LLM gains medical expertise in Brazilian healthcare
Researchers have trained a Large Language Model (LLM) with Brazilian clinical guidelines to enhance its ability to handle medical information specific to the country's healthcare system.
Vad har hänt
A study published on arXiv describes how researchers adapted the Qwen2.5-14B-Instruct model to the Brazilian clinical domain. This was achieved using 178 official clinical guidelines from Brazil's Unified Health System (SUS) as a foundation. The researchers generated approximately 70 million synthetic data tokens in formats such as reformulations, wiki articles, and question-answer pairs. The model then underwent continuous pre-training followed by Group Relative Policy Optimization (GRPO) to integrate this knowledge.
Key facts
| Antal kliniska riktlinjer | 178 |
|---|---|
| Genererade syntetiska tokens | ~70 miljoner |
| Basmodell | Qwen2.5-14B-Instruct |
| Antal invånare (SUS patienter) | Över 200 miljoner |
| HealthBench-BR prestanda | 83.9% |
| PCDT-QA prestanda | 85.4% |
”Brazil's Unified Health System (SUS) relies on official clinical guidelines that define diagnostic criteria, treatments, dosages, and monitoring procedures for over 200 million citizens. Yet current LLMs perform poorly on this guideline-specific knowledge, and no benchmark evalua”
Varför det spelar roll
Current LLMs perform poorly regarding domain-specific knowledge, particularly in medicine where local guidelines are critical. By integrating official Brazilian clinical guidelines, the model can potentially provide more relevant and accurate responses within the Brazilian healthcare framework. This is vital for ensuring that AI tools used in medicine are aligned with local protocols and standards.
Vem påverkas
This development affects researchers and developers in Natural Language Processing (NLP) working on medical applications, particularly those focused on the Brazilian market. Indirectly, it may impact healthcare professionals and patients in Brazil, as more well-informed AI systems can support diagnostics and treatment based on local guidelines.
EU-status
Not applicable to EU status.
Mer att veta
The researchers also introduced two new benchmarking tools: HealthBench-BR for evaluating clinical claims and PCDT-QA for open-ended clinical questions, measuring performance based on Brazilian protocols. The top-performing model achieved 83.9% on HealthBench-BR and 85.4% on PCDT-QA.
Quick answers about this story
Vad har hänt?
Länken öppnar i nytt fönster och leder till utgivarens egen sida.
Källan har spårats automatiskt från utgivaren via Aheadlines signalkedja.