research
The Bed Pressure Enters the Discharge Model
Operational workload and admission flow added independent signals to a multimodal readmission-risk model trained on 20,000 ICU stays.
Summary
Operational workload and admission flow added independent signals to a multimodal readmission-risk model trained on 20,000 ICU stays.
ICU discharge decisions reflect both patient state and hospital capacity, but many models represent only the patient. This system combines cleaned progress notes, departmental workload, admission flow, vital signs and diagnosis codes. Across twenty runs on 20,000 MIMIC-III stays with roughly 10.5 percent readmissions, its end-to-end transformer reached AUROC 0.8288, PR-AUC 0.4373 and F1 0.4757. Raw text embeddings hurt tree models, while tuned multimodal training reversed that transfer. This is retrospective benchmark evidence from one public dataset, not a prospective clinical validation or a deployment recommendation.
Why it matters
Operational workload and admission flow added independent signals to a multimodal readmission-risk model trained on 20,000 ICU stays.
Limits and context
- ICU discharge decisions reflect both patient state and hospital capacity, but many models represent only the patient.
- This is retrospective benchmark evidence from one public dataset, not a prospective clinical validation or a deployment recommendation.
Key claims
Operational workload and admission flow added independent signals to a multimodal readmission-risk model trained on 20,000 ICU stays.
Qualification: ICU discharge decisions reflect both patient state and hospital capacity, but many models represent only the patient.
Evidence: source-2026-11-10-013
Sources
- arXiv preprint 2611.11960, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.