대형언어모델의 의료분야 활용 - 의료접근성을 올리는 챗봇

English summary

The presentation explains the principles of large language models and their use in medical question answering, while examining the potential and limitations of chatbots intended to improve medical accessibility. Language models estimate the probability of the next token from preceding tokens and use transformer self attention and large scale self supervised pretraining. The presentation covers ChatGPT alignment, comparison with physician responses, ChatDoctor with retrieved medical knowledge, and the prospective development of multimodal generalist medical artificial intelligence.

Key points

• ChatGPT alignment is described as supervised fine tuning, reward model training, and reinforcement learning from human feedback.
• A 2023 JAMA Internal Medicine study compared ChatGPT with verified physician answers to 195 patient questions from Reddit AskDocs. Evaluators preferred chatbot responses in 78.6 percent of evaluations, and rated them higher for information quality and empathy. Chatbot responses averaged 211 words compared with 52 words for physician responses.
• ChatDoctor fine tuned LLaMA on approximately 100,000 patient physician dialogue pairs and augmented prompts with information retrieved from a database covering approximately 700 diseases and from Wikipedia. On the medical benchmarks described, it achieved higher Precision, Recall, F1, and BERTScore than vanilla ChatGPT and answered some questions involving knowledge after 2021. These are results reported in the presentation materials and do not establish clinical safety.

Audience

Graduate students, researchers, and clinical professionals in medicine and public health, particularly those studying community healthcare and digital health monitoring.

Scope and limitations

This summary is limited to the concepts, study designs, findings, and limitations contained in the presentation and the supplied evidence notes. Reddit and online consultation data do not provide the full electronic health record, verified history, physical examination, or laboratory monitoring available in clinical care. Chatbot responses were about four times longer than physician responses, which may have introduced a length bias, and empathy was assessed by healthcare professionals rather than patients. Open ended language models also lack deterministic systems that can reliably prevent hallucinations and medical errors. The systems discussed were not approved as medical devices, and evidence on effects on clinician burnout, patient adherence, and clinical outcomes remains lacking.

Research context

From a medical AI and data science perspective, the topic connects large language model question answering, external knowledge retrieval, multimodal data integration, and the validation challenges of generalist medical AI.