Emergency Department Triage Support System

Ethan Zhang · 2026 LACSEF

The Problem

Over 130 million patients visit U.S. emergency departments each year. Nurse-assigned ESI scores match actual outcomes only 19.9% of the time, and for the most critical level (ESI-1), over 50% of patients are under-triaged.

About This Work

All data, models, and code are open source and available via the links above. This work was done independently at home, including processing 67 GB of clinical data, manually verifying over 1,000 patient records, and training 3 models.

Key Findings

Uncovered concrete evidence that nurses unconsciously under-triage minority patients with the same symptoms — consistently across most clusters and all high-population clusters representing more than 50% of patients.

The model can assist nurses and alert them to unconscious bias. Fully implemented, EDTSS could reduce under-triage by 42% — saving an estimated 87,000 lives every year.

Built the EDTSS tool to augment nurse triage decisions with model-driven insights.

Three-Model Framework

Protocol-Compliant62%

ESI-compliant features only

Rogue Model66%

Adds demographics — 4pt gap quantifies bias

Retrospective ESI80%

Trained on actual outcomes — the achievable ceiling

Data Source

MIMIC-IV — a de-identified critical care database from MIT & Beth Israel Deaconess Medical Center. 98,396 patient records across 330M+ rows of data.