■ 学会名
2024 ISPE (International Society for Pharmacoepidemiology) Annual Meeting

■ 発表日
2024/08/27

■ 筆頭演者
Hisashi Urushihara
Department of Drug Development and Regulatory Science, Faculty of Pharmacy, Keio University

■ 共同演者
Shogo Yaegashi¹, Takayuki Ando², Takahiro Yako³, Nao Oishi³, Azusa Hara¹
1) Department of Drug Development and Regulatory Science, Faculty of Pharmacy, Keio University
2) Center for General Medicine Education, School of Medicine, Keio University
3) Graduate School of Science and Technology, Keio University　

■ 発表形態
Poster

■ 要旨
[Objective]
To construct a hypothesis-free approach to detecting safety signals by acquiring distributed representation of adverse events using a large-scale language model based on receipt data and clustering them.
[Methods]
The Bidirectional Encoder Representations from Transformer (BERT) was trained on claims data of patients treated with type 2 diabetes drugs from the claims databases provided by DeSC Healthcare and JMDC. The AE embeddings were clustered using the WARD method. The detected clusters were validated against their package inserts as signals.
[Results]
The number of detected clusters was 663, of which 21.6% contained adverse events listed in the package inserts.
[Conclusion]
The results suggest that this technique could be used for signal detection.