2023 Fiscal Year Final Research Report

Integrated AI analysis of natural language in nursing records by GPT and sensor data to support ward management

Research Project

PDF

Project/Area Number	22K19684
Research Category	Grant-in-Aid for Challenging Research (Exploratory)
Allocation Type	Multi-year Fund
Review Section	Medium-sized Section 58:Society medicine, nursing, and related fields
Research Institution	Tokyo University of Science (2023) The University of Tokyo (2022)
Principal Investigator	MORI Taketoshi 東京理科大学, 先進工学部機能デザイン工学科, 教授 (20272586)
Co-Investigator(Kenkyū-buntansha)	松原仁東京大学, 大学院情報理工学系研究科, 教授 (50325883) 武村雪絵東京大学, 医学部附属病院, 看護部長 (70361467) 野口博史大阪公立大学, 大学院工学研究科, 教授 (50431797)
Project Period (FY)	2022-06-30 – 2024-03-31
Keywords	看護記録 / SOAP / カルテ / ナースコール / バイタルセンサ / 病棟管理 / 生成AI / ChatGPT
Outline of Final Research Achievements	We aimed to construct a method for extracting and modeling data from electronic medical records, nursing worksheets, nurse call systems, and other systems for investigation and analysis of care. Specifically, we applied GPT-3 and GPT-2 to the natural language model of nursing records recorded by nurses in hospital wards, and constructed a method for estimating and predicting the state of patients and wards from nursing texts, as well as a system for interpolating and generating the continuation of nursing texts when part of a nursing text is described, and for pointing out the parts that are likely to be described in error. The system is composed of the GPT-4 (GPT-2) and the GPT-3 (GPT-3). The system was extended to be applicable to GPT-4 (ChatGPT) API, and showed that GPT-4 and 2023 generation AIs (Gemini, etc.) are more useful than previous generation models, including Japanese-language versions.
Free Research Field	看護理工学
Academic Significance and Societal Importance of the Research Achievements	本研究で看護記録や温度板・カルテ記述を分析するのに用いるGPTはBertと並ぶ文書分類，翻訳，質問応答などの自然言語処理の人工知能の最先端の一つである．これを看護ケア分野で開発導入することは，看護記録の計画，実践についての文章やメモ書き，カルテ上の箇条書きなど，正解アノテーションをつけた高品質の学習用テキストデータを得られない臨床現場の情報につき，この別ドメイン（別領域）の膨大なデータに基づく学習モデルを基に追加拡張できる可能性や，メモから構造的な文章を生成できるといった優秀な特長の有用性を示せ，機微情報を投入しない学習の検証の一つとなる。