Graph neural networks for individual treatment effect estimation

• Verfasst von: Sirazitdinov, Andrei [VerfasserIn]
• Verfasst von: Buchwald, Marcus [VerfasserIn]
• Verfasst von: Heuveline, Vincent [VerfasserIn]
• Verfasst von: Hesser, Jürgen [VerfasserIn]
• Titel: Graph neural networks for individual treatment effect estimation
• Titelzusatz: methods
• Verf.angabe: Andrei Sirazitdinov, Marcus Buchwald, Vincent Heuveline, and Jürgen Hesser, (Member, IEEE)
• E-Jahr: 2024
• Jahr: 02 August 2024
• Umfang: 11 S.
• Fussnoten: Gesehen am 22.10.2024
• Titel Quelle: Enthalten in: Institute of Electrical and Electronics EngineersIEEE access
• Ort Quelle: New York, NY : IEEE, 2013
• Jahr Quelle: 2024
• Band/Heft Quelle: 12(2024), Seite 106884-106894
• ISSN Quelle: 2169-3536
• DOI: doi:10.1109/ACCESS.2024.3437665
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Causal inference
• Sach-SW: Computational modeling
• Sach-SW: Estimation
• Sach-SW: graph neural networks
• Sach-SW: Graph neural networks
• Sach-SW: individual treatment effect estimation
• Sach-SW: Reviews
• Sach-SW: Task analysis
• Sach-SW: Training
• Sach-SW: Vectors
• K10plus-PPN: 1906378681

Abstract

Individual treatment effect (ITE) estimation is an important task for personalized decision-making in clinical settings. However, the data used to train an ITE estimation model may be limited. In this case, we expect that information regarding causal connectivity within features can facilitate model training and thus lead to better predictions. In this study, we incorporated causal information about the connectivity within features represented by a Directed Acyclic Graph (DAG) into the problem of ITE estimation. For this purpose, we propose a novel method based on Graph Neural Networks (GNN). Our results show that the proposed approach performs comparably to the current state-of-the-art methods on existing datasets. Using an artificial dataset, we demonstrate the potential advantages of using real graphs responsible for the data generation process over empty graphs with no edges. These advantages are particularly evident for datasets with limited training sizes and correctly defined DAGs. These findings highlight the potential of GNNs in personalized medicine for improving the assessment of individual treatment effects.