mortality prediction in critically ill hypertensive patients. Feature selection not only enhances interpretability and reduces computational complexity but may also contribute to improved model fairness. These findings support the integration of interpretable and equitable AI tools in critical care settings to assist with clinical decision-making.