Related Experiment Video
Updated: Mar 8, 2026
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
16.4K
Predictive modelling using neuroimaging data in the presence of confounds
Anil Rao1, Joao M Monteiro1, Janaina Mourao-Miranda1
1Department of Computer Science, University College London, United Kingdom; Max Planck University College London Centre for Computational Psychiatry and Ageing Research, University College London, London, United Kingdom.
Neuroimage
|February 2, 2017
Summary
This study defines and evaluates methods for handling confounding variables in neuroimaging predictive models. Current approaches, including instance weighting, did not improve prediction accuracy over a baseline model that ignored confounding factors.
Area of Science:
- Neuroimaging
- Machine Learning
- Biostatistics
Background:
- Predictive models trained on neuroimaging data often encounter confounding variables (e.g., age, gender) that influence imaging data but are not of primary clinical interest.
- Confounding variables can bias training samples, creating an association between the confound and target variable that differs from the target population.
- Accurate predictive models are crucial for clinical applications, necessitating effective strategies to mitigate confounding effects.
Purpose of the Study:
- To define 'confound' in the context of neuroimaging predictive modeling.
- To evaluate standard and novel methods (Instance Weighting) for addressing confounding in predictive models.
- To compare the performance of these methods against a baseline model in real-world neuroimaging datasets.
Main Methods:
- Defined a confound as a variable affecting imaging data and having a differing association with the target variable in the sample versus the target population.
- Implemented and assessed standard approaches: image adjustment and including the confound as a predictor.
- Developed and evaluated a novel Instance Weighting scheme to optimize model training for the target population.
Main Results:
- No confounding adjustment method improved prediction accuracy compared to a baseline model that ignored confounding.
- Including the confound as a predictor resulted in less accurate models than the baseline.
- Different methods focused predictions on specific population subsets, and predictive accuracy was higher when no confounding was present.
Conclusions:
- Current methods for handling confounding in neuroimaging predictive modeling do not consistently improve accuracy over ignoring confounds.
- Instance Weighting and other adjustment methods show potential for focusing predictions on specific population subsets.
- Further research is needed to optimize predictive models for clinical practice by effectively addressing confounding variables.
