GPT와 BERTopic을 이용한 MAUDE (Manufacturer and User Facility Device Experience) 데이터베이스 내 의료기기 사용 오류 분석 = Analysis of Medical Device Use Error in the Manufacturer and User Facility Device Experience (MAUDE) Database with GPT and BERTopic

다국어 초록 (Multilingual Abstract)

Medical device use errors are crucial to note because they directly impact patient safety and the quality of medical services. Previous studies have primarily relied on manual reviews by experts to identify and analyze medical device use errors from the MAUDE database, which collects adverse event reports. To overcome these limitations, this study utilized the large language model GPT-3.5-turbo and advanced prompting techniques to automatically extract use errors from unstructured narrative data in the MAUDE database and performed topic analysis using BERTopic. This approach aimed to minimize subjectivity and inefficiency that might arise in previous studies and to more clearly identify patterns and major causes of medical device use errors. Applying this method to adverse event data for the 'Automatic Delivery Peritoneal System,' the main causes of use errors were found to be device setup errors and user interface issues, which were mainly attributed to insufficient user training. This study provides important foundational data that can contribute to medical device safety by systematically analyzing medical device use errors, offering a structured framework for design and manufacturing improvements, accident prevention and management, and enhancing the consistency and reproducibility of data analysis.

목차 (Table of Contents)

• I. 서론 1
• 1.1. 연구 배경 1
• 1.2. 연구 목적 및 내용 3
• II. 선행 연구 5
• 2.1. 대규모 언어 모델(Large Language Model, LLM) 5
• 2.2. 대규모 언어 모델(Large Language Model, LLM)의 발전과 한계점 10
• 2.3. 의료 분야에서 대규모 언어 모델(Large Language Model, LLM) 12
• 2.4. 프롬프트 엔지니어링 14
• 2.5. 토픽 모델링 17
• 2.6. 의료기기 사용 오류에 대한 선행 연구 21
• III. 연구 방법 29
• 3.1. 연구 프로세스 개요 29
• 3.2. 데이터 수집 및 전처리 30
• 3.3. 프롬프트 설계 32
• 3.4. 프롬프트 선택 41
• 3.5. GPT-3.5 turbo를 사용한 텍스트 처리 43
• 3.6. BERTopic을 사용한 토픽 모델링 45
• IV. 연구 결과 47
• 4.1. 프롬프트와 텍스트 처리에 대한 전문가 평가 결과 47
• 4.2. BERTopic을 활용한 토픽 모델링 결과 53
• V. 결론 및 논의 69
• Reference 71
• Appendix 79