Researchers in Spain have tested whether small artificial intelligence language models can turn medical reports into plain language for patients. The study found the tools can help, but every model tested made errors that a human would need to catch.
The work comes from a team of researchers at Universidad Carlos III de Madrid (UC3M). It was published in the journal Scientific Reports.
The project targets a common problem. According to the researchers, 46% of European citizens have trouble understanding medical documents.
“We developed a set of tools to help improve doctor-patient communication,” said study co-author Paloma Martínez, a Full Professor in UC3M’s Department of Computer Science. She said language models can produce one version of a report for doctors and a second version for patients.
The patient version is meant to explain abbreviations and complex phrasing. The goal is to help people understand their health and follow their prescribed treatments.
How the Models Were Tested
The team used 50 medical reports based on clinical cases from the Andalusian Society of Family and Community Medicine. Cognitive accessibility experts adapted each report three separate times.
The experts followed the international plain language standard ISO 24495-1:2023. That process produced 150 adapted texts, which were used to fine-tune and evaluate the models.
Six models from four families were tested. They included Llama 3.2 1B and 3B, Salamandra 2B and 7B, RigoChat-v2 7B, and Medical-mT5 3B.
The researchers measured simplification, meaning, factual accuracy, readability and environmental impact. No single model came out on top in every area.
Llama-3.2-3B showed the most balanced results for efficiency, robustness and overall performance. RigoChat-v2-7B performed best when output quality and readability were the main goals.
Three Types of Errors Found
The team found three main error types across all four model families. The first was oversimplification, where models removed clinical details that patients need to know.
The second was clinical hallucination, where models invented or changed data from the original report. The third was incoherence, such as broken sentences or sudden jumps that made the text hard to follow.
“It is precisely because of these errors that human oversight remains essential for these models, particularly in a context as sensitive as healthcare,” Martínez said.
Co-author Lourdes Moreno, also a Full Professor at UC3M, said plain language is more than shorter sentences and easier words. She said it means organizing content clearly, adapting technical terms and guiding readers based on what they need to know.
The team has released its dataset, fine-tuned models and prompts on public platforms, including GitHub and Zenodo. The work was part of HumanAI, a project under Spain’s State Plan for Scientific, Technical and Innovation Research, funded by the State Research Agency.
The researchers said the study is timely because of Spain’s Royal Decree 707/2026. Published in the Official State Gazette on September 3, it sets the country’s first rules on basic cognitive accessibility.
The decree requires documents, procedures and services to be easier for citizens to understand. It covers public bodies in areas such as healthcare, justice, employment, digital services and transport, and it takes effect on January 2, 2027.
