SHUB-8/Triage-BERT

TriageBERT - ESI Triage Classification Model

A BERT-based model fine-tuned for Emergency Severity Index (ESI) classification on PubMedBERT.

Model Description

This model classifies emergency medical text into 5 ESI levels:

• ESI 1: Immediate (life-threatening)
• ESI 2: Emergent (high risk)
• ESI 3: Urgent (stable but needs multiple resources)
• ESI 4: Less Urgent (single resource needed)
• ESI 5: Non-Urgent (no resources needed)

Training

• Base Model: PubMedBERT (biomedical domain)
• Training Data: MIMIC-IV ED-Triage data + synthetic data
• Optimization: Recall-optimized for critical cases (ESI 1-2)

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model
tokenizer = AutoTokenizer.from_pretrained("SHUB-8/Triage-BERT")
model = AutoModelForSequenceClassification.from_pretrained("SHUB-8/Triage-BERT")

# Predict
text = "Patient has chest pain and difficulty breathing"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    esi_level = torch.argmax(probs).item() + 1  # ESI 1-5

print(f"ESI Level: {esi_level}")

Performance

Metric	Value
Accuracy	~85%
Recall (ESI 1-2)	~92%
F1 Score	~83%

Intended Use

• Emergency department triage assistance
• Medical emergency prioritization
• Healthcare AI research

Limitations

• English language only
• Should be used as decision support, not replacement for clinical judgment
• Performance may vary on out-of-distribution data

Citation

If you use this model, please cite:

@misc{triage-bert-esi,
  author = {Your Name},
  title = {TriageBERT: ESI Triage Classification Model},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/SHUB-8/triage-bert-esi-recall-optimized}
}