Using large language models and knowledge graphs to improve medical drug prescription and healthcare

Abstract

Healthcare is a critical pillar of societal well- being, ensuring that individuals receive the appropriate care to maintain or improve their health. In this context, accurate prescription of medication plays a crucial role in patient safety and treatment outcomes. However, medication errors (MEs) remain a persistent issue. These errors can lead to adverse drug reactions that jeopardize patient wellness in healthcare systems. The rapid expansion of medical data, combined with challenges in retrieving accurate information, compounds the issue. This study addresses the challenge of dosage and administration errors by improving access to accurate prescription information using a Large Language Model (LLM)-powered chatbot. The study used Design Science Research methodology. We fine-tuned the Llama Meditron- 7B model with 25,547 labeled prescription entries from the PubMed database. Using the Retrieval Augmented Generation (RAG) framework, we embedded data from the 1st Edition of the Kenya National Medicine Formulary 2023 and stored it in a Qdrant vector store for efficient retrieval. The model was trained on the NVIDIA A100 GPU 16GB in Google Colab. The model’s performance was evaluated using Recall-Oriented Understudy for Gisting (ROUGE) and Bilingual Evaluation Understudy (BLEU) metrics, and the results were compared to the baseline Meditron and Generative Pre-trained Transformer (GPT)-3.5 models. The Meditron-7B model significantly outperformed its counterparts, achieving ROUGE-1 scores of 0.3648 for recall and 0.5321 for precision, with a BLEU score of 0.1268. This demonstrates a marked improvement in the precision and efficiency of prescription information retrieval, offering an enhanced decision support tool for healthcare providers. This decision support system can reduce MEs and improve access to reliable prescription data. This LLM-powered chatbot has the potential to transform healthcare, especially in resource-constrained medical environments.