Early Prediction of Vasopressor Initiation in ICU Patients Using Machine Learning on the MIMIC-IV Dataset

This study develops machine learning models to predict vasopressorinitiation within 24 hours of ICU admission using early physiologicand laboratory measurements from the MIMIC-IV database. Using 73,834 ICU stays, logistic regression and LightGBM models weretrained using features aggregated from the first six hours of ICUadmission. Model performance was evaluated using ROC-AUC, PR-AUC,bootstrap confidence intervals, calibration analysis, and decisioncurve analysis. The LightGBM model achieved a ROC-AUC of 0.835 and PR-AUC of 0.701,outperforming logistic regression while demonstrating improvedprobability calibration and clinical utility. These findings demonstrate that early physiologic signals capturedwithin routinely collected EHR data can support early detection ofhemodynamic deterioration in critically ill patients. References are included in the attached manuscript.