Multicenter validation of a machine learning model to predict intensive care unit readmission within 48 hours after discharge.

Summary

The iREAD ensemble model using 30 discharge-time features predicted ICU readmission within 48 hours with AUROC 0.820 internally and 0.768/0.725 on two external cohorts, outperforming traditional scores and conventional ML baselines. High-risk patients identified by iREAD had over four-fold higher readmission rates in Kaplan–Meier analyses.

Key Findings

• Internal AUROCs: 0.771 (≤48 h), 0.834 (>48 h), 0.820 (overall).
• External AUROCs: 0.768 (MIMIC-III overall) and 0.725 (eICU-CRD overall), outperforming traditional scores (all P < 0.001).
• Kaplan–Meier: >40% of iREAD high-risk group readmitted within 48 h, >4-fold higher than traditional score-based stratification.

Clinical Implications

Integrate iREAD into discharge workflows to flag high-risk patients for interventions (e.g., delayed transfer, step-down monitoring). Prospective implementation trials are warranted.

Limitations

• Retrospective model development; performance degradation on external datasets
• Clinical impact untested; prospective implementation and calibration needed

Future Directions

Prospective, cluster-randomized implementation to test outcome impact; fairness auditing; local recalibration; integration with discharge checklists and EHR.