SeltaSquare

1. 
2. 

• 게시판 리스트 옵션
  • 검색

   · Background: The U.S. FDA is increasingly focused on tools that can automatically identify Adverse Drug Events (ADEs) within prescribing information. In collaboration with the MITRE Corporation, the FDA organized a joint initiative titled "Adverse Drug Event Evaluation (ADE Eval)" to determine whether current Natural Language Processing (NLP) algorithms are robust enough for real-world application.

   · Objective: The primary goal of the ADE Eval was to evaluate various NLP techniques for identifying Adverse Drug Reactions (ADRs) mentioned in publicly available FDA-approved drug labels. This study specifically aimed to assess the feasibility of models that could replicate and support pharmacovigilance workflows within the FDA.

   · Methodology: The research involved developing specific drug safety annotation guidelines and an annotated corpus. Two key metrics were used to model the FDA’s expertise: ① 

The algorithm’s ability to identify the correct MedDRA® terms within the annotated corpus. ② The quality of the extracted evidence, measuring the ability to correctly identify text segments supporting the selected MedDRA® terms. ③ An additional metric evaluated the correction cost of system outputs for subsequent iterative learning.

   · Results: A total of 13 teams participated with 23 submissions. The top-performing models achieved a MedDRA® coding F1-measure of 0.79, a quality score of 0.96, and an entity recognition (mention-finding) F1-measure of 0.89.

   · Conclusion: While current NLP technologies have not yet reached a level of performance allowing for unsupervised intervention, the study confirmed that NLP-generated outputs provide significant value when integrated into human-led pharmacovigilance workflows.

Ref.> Samuel Bayer, Oanh Dang, John Aberdeen, Sonja Brajovic, Kimberley Swank, Lynette Hirschman, Robert Ball. (Drug Safety, 2021)

https://link.springer.com/article/10.1007/s40264-020-00996-3