Large languages (LLM) models have shown a potential for medical answers and questions about various health -related tests and covering different formats and sources. Indeed, we were at the forefront of efforts to extend the usefulness of LLM for health and medical applications, as demonstrated in our recent work on Med-Gemini,, Medalm,, Friend,, Multimodal medicaland our release of New assessment tools and methods To assess the performance of the model in various contexts. In particular in low -resources contexts, LLM can potentially serve as precious decision -making tools, improve clinical diagnosis accuracy, accessibility and support for multilingual clinical decision, and health training, especially at community level. However, despite their success on existing medical references, there is still a certain uncertainty as to how these models are generalized in tasks involving distribution changes in the types of diseases, the medical knowledge specific to the region and the contextual variations between symptoms, language, location, linguistic diversity and localized cultural contexts.
Tropical and infectious diseases (Trinds) are an example of such an out-of-distributed disease subgroup. The trinds are very widespread in the poorest regions in the world, Assuming 1.7 billion people worldwide With disproportionate impacts on women and children. The prevention and treatment challenges of these diseases include limits Monitoring, early detection, precise initial diagnosis,, Management and vaccines. The LLM for the answer to health related questions could potentially allow early detection and monitoring depending on the symptoms, location and risk factors. However, only limited studies have been conducted to understand LLM performance on Trinds with few existing data sets for a rigorous LLM assessment.
To fill this gap, we have developed synthetic characters -that is to say, sets of data that represent profiles, scenarios, etc., which can be used to assess and optimize models-and reference methodologies for diseases of disease outside distribution. We have created a set of Trinds data which consists of more than 11,000 manual personalities generated by LLM representing a wide range of tropical and infectious diseases through demographic, contextual, language, language, clinical and consumption increases. Part of this work was recently presented in the 2024 Neirips workshops AI Generative for health And Progress of medical foundation models.