Desarrollo de una aplicación web para la identificación de contraindicaciones médicas utilizando modelos de lenguaje natural en español

Abstract

Abstract: Identifying contraindications between medications is a fundamental process for preventing serious adverse medical effects. However, manually consulting pharmacological sources is time-consuming and prone to errors; furthermore, the common practice of self-medication a recurring issue in the country is recognized as a potential risk. In this context, this capstone project aimed to develop a web application prototype for identifying contraindications between medications, utilizing natural language processing models in Spanish. To achieve this, an experimental design was implemented to compare two approaches: a traditional large language model (LLM) and a model based on Retrieval-Augmented Generation (LLM + RAG). This implementation utilized the Mistral 7B Instruct v0.2 language model, a knowledge base built from specialized pharmacological information, and a vector database powered by ChromaDB. The system was subsequently evaluated. The results demonstrated that the LLM + RAG model generated more accurate and relevant responses, supported by clinical evidence, compared to the traditional model; the system also exhibited an adequate level of quality, stability, and usability.