A structural characterization of shortcut features for prediction.
Causal inference
Machine learning
Prediction models
Journal
European journal of epidemiology
ISSN: 1573-7284
Titre abrégé: Eur J Epidemiol
Pays: Netherlands
ID NLM: 8508062
Informations de publication
Date de publication:
Jun 2022
Jun 2022
Historique:
received:
15
06
2022
accepted:
19
06
2022
pubmed:
7
7
2022
medline:
20
7
2022
entrez:
6
7
2022
Statut:
ppublish
Résumé
With the rising use of machine learning for healthcare applications, practitioners are increasingly confronted with the limitations of prediction models that are trained in one setting but meant to be deployed in several others. One recently identified limitation is so-called shortcut learning, whereby a model learns to associate features with the prediction target that do not maintain their relationship across settings. Famously, the watermark on chest x-rays has been demonstrated to be an instance of a shortcut feature. In this viewpoint, we attempt to give a structural characterization of shortcut features in terms of causal DAGs. This is the first attempt at defining shortcut features in terms of their causal relationship with a model's prediction target.
Identifiants
pubmed: 35792990
doi: 10.1007/s10654-022-00892-3
pii: 10.1007/s10654-022-00892-3
pmc: PMC9256901
doi:
Types de publication
Journal Article
Langues
eng
Sous-ensembles de citation
IM
Pagination
563-568Subventions
Organisme : NHLBI NIH HHS
ID : K01 HL141771
Pays : United States
Informations de copyright
© 2022. Springer Nature B.V.
Références
JAMA. 2016 Dec 13;316(22):2402-2410
pubmed: 27898976
Nature. 2015 May 28;521(7553):436-44
pubmed: 26017442
Nat Biomed Eng. 2018 Oct;2(10):719-731
pubmed: 31015651
Spine J. 2021 Oct;21(10):1606-1609
pubmed: 32858170
IEEE Trans Pattern Anal Mach Intell. 2013 Aug;35(8):1798-828
pubmed: 23787338
NPJ Digit Med. 2019 Apr 30;2:31
pubmed: 31304378
JAMA. 2018 Apr 3;319(13):1317-1318
pubmed: 29532063
medRxiv. 2020 Oct 07;:
pubmed: 32995822
Nature. 2017 Jun 28;546(7660):686
pubmed: 28658222