Predicting Mortality in ICU Patient with Hypertension Using Machine Learning on EHR Data

Rishabh Goel, Andrew B. Gumbert, Stephanie Chen

Keywords: machine learning, time series, mortality prediction, ICU data, hypertension, ensemble learning, MIMIC-IV

TL;DR: We develop a time-aware ensemble model that predicts ICU mortality in hypertensive patients using only early EHR data, achieving strong performance across multiple time horizons with minimal benefit from later observations.

Abstract: Hypertension is a leading preventable cause of mortality, with over 120 million patients affected in the US alone. Yet current predictive models are often limited to specific subtypes of the disease and only predict short-term outcomes. This study introduces a unified, time-aware, ensemble for mortality risk prediction that generalizes across all hypertension subtypes and multiple horizons while addressing selection effects from longer observation windows. Using MIMIC-IV EHR date, we trained models at 24, 48, and 72 hours and combined their outputs with learned convex weights. Evaluated with stratified 5-fold cross-validation for in-hospital, 30-, 60-, 90-day, and 1-year mortality, the ensemble achieved AUCs of 0.854, 0.847, 0.841, and 0.801, with F1 > 0.50 for all horizons except in-hospital. Learned temporal weights placed minimal weight on later windows, indicating limited incremental signal once selection bias is controlled.

Submission Number: 34