README.md

---
language:
- en
task_categories:
- text-classification
task_ids:
- multi-class-classification
pretty_name: SymCAT Medical Triage Dataset
tags:
- medical-triage
- symptom-classification
- healthcare
- pre-triage
- symcat-derived
- natural-language-processing
- academic-project
size_categories:
- n<1K
---

# SymCAT Medical Triage Dataset

This repository contains a SymCAT-derived dataset prepared for symptom-based medical pre-triage text classification.

The dataset was created and processed as part of a bachelor's thesis project by **Cristian Untaru** at the **West University of Timișoara, Faculty of Informatics**.

## Dataset Description

The dataset contains natural-language symptom descriptions labeled into three triage-oriented categories:

- `self_monitor`
- `consult_gp`
- `urgent`

Each example is generated from SymCAT-derived condition and symptom information and converted into a text classification format suitable for fine-tuning transformer-based language models.

The dataset is intended for academic and experimental work on medical pre-triage classification. It is not intended to be used as a standalone clinical dataset or as a source for real-world medical decision-making.

## Intended Use

This dataset can be used for:

- fine-tuning text classification models for symptom-based pre-triage;
- comparing transformer-based models such as DistilBERT, BioBERT, and RoBERTa;
- academic experiments in natural language processing for healthcare;
- prototyping medical pre-triage assistant systems.

The dataset was used to fine-tune transformer-based triage classification models in the broader pre-triage assistant project.

## Labels

| Label | Meaning |
|---|---|
| `self_monitor` | The symptoms may be monitored by the patient, assuming no worsening or additional warning signs. |
| `consult_gp` | The patient should consider consulting a general practitioner or a non-emergency medical professional. |
| `urgent` | The symptoms may require urgent medical attention or emergency evaluation. |

The label mapping is provided in `label_map.json`.

## Dataset Splits

The dataset is provided using stratified train, validation, and test splits.

| Split | File | Number of examples |
|---|---|---:|
| Train | `train.csv` | 490 |
| Validation | `validation.csv` | 105 |
| Test | `test.csv` | 106 |
| Full dataset | `full.csv` | 701 |

The split ratio is approximately 70% train, 15% validation, and 15% test.

## Class Distribution

The class distribution in the full dataset is:

| Class | Number of examples |
|---|---:|
| `self_monitor` | 220 |
| `consult_gp` | 266 |
| `urgent` | 215 |

Additional dataset statistics:

| Statistic | Value |
|---|---:|
| Total examples | 701 |
| Duplicate texts | 0 |
| Imbalance ratio | 1.237 |
| SymCAT-derived conditions before final subset selection | 801 |
| Condition-level overrides applied before final subset selection | 75 |

## Data Fields

The CSV files include the following fields:

| Field | Description |
|---|---|
| `condition_name` | Name of the medical condition associated with the generated example. |
| `condition_slug` | Normalized condition identifier. |
| `symptom_names` | List of symptom names associated with the condition. |
| `symptom_slugs` | Normalized symptom identifiers. |
| `symptom_probabilities` | Symptom probabilities extracted from the SymCAT-derived source. |
| `text` | Natural-language input text used for classification. |
| `num_symptoms` | Number of symptoms included in the generated text. |
| `max_symptom_probability` | Maximum symptom probability for the example. |
| `mean_symptom_probability` | Mean symptom probability for the example. |
| `urgent_score` | Rule-based urgency score used during weak labeling. |
| `gp_score` | Rule-based general-practitioner consultation score used during weak labeling. |
| `self_score` | Rule-based self-monitoring score used during weak labeling. |
| `label` | Final triage label. |
| `override_applied` | Whether a condition-level override was applied during labeling. |

## Labeling Methodology

SymCAT does not provide direct triage labels. Therefore, the triage labels in this dataset were derived using a weak-supervision approach based on symptom-level risk cues and condition-level overrides.

The labeling process assigns each example to one of three triage categories:

1. `urgent`
2. `consult_gp`
3. `self_monitor`

The general logic is:

- symptoms or conditions associated with emergency warning signs are assigned to `urgent`;
- symptoms that suggest the need for non-emergency medical evaluation are assigned to `consult_gp`;
- lower-risk symptom combinations are assigned to `self_monitor`.

