README.md

metadata

language:
  - en
license: apache-2.0
base_model: emilyalsentzer/Bio_ClinicalBERT
tags:
  - medical
  - clinical
  - ssi
  - classification
  - surveillance
  - multi-label
metrics:
  - accuracy
  - f1
  - precision
  - recall
model-index:
  - name: SSIBERT-multiclass
    results:
      - task:
          type: text-classification
          name: Multi-Label SSI Detection
        dataset:
          name: Synthetic UK NHS Clinical Notes (Multi-Label)
          type: synthetic
          split: test
        metrics:
          - name: F1 (Micro)
            type: f1
            value: 1

Model Card for Ch3DS/SSIBERT-multiclass

Model Details

Model Description

This model is a fine-tuned version of Bio_ClinicalBERT designed for multi-label classification of postoperative clinical notes. Unlike the binary SSI model, this model identifies specific clinical indicators of infection:

1. Purulence: Presence of pus or purulent discharge.
2. Redness: Erythema, spreading redness, or inflammation.
3. Fever: Pyrexia, rigors, or elevated temperature.
4. Antibiotics: Prescription of antibiotics (treatment or prophylaxis).
5. SSI: Overall determination of Surgical Site Infection.

It is tailored to UK NHS terminology.

• Developed by: Daryn Sutton
• Model type: Multi-Label Text Classification (BERT)
• Language(s) (NLP): English
• License: Apache 2.0
• Finetuned from model: emilyalsentzer/Bio_ClinicalBERT
• Repository: https://huggingface.co/Ch3DS/SSIBERT-multiclass

Uses

Direct Use

This model extracts structured data from unstructured clinical notes, allowing for more granular surveillance.

• Input: Clinical note text.
• Output: Probabilities for [Purulence, Redness, Fever, Antibiotics, SSI].

Out-of-Scope Use

• Diagnosis: This is a surveillance tool, not a diagnostic device.
• Non-UK Contexts: May perform poorly on non-NHS terminology.

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "Ch3DS/SSIBERT-multiclass"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "Day 3 post THR. Wound oozing pus. Patient pyrexial. Plan: Start Flucloxacillin."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits)

labels = ["Purulence", "Redness", "Fever", "Antibiotics", "SSI"]
for i, label in enumerate(labels):
    print(f"{label}: {probs[0][i]:.2%}")

Training Details

Training Data

• Source: 5 million synthetic clinical notes.
• Methodology: Generated using templates based on UK NHS terminology and the PRAISE network's surveillance definitions.
• Labels: Multi-hot encoded.

Training Procedure

• Epochs: 3
• Batch Size: 64
• Hardware: NVIDIA GeForce RTX 5070 Ti

Evaluation

Evaluated on a held-out test set of synthetic data. Achieved near-perfect performance on the synthetic distribution.

Model Card Contact

Daryn Sutton