← Back to projects
Pilot study - Published

TIAEU

Intelligent Triage at Emergency Entry

Project Description

TIAEU is a single-center retrospective study including 657 patients at Lille University Hospital over 7 months (2024). This study demonstrated the superiority of 3 AI models (NLP-TRIAGEMASTER, LLM-URGENTIAPARSE, JEPA-EMERGINET) over standard nursing triage. Concordance with final medical decision (FRENCH scale) reached 90% for the best AI vs 30% for standard nursing triage (p<0.001).

Objectives

Evaluate AI model performance for ED triage compared to standard nursing triage using FRENCH scale

Methodology

Single-center retrospective study (CHU Lille) including 657 patients with 10-fold cross-validation

Results

AI concordance: 78% vs Nurse: 30% (p<0.001)
Under-triage: 11% vs 18% (RR=0.61)

Publication

IEEE/ACM BDCAT 2025
Nantes, France

Technical Details

  • TRIAGEMASTER (NLP): Doc2Vec-based with neural regression, paragraph vector analysis
  • URGENTIAPARSE (LLM): Pre-trained BERT model (FlauBERT) fusing text and vital signs
  • EMERGINET (JEPA): Joint Embedding Predictive Architecture with dual encoding

Key Results

Overall Performance

Exact concordance: 90%
Concordance ±1 class: 89%
Weighted Kappa: 0.80

Best Model

URGENTIAPARSE (LLM)
AUC-ROC: 0.879
F1-score macro: 0.894

Project Team

Dr Edouard Lansiaux

Principal Investigator
Lille University Hospital

Dr Ramy Azzouz

Co-director
Lille University Hospital

Pr Emmanuel Chazard

Scientific Committee
METRICS ULR 2694

Publications

  • Journal Artificial Intelligence models for predicting triage in Emergency Departments

    Lansiaux E, Azzouz R, Chazard E, Vromant A, Wiel E.

    JMIR Medical Informatics