Agentic-AI-based emergency room admission optimization
“The data was already there. What wasn’t organized were the criteria bed operations actually run on.”
Problem
An emergency patient goes through registration, initial assessment, tests and consults before admission is decided and they move to a ward or ICU. Tracking bed status and coordinating those moves ran on phone calls, paper and individual experience, so bottlenecks only became visible after the fact.
- Transfer delay — admitted patients keep occupying ER beds
- Fragmented data — ER, ward, admin, nursing and transport data split across OCS, EMR and ERP
- Unstructured tacit knowledge — expert criteria are not in a form AI can use
What BeSir did
Go beyond improving one hospital and develop a standard ER bed-management AI model applicable to regional emergency centers and mid-sized general hospitals nationwide.
- 1 · Extracting tacit knowledge — interviews and job shadowing with emergency medicine, nursing, admin and transport staff capture the criteria. Following the patient and family journey also surfaces bottlenecks outside clinical care.
- 2 · Building a sovereign ontology — beds, patients, wards and departments are connected into a structure AI can query and reason over.
- 3 · Validation by the hospital’s own team — MJAX Avengers, the hospital’s execution unit, reviews and refines the ontology directly.
- 4 · Agentic orchestration — bed lookup, admission-wait prediction, transfer feasibility, proactive recommendations and department alerts run as one flow, with clinicians verifying the output.
Targets
- Target
Bottleneck detection accuracy ≥ 80%
- Target
ER acceptance rate improved by 10%
- Target
Admission-wait prediction accuracy ≥ 75%
- Target
AI recommendation adoption ≥ 50%
A baseline is fixed after kickoff and improvement measured against it. In PoC scope, Excel logs, bed-occupancy rates and mockup data replace full real-time EMR integration; all data is de-identified. Ministry of SMEs and Startups public–private open innovation project · 2026.06–2027.01.
Verified confirmed in the operating environment · PoC confirmed within PoC scope · Target a target set on work in progress · Estimate calculated from the stated assumptions