fundacionctic/predict-dermat

Model Card for Model ID

This is a fine-tuned version of the pre-trained biomedical language model bsc-bio-ehr-es in Spanish, tailored for text classification tasks. We used two NVIDIA GPUs for training.

Model Details

Model Description

This model has been fine-tuned for text classification on dermatological Spanish electronic health records (EHR). It leverages the pre-trained biomedical language understanding from the bsc-bio-ehr-es model and adapts it to classify dermatology-related texts effectively. The model is intended to predict among 25 different skin diseases from a medical record. It could be a first visit or a follow-up visit. It takes as input four features:

• textual medical record: the EHR written by a doctor
• disease type: the type of disease associated with the EHR
• disease location: the location in the body of the disease
• disease severity: how severe or lethal is the disease

It is IMPORTANT to load and concatenate them in this specific order. The details to reproduce the cascade predictions are available in the Training section.

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

• Developed by: Fundacion CTIC
• Funded by [optional]: SATEC
• Model type: Fine-tuned LM Encoder
• Language(s) (NLP): Spanish
• License: CC-BY-NC
• Finetuned from model [optional]: bsc-bio-ehr-es

Model Sources [optional]

• Repository:
• Paper [optional]: Coming soon...
• Demo [optional]: [More Information Needed]

Uses

The Model is industry-friendly and the best model of the dermat collection. The vanilla version of the model is called vanilla-dermat and is meant to predict not only the disease but also the 3 features mentionned above. We DO NOT recommend to fine-tune this model. It is already meant to be a downstream task.

Direct Use

This model can be directly used for classifying dermatological text data in Spanish EHRs.

Downstream Use

The model can be integrated into healthcare applications for automatic classification of dermatological conditions from patient records.

Out-of-Scope Use

The model is not suitable for non-medical text classification tasks or for texts in languages other than Spanish.

Bias, Risks, and Limitations

This model is fine-tuned on a specific dataset and may not generalize well to other types of medical texts or conditions. Users should be cautious of biases in the training data that could affect the model's performance.

Recommendations

Users should validate the model's performance on their specific data and consider any ethical implications of deploying a machine learning model in a healthcare setting.

How to Get Started with the Model

from transformers import RobertaTokenizerFast, RobertaForSequenceClassification,

tokenizer = RobertaTokenizerFast.from_pretrained("fundacionctic/predict-dermat")
model = RobertaForSequenceClassification.from_pretrained("fundacionctic/predict-dermat")


inputs = tokenizer("Ejemplo de texto dermatológico + tipo + localizacion + gravedad".tolist(),
            truncation=True,
            padding='max_length',
            max_length=max_length,  # Replace with your desired maximum sequence length
            return_tensors='pt',
            return_attention_mask=True,
        ))
outputs = model(input_ids, attention_mask=attention_mask)

Training Details

Training Data

The model was fine-tuned on the DermatES dataset from Fundación CTIC, which contains Spanish dermatological EHRs.

Training Procedure

In order to reproduce the experiment it is ESSENTIAL to respect the order of prediction of the three ontology-base features. More details in the original paper of Dermat

from transformers import RobertaTokenizerFast, RobertaForSequenceClassification,

tokenizer = RobertaTokenizerFast.from_pretrained("PlanTL-GOB-ES/bsc-bio-ehr-es")
model = RobertaForSequenceClassification.from_pretrained("PlanTL-GOB-ES/bsc-bio-ehr-es")

def reset_model():
  model =  RobertaForSequenceClassification.from_pretrained("PlanTL-GOB-ES/bsc-bio-ehr-es")

def cascade(inputs,information_list,model,tokenizer,predictions=None):
  if not information_list:
    return predictions
  else : 
    inputs = tokenizer("Ejemplo de texto dermatológico".tolist(),
                truncation=True,
                padding='max_length',
                max_length=max_length,  # Replace with your desired maximum sequence length
                return_tensors='pt',
                return_attention_mask=True,
            ))
    labels = information_list[0].tolist()
    outputs = model(input_ids, attention_mask=attention_mask,labels=labels)
    predictions = torch.argmax(outputs.logits,dim=1)
    inputs = [tokenizer.decode(input+predictions[i]) for i,input in enumerate(inputs)]
    model = reset_model()
    return cascade(inputs,information_list[1:],model,tokenizer,predictions)
inputs = ["un informe,","otro informe"]
information_list = [[tipo1,tipo2],[sitio1,sitio2],[gravedad1,gravedad2]]
predicted_diseases = cascade(inputs,information_list,model,tokenizer)

Preprocessing

Lowercased, anonymized and accents removed texts

Training Hyperparameters

• Training regime: fp32

Speeds, Sizes, Times

Epochs: 7
Batch size: 64
Learning rate: 0.0001

Evaluation

Testing Data, Factors & Metrics

Testing Data

The evaluation was performed on 0.2 of the DermatES dataset.

Factors

[More Information Needed]

Metrics

- *Accuracy:* 0.51
- *F1 Score:* 0.42
- *top-k (k=2) accuracy:* 0.67
- *top-k (k=2) f1 Score:* 0.61

Summary

The model achieves poor accuracy and F1 score on dermatological text classification, demonstrating the need of using ontologies (see oracle-dermat ) for this specific medical domain.

Model Examination [optional]

[More Information Needed]

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: GPU and CPU with avx
• Hours used: >96
• Cloud Provider: [More Information Needed]
• Compute Region: EU
• Carbon Emitted: [More Information Needed]

Technical Specifications [optional]

Model Architecture and Objective

The model is based on the RoBERTa architecture, fine-tuned for the objective of text classification in the biomedical domain.

Compute Infrastructure

Hardware

Two NVIDIA GPUs were used for the fine-tuning process.

Software

The fine-tuning was performed using the 🤗 Transformers library.

Citation

BibTeX:

Original paper

@inproceedings{schaub2024,
  author    = {L.-P. Schaub Torre and P. Quiros and H. Garcia-Mieres},
  title     = {Detección Automática de Patologías en Notas Clínicas en Español Combinando Modelos de Lenguaje y Ontologías Médicos},
  booktitle = {SEPLN-2024: 40th Conference of the Spanish Society for Natural Language Processing},
  year      = {2024},
  pages     = {99--120},
  publisher = {Universidad de Valladolid, ECA-SIMM Research Group, Departamento de Informática},
  address   = {Valladolid},
  url       = {https://ceur-ws.org/Vol-3846/},
}

English version

@misc{torre2024automaticdetectiondiseasesspanish,
      title={Automatic detection of diseases in Spanish clinical notes combining medical language models and ontologies}, 
      author={Leon-Paul Schaub Torre and Pelayo Quiros and Helena Garcia Mieres},
      year={2024},
      eprint={2412.03176},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.03176}, 
}

APA:

Schaub Torre, L.-P., Quiros, P., & Garcia-Mieres, H. (2024). Detección automática de patologías en notas clínicas en español combinando modelos de lenguaje y ontologías médicos. In SEPLN-2024: 40th Conference of the Spanish Society for Natural Language Processing (pp. 99–120). Universidad de Valladolid, ECA-SIMM Research Group, Departamento de Informática. https://ceur-ws.org/Vol-3846/

Glossary [optional]

More Information [optional]

[More Information Needed]

Model Card Authors

Leon-Paul Schaub Torre, Pelayo Quiros and Helena Garcia-Mieres

Model Card Contact

leon.schaub@fundacionctic.org pelayo.quiros@fundacionctic.org