Condition-level overrides were applied in selected cases where a condition required a stronger triage label than the symptom-level rules alone would suggest.

This means that the labels are suitable for academic experimentation and model comparison, but they should not be interpreted as clinically validated triage decisions.

## Files

| File | Description |
|---|---|
| `train.csv` | Training split used for fine-tuning. |
| `validation.csv` | Validation split used during model selection and early stopping. |
| `test.csv` | Held-out test split used for final evaluation. |
| `full.csv` | Complete processed dataset before splitting into train, validation, and test files. |
| `label_map.json` | Mapping between labels and numeric class IDs. |
| `dataset_stats.json` | Dataset-level statistics generated during preprocessing. |
| `.gitattributes` | Git LFS configuration automatically used by Hugging Face/Git. |
| `README.md` | Dataset Card documentation. |

## Relationship to MedQuAD

This dataset is separate from MedQuAD.

The SymCAT-derived dataset in this repository was used for fine-tuning triage classification models. MedQuAD was processed separately in the broader project as a retrieval corpus for contextual medical question-answer information.

The processed MedQuAD retrieval dataset is available separately as:

[`cristian-untaru/medquad-retrieval-pretriage`](https://huggingface.co/datasets/cristian-untaru/medquad-retrieval-pretriage)

Therefore:

- SymCAT-derived dataset: used for training/fine-tuning triage classifiers;
- MedQuAD retrieval dataset: used for contextual retrieval of medical Q&A information;
- the triage classifier is not fine-tuned on MedQuAD.

This distinction is important because the SymCAT-derived dataset provides supervised triage labels, while MedQuAD provides medical question-answer context that can be retrieved and displayed alongside the classifier prediction.

## How to Load the Dataset

The dataset can be loaded with the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("cristian-untaru/symcat-medical-triage-dataset")

print(dataset)
print(dataset["train"][0])
```

Expected splits:

```text
train
validation
test
```

The repository also includes `full.csv`, which contains the complete processed dataset before splitting.

## Example Record

A typical record contains a generated symptom-based text and its corresponding triage label.

Example structure:

```text
text: Condition: Example condition. Patient reports: symptom 1, symptom 2, symptom 3.
label: consult_gp
```

The exact fields available in the CSV files are described in the **Data Fields** section.

## Related Dataset and Model Repositories

Related dataset repositories:

- [`cristian-untaru/symcat-medical-triage-dataset`](https://huggingface.co/datasets/cristian-untaru/symcat-medical-triage-dataset)
- [`cristian-untaru/medquad-retrieval-pretriage`](https://huggingface.co/datasets/cristian-untaru/medquad-retrieval-pretriage)

Related model repositories:

- [`cristian-untaru/distilbert-medical-triage`](https://huggingface.co/cristian-untaru/distilbert-medical-triage)

Additional BioBERT and RoBERTa model repositories may be added separately after training and publication.

## Limitations

This dataset has several important limitations:

- It is a small academic dataset.
- The labels were produced using weak supervision and rule-based triage logic.
- It should not be treated as a clinically validated triage dataset.
- It does not replace professional medical judgment.
- It does not include patient history, age, vital signs, physical examination findings, comorbidities, medication history, or laboratory results.
- Some generated examples may be simplified and may not reflect the full complexity of real patient descriptions.
- The dataset is intended for academic experimentation and prototype development, not real-world clinical deployment.
- The dataset is derived from SymCAT-based condition and symptom information, and users should verify any source-specific usage requirements before reuse.

## Medical Disclaimer

This dataset is intended only for academic, research, and prototype development purposes.

It must not be used as the sole basis for medical diagnosis, treatment, triage, or emergency decision-making. In real-world scenarios, medical triage should be performed by qualified healthcare professionals.

In case of severe, worsening, or life-threatening symptoms, users should contact emergency medical services or a qualified healthcare professional.

## Author

**Cristian Untaru**  
Faculty of Informatics  
West University of Timișoara