Knowledge-Grounded Medical Dialogue Generation

With the introduction of the first FDA-approved treatment for Alzheimer’s disease (AD), genetic testing for APOE, a major genetic risk factor for AD, has become a critical step to assess treatment eligibility. In order to address the increasing volume of APOE testing, tools to help patients understand genetic risk factors and their implications are urgently needed. Conversational agents powered by large language models (LLMs) can help triage patients and supplement human counselors. However, deploying such agents poses challenges: institutional barriers prevent the input of clinical data, including protected health information (PHI), into commercial LLMs, LLMs potentially hallucinate critical medical facts, and LLMs should mimic the communication style of clinicians in order to be trusted supplements. We introduce a dual-method approach to enhance LLMs’ accuracy and clinical communication effectiveness. First, we build a knowledge bank of recorded patient-provider genetic counseling sessions and leverage an open-source LLM to extract and summarize relevant information. We leverage this knowledge bank to develop a retrieval-augmented system for answering patient questions. We find that responses generated from our pipeline are more readable and better resemble human responses compared to those directly from GPT-4, suggesting that this pipeline enhances both accuracy and a clinician-like tone of communication.