KG-Infused LLM for Virtual Health Assistant: Accelerated Inference and Enhanced Performance

Virtual Health Assistants (VHAs) represent a significant advancement in patient care, leveraging artificial intelligence to provide continuous support and interaction. Despite the integration of advanced Large Language Models (LLMs), which have greatly improved conversational capabilities, VHAs still struggle to fully replicate the nuanced expertise of human medical professionals. This limitation is primarily due to their reliance on broad training data, which can result in responses that lack the necessary reliability and contextual relevance required in health-care, where precision is paramount. To address this challenge, recent research has focused on integrating healthcare databases, such as the Unified Medical Language System (UMLS), a comprehensive biomedical knowledge source, to enhance the reasoning capabilities and reliability of LLMs. However, previous efforts have been impeded by high inference times and suboptimal performance on various evaluation metrics due to inefficient retrieval of pertinent information from extensive databases. In this study, we propose a novel methodology that involves constructing a Knowledge Graph (KG) within the Neo4j graph database using UMLS data to facilitate faster retrieval. This approach is augmented by employing named entity recognition techniques to accurately identify relevant entities and by applying learning-to-rank and semantic matching algorithms to effectively rank the retrieved information. We validated our approach using various LLMs, including GPT-3.5 Turbo, GPT-4, LLaMA-7b, and LLaMA-13b, across the BioASQ, MedicationQA, and ExpertQA datasets. Our experiments demonstrated a 23% improvement in ROUGE-L scores, an 8-10% improvement in BERTScores, and a 20-30% improvement in BLEU scores, achieving up to an 80% reduction in inference time.