News & events
MEDICLARO: an expert-created multi-reference corpus.
A benchmark for Small Language Models in Spanish clinical simplification.
Open resources for the research community.
A new HumanAI study has been published in Scientific Reports: “Benchmark of Small Language Models for Plain-Language Simplification in Spanish Clinical Texts”, by Paloma Martínez, Jesús M. Sánchez-Gómez and Lourdes Moreno. The work addresses the still limited availability of resources and evaluation frameworks for plain-language adaptation of clinical information in Spanish.
The study makes two main contributions.
The first is MEDICLARO, an openly available, multi-reference corpus specifically created for Spanish clinical plain-language adaptation. It contains 50 clinical texts and 150 expert-written adaptations, with three independent plain-language versions for each original text. The adaptations were produced by experienced specialists in plain language and cognitive accessibility following a documented methodology grounded in ISO 24495-1:2023. This multi-reference design is especially valuable because plain-language adaptation does not have a single correct solution: different expert reformulations can be valid while preserving the same clinical meaning.
The second contribution is the first systematic benchmark of Small Language Models for Spanish clinical plain-language adaptation. Six models from four families were compared using fine-tuning and prompting strategies and evaluated across simplification, semantic preservation, factual consistency, readability, human assessment and computational efficiency. This provides a reproducible evaluation framework for a task for which Spanish clinical resources and benchmarks remain scarce. Importantly, the study treats plain language as a defined methodological framework, rather than simply equating simplification with shorter sentences or easier vocabulary, and grounds the adaptation process in ISO 24495-1.
The results show that Llama-3.2-3B offers the most balanced profile across robustness and efficiency, while RigoChat-v2-7B performs particularly well when meaning preservation and output quality are prioritised. The study also shows why automatic metrics alone are insufficient for clinical simplification and why human evaluation remains necessary.
Read the article in Scientific Reports
MEDICLARO corpus on Zenodo
Evaluation framew
The MEDICLARO corpus is in Spanish. The following example is shown in its original language.
| Versión | Fragmento |
|---|---|
| Texto clínico original |
Paciente de 17 años acude a consulta de atención primaria para recoger resultados de una analítica.
Anorexia nerviosa. |
| Adaptación experta 1 |
La paciente, de 17 años de edad, viene a consulta para recoger unos análisis.
La paciente viene acompañada de su madre.
Anorexia nerviosa: Trastorno de la alimentación caracterizado por un temor intenso de aumentar de peso, un rechazo a mantener un peso normal y una imagen del cuerpo deformada. |
| Adaptación experta 2 |
La paciente viene acompañada por su madre para recoger los resultados de unos análisis.
La paciente tiene anorexia nerviosa, es decir, no come por miedo a engordar. |
| Adaptación experta 3 |
Una paciente de 17 años va a la consulta del médico de cabecera a recoger los resultados del análisis de sangre.
Anorexia nerviosa, es decir, trastorno del comportamiento alimentario por el que la paciente come pocas cantidades de alimentos. |
Each MEDICLARO source text has three independent plain-language adaptations produced by experts. The multi-reference design captures different valid adaptations of the same clinical information following a common plain-language methodology grounded in ISO 24495-1:2023.
Project PID2023-148577OB Funded by:
Accessibility
Este sitio web se ha diseñado para cumplir con el Real Decreto 1112/2018, de 7 de septiembre, sobre accesibilidad de los sitios web y aplicaciones para dispositivos móviles del sector público, en el que se indica que hay que ser conforme con la norma UNE-EN 301 549, que en lo que se refiere al contenido web, es cumplir con las Web Content Accessibility Guidelines (WCAG) 2.1.
Social networks
