PCN234 DEVELOPMENT AND VALIDATION STUDY OF ARTIFICIAL INTELLIGENCE SUPPORTING SYSTEMATIC LITERATURE REVIEW FOR CONDUCTING COST-EFFECTIVENESS ANALYSIS

Abstract

Systematic literature review is an essential step to identify available evidence for cost-effectiveness evaluations. Recent development of artificial intelligence (AI) may help such pre-work which usually takes time and effort. This study aimed to develop the AI on-board efficient and comprehensive literature search system to support cost-effectiveness analysis. We developed Named Entity (NE) extraction and document classifier engine and scoring engine with document classifier with AI. Firstly, we used NE extraction engine (Method-1) to extract specific terms as PICO (Patient, Intervention, Comparison, and Outcome) from the text data. Secondly, the terms labeled with the diseases and drugs were extracted, then screened the description with outcomes and costs. Thirdly, we used scoring engine with document classifier (Method-2) scoring the relevance to cost-effectiveness research by the specific sentences. Finally, the abstracts with higher total scores were selected. We used abstracts in PubMed database for the evaluation. Systematic literature search has conducted on the articles published until January 2017 on MEDLINE®, MEDLINE® In-Process, Embase® and Cochrane collaboration. PICO was settled with hepatocellular carcinoma (Patient), antineoplastic systemic therapy (Intervention/Comparators) for health economic and related studies (Outcomes). This manual work identified 22 health economic and related studies out of 3,884 records identified by search term algorisms. Method-1 identified 538 studies (include 2,690 expected identified words). The precision rate, recall rate and accuracy rate (F-measure) was 0.900, 0.820, and 0.859, respectively, compared with the studies identified by manual. Method-2 identified 10 studies (include 135 sentences). The precision rate, recall rate and accuracy rate was 0.810, 0.810, and 0.808, respectively, compared with the studies identified by manual. Our study suggested that AI have a progress of contributing systematic literature review required for cost-effectiveness analysis. Improving the precision and recall rates will allow us use the AI system widely to various diseases and drugs.