CodiEsp cross-encoder for the learned IMPLICIT layer (Spanish)

Scores a (mention in context, candidate ICD-10 code) pair: how likely the CodiEsp annotators are to give that code. It is one feature of the learned IMPLICIT layer of CEI-Trace, which chooses among the codes the official books offer for each mention; it never proposes codes by itself.

  • Base model: PlanTL-GOB-ES/roberta-base-biomedical-clinical-es (Apache-2.0), a 125-million-parameter Spanish clinical encoder, with a sequence-classification head.
  • Training: CodiEsp training and development annotations (the configuration used for the test run).
  • Load with transformers (AutoModelForSequenceClassification); pairs are built by scripts/convention_crossenc.py.

This model learns one dataset's annotation practice, mistakes included, by design: use it only as a labelled IMPLICIT layer.

Use

Part of CEI-Trace (https://github.com/putssander/CEI-Trace), which codes clinical text by the official ICD-10 books with local language models and adds each setting's guidelines and practice as separate layers. Article: Puts S, Dekker A. CEI-Trace: fitting ICD-10 coding to a setting's guidelines and practice in layers around one official core (2026).

Not for clinical use. Trained on published Spanish clinical case reports (CodiEsp, CC BY 4.0; Miranda-Escalada et al., CLEF eHealth 2020, doi:10.5281/zenodo.3837305).

Downloads last month
52
Safetensors
Model size
0.1B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sander-puts/codiesp-convention-crossencoder-es

Finetuned
(19)
this model