ReAdmit: Predicting Early Unplanned ICU Readmission Using Radiology Notes and Structured Data.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Predictive Analytics

Background:

• Unplanned intensive care unit (ICU) readmissions negatively impact patient outcomes and reflect hospital care quality.
• Accurate prediction of readmission risk is essential for optimizing ICU discharge decisions and patient management.

Purpose of the Study:

• To develop a predictive model for ICU readmission within 72 hours post-discharge.
• Utilize structured data and radiology notes from before ICU discharge to support discharge decisions and transitional care for high-risk patients.

Main Methods:

• Retrospective study using MIMIC-IV database data from 30,714 ICU patients.
• Developed the ReAdmit model using structured data and radiology notes processed by Clinical BERT and XGBoost.
• Model calibrated with Isotonic Regression and validated using AUROC and standardized readmission rate.

Main Results:

• The 72-hour readmission rate was 5.97%.
• The ReAdmit model achieved an AUROC of 0.783, outperforming models using only structured data (AUROC=0.625) and the SWIFT score.
• Subgroup analysis showed median AUROC exceeding 0.77, with a standardized readmission rate of 1.103.

Conclusions:

• The ReAdmit model accurately predicts 72-hour ICU readmission risk using routine clinical features and radiology notes.
• Enables early identification of high-risk patients, supporting safer transfer decisions and guiding post-discharge care to reduce preventable readmissions.