In-Hospital Cardiovascular Mortality Risk Prediction in Pará, Brazil: A Comparative Analysis of XGBoost and Logistic Regression Using Administrative Hospital Data

Abstract

This study develops and compares XGBoost and logistic regression models for predicting in-hospital cardiovascular mortality from SIH-SUS administrative hospital records in Pará, Brazil (2019–2023). XGBoost outperformed logistic regression (AUC 0.784 vs. 0.674, primary cohort; 0.795 vs. 0.736, sensitivity cohort). Most predictive signal is recorded only at or near discharge, limiting real-time clinical utility. Municipal care-access disparities identified in a companion geographic study persisted as a patient-level predictor after clinical adjustment. Analysis code available at: github.com/g4bfernandoo/cardiac-mortality-prediction