Intended use
Workflow engineering and hackathon demonstration of an uncertainty-aware ED waiting-room reassessment board using synthetic data. It may help study whether users can distinguish current risk from information quality and reassessment value.
Prohibited use
- Clinical care, diagnosis, treatment, or autonomous escalation
- Replacing ESI, standard reassessment protocols, or clinical judgment
- Claims about real-world waiting-room deterioration or prospective efficacy
- Deployment across hospitals without external validation and governance
Outcome framing
The estimated probability refers to a generated retrospective critical-outcome proxy. It does not predict a diagnosis and cannot establish that an event occurred while a person waited.
Data provenance
The displayed shift and development cohort are entirely synthetic. Scenario distributions are correlated by age, acuity, complaint, workload, and latent severity. No protected health information, MIMIC records, or private datasets are included.
Uncertainty & abstention
The plausible range combines bootstrap model spread with missingness and observation-age penalties. It is a prototype uncertainty indicator—not a clinically validated confidence interval. Excessive uncertainty or out-of-distribution inputs trigger abstention.
Known limitations
- Synthetic data cannot establish clinical performance or fairness.
- The conditional observation sampler is inspectable but simplified.
- Policy weights encode a prototype workflow preference and require sensitivity analysis with clinicians.
- The local SQLite fallback is designed for reproducible demo state, not hospital-scale concurrency.
- The failure scenario is deliberately injected to verify safe fallback behavior.
Safety behavior
When inference fails, old predictions are not shown as current. Freshness remains visible, the queue falls back to documented acuity, overdue reassessment, and staleness, and the failure is recorded in the audit log.