hsilvosa/BETO-ATC-Classifier

BETO ATC Level-1 Classifier

Fine-tuned dccuchile/bert-base-spanish-wwm-cased (BETO) that maps a Spanish drug description (name, dose, dosage form, active ingredients, route) to its ATC level-1 anatomical group (single-label, 14 classes). Levels 2-5 are not modelled. Training code: src/train_atc_classifier.py.

Evaluation

Held-out test set: 7996 rows (2518 unique texts), split by active-ingredient set. No ingredient combination in the test set appears in training.

Metric	Value
Top-1 accuracy	79.3%
Top-3 accuracy	87.7%
Macro F1	0.691

Limits:

• One split seed was run, so the variation between seeds is unknown.
• Class P (antiparasitics, 143 rows in total) has no rows in the test set and is not measured.
• The dataset has one row per medicine presentation, so many rows repeat. A random row split gives about 99.9% accuracy because almost every test text also appears in training. Do not read that number as generalisation.
• The model predicts the ATC group from the description. It does not replace the official AEMPS classification.

Intended Use & Disclaimer

Intended only for medical research and pharmaceutical data analytics. Not for clinical prescribing or medical advice.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo = "hsilvosa/BETO-ATC-Classifier"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)

text = "Medicamento: Omeprazol 20 mg | Forma farmacéutica: COMPRIMIDO | Principios activos: Omeprazol"
inputs = tokenizer(text, return_tensors="pt")
probs = torch.softmax(model(**inputs).logits, dim=-1)
print(probs)