Author(s)
Sakata Y1, Inoue K2, Nagasawa T3, Ooishi M3, Azuma M4, Kitabayashi H5, Kusaba S6, Tanaka R7, Nawata S8, Takizaki K9, Sudo M3, Okamoto R10, Abe R11, Nakatsui M10, Okuno Y10
1Eisai Co., Ltd., Tokyo, Japan, 2G-Search Limited, Tokyo, Japan, 3Fujitsu Limited, Osaka, Japan, 4Eisai Co,. Ltd., Tokyo, Japan, 5Kyowa Kirin Co., Ltd., Tokyo, Japan, 6Kusaba Consulting Office, Osaka, Japan, 7Chugai Pharmaceutical Co., Ltd., Tokyo, Japan, 8Kyorin Pharmaceutical Co. Ltd., Tokyo, Japan, 9Mitsubishi Tanabe Pharma Corporation, tokyo, Japan, 10Kyoto University, Kyoto, Japan, 11RIKEN, Tokyo, Japan
OBJECTIVES: Systematic literature review is required to identify literature information prior conducting cost-effectiveness analysis. Now that we have been developing an artificial intelligence (AI) for efficient literature search system, this study aimed to provide comprehensive reporting system by using further AI on-board natural language processing (NLP) of structuring sentences and their relationship. METHODS: We used Named Entity Extraction (NEE) engine and Relation Extraction (RE) engine developed in Fujitsu Limited to structure text data into table format. NEE engine is an AI engine that can extract specific terms as PICOs (Patient, Intervention, Comparison, and Outcome) from the text data. RE engine is another AI engine that can distinguish relationship between PICOs from the text data. An AI model was generated by training the appearance patterns of PICO words and the relationships in sentences. Our system create tables by arranging extracted PICO words and the relationships. RESULTS: We used 495 sentences in 45 abstracts of hepatocellular carcinoma searched from PubMed® for our evaluation. PICO words and the relationships extracted by two AI engines were evaluated by health technology assessment specialists. Targeted terms of the NEE engine were disease, intervention, outcome index, and the value, and the accuracy rate (F-value) was 0.860, 0.731, 0.767, and 0.827, respectively, compared with the evaluation by the specialists. Targeted relationships of RE engine were between disease and intervention, intervention and outcome index, and outcome index and the value. The F-value of RE engine was 0.872, 0.881, and 0.735, respectively. Based on those results, PICO relationships were arranged to make tables, which user could capture the contents visually. CONCLUSIONS: Our study suggested that AI could provide a support for summarizing the comprehensive data information based on the systematic literature review. Improving the precision and recall rates will allow us to use the AI system widely to various diseases and drugs.
Conference/Value in Health Info
2020-11, ISPOR Europe 2020, Milan, Italy
Value in Health, Volume 23, Issue S2 (December 2020)
Code
PCN267
Topic
Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Systems & Structure
Disease
Oncology
