Frozen Encoder BioMedBERT Medical Triage Classifier

Model Description

This repository contains a frozen-encoder BioMedBERT sequence classification model. The pretrained encoder is kept frozen and only the classification head is trained.

The model receives a short English symptom description and predicts one of the following triage-oriented labels:

self_monitor
consult_gp
urgent

This model was developed as part of the SortMed academic project, a medical pre-triage assistant prototype built for a bachelor's thesis by Cristian Untaru at the West University of Timișoara, Faculty of Informatics.

The model is intended for research and educational use. It is not a diagnostic system and is not a certified medical device.

Intended Use

This model is intended to be used in an academic medical pre-triage prototype for:

classifying English symptom descriptions into broad urgency categories;
comparing transformer-based models for medical text classification;
evaluating full fine-tuning against parameter-efficient fine-tuning methods;
supporting experimental NLP research in symptom-based medical pre-triage.

Example input:

I have chest pain, shortness of breath, and I feel dizzy.

Example output:

urgent

Out-of-Scope Use

This model must not be used as:

a replacement for professional medical advice;
a diagnostic system;
an emergency decision-making tool;
a certified clinical triage system;
a standalone healthcare product;
a tool for non-English medical text without additional validation.

For severe, worsening, or emergency symptoms, users should contact emergency services or a qualified healthcare professional.

Labels

The model predicts one of three labels:

Label	Meaning
`self_monitor`	Symptoms appear mild and may be monitored, assuming no worsening or additional warning signs.
`consult_gp`	A general practitioner or non-emergency medical professional should be consulted.
`urgent`	Symptoms may require urgent medical attention or emergency evaluation.

The label mapping is included in label_map.json and in the model configuration.

Expected mapping:

{
    0: "self_monitor",
    1: "consult_gp",
    2: "urgent",
}

Training Data

The model was trained on the cristian-untaru/symcat-medical-triage-dataset, a SymCAT-derived dataset prepared for academic experiments in medical pre-triage classification.

The dataset contains English symptom-based text examples labeled into three triage classes:

self_monitor
consult_gp
urgent

Dataset split used for training:

Split	Examples
Train	490
Validation	105
Test	106
Total	701

The dataset labels were produced using rule-based triage logic for academic experimentation. They are not clinically validated annotations from medical professionals.

MedQuAD was not used to train this classifier. In the broader SortMed system, MedQuAD is used separately for retrieval of related medical information through the cristian-untaru/medquad-retrieval-pretriage dataset.

Training Procedure

Base checkpoint:

microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext

Parameter	Value
Fine-tuning method	Frozen Encoder + Trainable Classification Head
Maximum sequence length	128
Epochs	8
Train batch size	16
Evaluation batch size	32
Learning rate	0.0002
Weight decay	0.01
Warmup ratio	0.1
Mixed precision	FP16
Early stopping patience	3
Best checkpoint metric	f1
Random seed	42
Trainable modules	pooler, classifier

Additional training metadata is available in training_config.json.

Evaluation Results

The model was evaluated on the held-out test split.

Metric	Test Score
Accuracy	0.6038
Macro Precision	0.6201
Macro Recall / Sensitivity	0.6058
Macro Specificity	0.7995
Macro F1	0.6107
Macro AUC OvR	0.8076
Macro IoU / Jaccard	0.4444
Test Loss	0.8381

Per-Class Results

Per-class metrics are not currently published as a separate artifact in this repository. The aggregate test metrics above are the available published evaluation results.

These results should be interpreted in the context of the small dataset size and the academic nature of the experiment.

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo_id = "cristian-untaru/frozen-encoder-biomedbert-medical-triage"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

text = "I have chest pain, shortness of breath, and I feel dizzy."

inputs = tokenizer(
    text,
    return_tensors="pt",
    truncation=True,
    padding=True,
    max_length=128
)

with torch.no_grad():
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)
    predicted_class_id = torch.argmax(probabilities, dim=-1).item()

predicted_label = model.config.id2label[predicted_class_id]
confidence = probabilities[0][predicted_class_id].item()

print("Predicted label:", predicted_label)
print("Confidence:", round(confidence, 4))

This model loads with the standard Transformers API and does not require trust_remote_code=True.

To verify the label mapping:

print(model.config.id2label)

Expected output:

{0: "self_monitor", 1: "consult_gp", 2: "urgent"}

Repository Files

File	Description
`config.json`	Model architecture, label mapping, and classification configuration.
`model.safetensors`	Model weights in SafeTensors format.
`tokenizer.json`	Tokenizer vocabulary and processing pipeline.
`tokenizer_config.json`	Tokenizer configuration.
`training_config.json`	Human-readable training configuration.
`label_map.json`	Mapping between numeric class IDs and triage labels.
`README.md`	Model card documentation.
`.gitattributes`	Git LFS configuration for large model files.

Limitations

This model has several important limitations:

It was trained on a small academic dataset.
The labels are rule-based and not clinically validated by medical experts.
It does not perform medical diagnosis.
It does not consider patient age, sex, medical history, medication history, vital signs, comorbidities, pregnancy status, allergies, or physical examination findings.
It may produce incorrect predictions for rare, ambiguous, incomplete, severe, or contradictory symptom descriptions.
It was trained for English text only.
It should not be used for emergency triage without medical supervision.
Its predictions depend on the quality and completeness of the user-provided text.

Ethical and Safety Considerations

Medical pre-triage systems can influence user decisions in sensitive healthcare situations. This model should therefore be used only with clear disclaimers and with safeguards that encourage users to seek professional medical care when appropriate.

Any real-world deployment would require:

clinical expert review;
validation on larger and more representative medical datasets;
safety testing;
bias and fairness evaluation;
monitoring for incorrect or unsafe predictions;
clear user-facing warnings;
escalation rules for emergency symptoms.

Medical Disclaimer

This model is intended only for academic, research, and prototype development purposes. It is not a substitute for professional medical advice, diagnosis, treatment, or emergency care.

If symptoms are severe, sudden, worsening, or potentially life-threatening, users should contact emergency services or a qualified healthcare professional immediately.

Related Datasets

Related Models

Full Fine-Tuned Models

LoRA Models

Bottleneck MLP Adapter Models

Frozen Encoder Models

Downloads last month: 13

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for cristian-untaru/frozen-encoder-biomedbert-medical-triage

Base model

microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext

Finetuned

(190)

this model

Dataset used to train cristian-untaru/frozen-encoder-biomedbert-medical-triage

Evaluation results

Accuracy on SymCAT Medical Triage Dataset
self-reported

0.604
Macro F1 on SymCAT Medical Triage Dataset
self-reported

0.611
Macro Precision on SymCAT Medical Triage Dataset
self-reported

0.620
Macro Recall on SymCAT Medical Triage Dataset
self-reported

0.606
Macro AUC OvR on SymCAT Medical Triage Dataset
self-reported

0.808