Benchmarking Small Language Models for Spanish Clinical Text Simplification and Health Literacy

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Scientific Reports https://doi.org/10.1038/s41598-026-65740-w
Paloma Martínez et al.

Points

Universidad Carlos III de Madrid researchers developed the MEDICLARO corpus to evaluate small artificial intelligence language models for simplifying Spanish clinical notes into accessible plain language.
The benchmark dataset incorporated fifty clinical notes with one hundred fifty human written adaptations crafted by cognitive accessibility experts following international ISO plain language standards.
Comparative evaluations across four model families demonstrated that Llama three point two three B offered the most balanced profile while RigoChat v two seven B prioritized readability.
Qualitative error analysis revealed recurring risks of oversimplification factual clinical hallucinations and sentence fragmentation across all tested artificial intelligence model architectures.
Authors concluded that while small language models provide a sustainable foundation for digital health accessibility human clinician verification remains mandatory before patient distribution.

Summary

This study evaluated the performance of small artificial intelligence language models in simplifying Spanish clinical reports into plain language for nonexpert audiences. Published in Scientific Reports by researchers at Universidad Carlos III de Madrid led by Paloma Martínez and Lourdes Moreno, the investigation addressed the widespread challenge of low health literacy, given that approximately 46% of European citizens struggle to comprehend medical documentation. Seeking compliance with upcoming cognitive accessibility regulations under Royal Decree 707/2026, the research sought to establish the first open-source benchmark for Spanish clinical text simplification across multiple small language model architectures.

To conduct the benchmark, investigators created the MEDICLARO corpus, comprising 50 clinical notes sourced from the Andalusian Society of Family and Community Medicine. Cognitive accessibility experts adapted each note into three independent plain-language versions according to ISO 24495-1:2023 international standards, generating 150 gold-standard reference texts. Fine-tuning and prompt-based strategies were evaluated across four model families (Llama 3.2 1B and 3B, Salamandra 2B and 7B, RigoChat-v2 7B, and Medical-mT5 3B). Quantitative and qualitative metrics assessed simplification, semantic similarity, factual consistency, readability, and environmental footprint. Comparative analysis established Llama-3.2-3B as offering the most balanced overall profile in efficiency, robustness, and computational performance, whereas RigoChat-v2-7B demonstrated superior output quality and readability.

The authors conclude that fine-tuned small language models offer a sustainable, patient-centered foundation for integrating plain-language simplification into Spanish clinical workflows. Qualitative error analyses identified persistent risks of oversimplification, clinical hallucinations, and structural incoherence across all evaluated model families. Consequently, the authors emphasize that expert human oversight remains a non-negotiable requirement to verify factual accuracy and prevent clinical miscommunication before AI-simplified reports are delivered to patients.

Link to the article: https://www.nature.com/articles/s41598-026-65740-w

References

Martínez, P., Sanchez-Gomez, J. M., & Moreno, L. (2026). Benchmark of small language models for plain-language simplification in Spanish clinical texts. Scientific Reports, 16(1), 37600. https://doi.org/10.1038/s41598-026-65740-w