% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Komeyer:1034775,
author = {Komeyer, Vera and Eickhoff, Simon and Rathkopf, Charles and
Grefkes, Christian and Patil, Kaustubh and Raimondo,
Federico},
title = {{C}orrect deconfounding enables causal machine learning for
precision medicine and beyond},
reportid = {FZJ-2024-07529},
year = {2024},
abstract = {Artificial intelligence holds promise for individualized
medicine. Yet, predictive models in the neurobiomedical
domain suffer from a lack of generalizability and
replicability so that transitioning models from prototyping
to clinical applications still poses challenges. Key
contributors to these challenges are confounding effects; in
particular the oftentimes purely statistical perspective on
confounding. However, complementing these statistical
considerations with causal reasoning from domain knowledge
can make predictive models a tool for causal biomedical
inference beyond associative insights. Such causal insights
give answers to biomedical questions of how and why,
arguably what most biomedical investigations ultimately seek
for. Here, we suggest a 5-step approach for targeted,
context-informed deconfounding. We exemplify the 5-step
approach with a real-world neurobiomedical predictive task
using data from the UK Biobank. The core of this approach
constitutes a bottom-up causal analysis to identify a
correct set of deconfounders and the appropriate
deconfounding method for a given causal predictive
endeavour. Using the 5-step approach to combine causal with
statistical confounder considerations can make predictive
models based on observational (big) data a technique
comparable to Randomized Control Trials (RCTs). Through
causally motivated deconfounding we aim at facilitating the
development of reliable and trustworthy AI as a medical
tool. In addition, we aim to foster the relevance of low
performing or even null result models if they originate from
a “skilful interrogation of nature”, i.e. a
deconfounding strategy derived from an adequate causal and
statistical analysis. Ultimately, causal predictive
modelling through appropriate deconfounding can contribute
to mutual recursive feedback loops of causal insights across
disciplines, scales and species that enable the field to
disentangle the cause-effect structure of neurobiomedical
mechanisms.},
cin = {INM-7 / INM-3},
cid = {I:(DE-Juel1)INM-7-20090406 / I:(DE-Juel1)INM-3-20090406},
pnm = {5254 - Neuroscientific Data Analytics and AI (POF4-525) /
5251 - Multilevel Brain Organization and Variability
(POF4-525) / DFG project G:(GEPRIS)431549029 - SFB 1451:
Schlüsselmechanismen normaler und krankheitsbedingt
gestörter motorischer Kontrolle (431549029)},
pid = {G:(DE-HGF)POF4-5254 / G:(DE-HGF)POF4-5251 /
G:(GEPRIS)431549029},
typ = {PUB:(DE-HGF)25},
doi = {10.1101/2024.09.20.24314055},
url = {https://juser.fz-juelich.de/record/1034775},
}