How to control for confounds in decoding analyses of neuroimaging data

Snoek, Lukas; Miletić, Steven; Scholte, H. Steven

doi:10.1016/j.neuroimage.2018.09.074

Cited by 113 publications

(145 citation statements)

References 62 publications

Supporting

Mentioning

138

Contrasting

Order By: Relevance

“…Deconfounding approaches used in standard GLM-based analysis must be adapted to out-sample data, and can easily lead to pessimistic evaluations [7]. Indeed, when there is shared signal between the effect of interest and the confound in the brain data, deconfounding fully removes this signal and makes it impossible to build a prediction from it.…”

Section: Discussionmentioning

confidence: 99%

“…One solution is to apply deconfounding jointly on the train and the test set, but it breaks the statistical validity of cross-validation [11] by coupling the train and the test set, hence it gives biased results. Another solution is to apply deconfounding separately on the train and the test set, but this can be very pessimistic [7]. Indeed the shared variance removed is then an in-sample property, and grows larger in an uncontrolled way for small test sets.…”

Section: Method: Anti Mutual-information Subsampling a Formalizimentioning

confidence: 99%

“…Second, it is designed to control in-sample properties, while predictive models are designed for out-of-sample prediction. A two step approach based on applying a classical GLM to deconfound as a first step followed by a predictive model is often pessimistic, leading to below-chance prediction [2], [7].…”

Section: Introduction: Confounded Predictionsmentioning

confidence: 99%

See 2 more Smart Citations

Controlling a confound in predictive models with a test set minimizing its effect

Chyzhyk

Varoquaux

Thirion

et al. 2018

2018 International Workshop on Pattern Recognition in Neuroimaging (PRNI)

View full text Add to dashboard Cite

Predictive models applied on brain images can extract imaging biomarkers of pathologies or psychological traits. Yet, a successful prediction may be driven by a confounding effect that is correlated with the effect of interest. For instance fluid intelligence is strongly impacted by age; age is well predicted from brain images; hence successful prediction of fluid intelligence from brain images might have captured nothing more than a biomarker of aging. Here we introduce a nonparametric approach to control for a confounding effect in a predictive model. It is based on crafting a test set on which the effect of interest is independent from the confounding effect. We name this strategy "anti mutual-information subsampling". We demonstrate the approach with a large sample resting-state fMRI and psychometric data of healthy aging subjects (n = 608). We show that using a linear model to remove the effect of age on the brain signals ("deconfounding") leads to pessimistic scores, as previously reported. Anti mutual-information subsampling does not require to remove from the brain signals the shared variance between aging and fluid intelligence, and hence does not display this pessimistic behavior. In addition, it is non-parametric and hence robust to violations of the linear hypothesis.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Method: Anti Mutual-information Subsampling a Formalizimentioning

confidence: 99%

See 1 more Smart Citation

Controlling a confound in predictive models with a test set minimizing its effect

Chyzhyk

Varoquaux

Thirion

et al. 2018

2018 International Workshop on Pattern Recognition in Neuroimaging (PRNI)

View full text Add to dashboard Cite

show abstract

“…In keeping with the recommended strategy (48), both the symptom dimensional-scores and rs-FC features were adjusted for confounding effects of age, gender, site, and head-motion (DVARS). To avoid data-leakage within cross-validation (49), confound regression models were learned only on the training-set and then applied on both training and test data (50,51). The RVM model was then trained on confound-adjusted training data and applied to the confound-adjusted held-out test data.…”

Section: Prediction Of Symptom Dimensions Using Network Rs-fcmentioning

confidence: 99%

Connectivity patterns of task-specific brain networks allow individual prediction of cognitive symptom dimension of schizophrenia and link to molecular architecture

Chen

Mueller

Dukart

et al. 2020

Preprint

View full text Add to dashboard Cite

AbstractBackgroundDespite the marked inter-individual variability in the clinical presentation of schizophrenia, it remains unclear the extent to which individual dimensions of psychopathology may be reflected in variability across the collective set of functional brain connections. Here, we address this question using network-based predictive modeling of individual psychopathology along four data-driven symptom dimensions. Follow-up analyses assess the molecular underpinnings of predictive networks by relating them to neurotransmitter-receptor distribution patterns.MethodsWe investigated resting-state fMRI data from 147 schizophrenia patients recruited at seven sites. Individual expression along negative, positive, affective, and cognitive symptom dimensions was predicted using relevance vector machine based on functional connectivity within 17 meta-analytic task-networks following a repeated 10-fold cross-validation and leave-one-site-out analyses. Results were validated in an independent sample. Networks robustly predicting individual symptom dimensions were spatially correlated with density maps of nine receptors/transporters from prior molecular imaging in healthy populations.ResultsTen-fold and leave-one-site-out analyses revealed five predictive network-symptom associations. Connectivity within theory-of-mind, cognitive reappraisal, and mirror neuron networks predicted negative, positive, and affective symptom dimensions, respectively. Cognitive dimension was predicted by theory-of-mind and socio-affective-default networks. Importantly, these predictions generalized to the independent sample. Intriguingly, these two networks were positively associated with D1 dopamine receptor and serotonin reuptake transporter densities as well as dopamine-synthesis-capacity.ConclusionsWe revealed a robust association between intrinsic functional connectivity within networks for socio-affective processes and the cognitive dimension of psychopathology. By investigating the molecular architecture, the present work links dopaminergic and serotonergic systems with the functional topography of brain networks underlying cognitive symptoms in schizophrenia.

show abstract

“…//github.com/lukassnoek/MVCA), as described in the author's article74 . Responses were z-scaled using training set means and standard deviations.…”

mentioning

confidence: 99%

Functional and diffusion MRI reveal the functional and structural basis of infants’ noxious-evoked brain activity

Baxter

Moultrie

Fitzgibbon

et al. 2020

Preprint

View full text Add to dashboard Cite

Understanding the neurophysiology underlying pain perception in infants is central to improving early life pain management. In this multimodal MRI study, we use resting-state functional and white matter diffusion MRI to investigate individual variability in infants’ noxious-evoked brain activity. In an 18-infant nociception-paradigm dataset, we show it is possible to predict infants’ cerebral haemodynamic responses to experimental noxious stimulation using their resting-state activity across nine networks from a separate stimulus-free scan. In an independent 215-infant Developing Human Connectome Project dataset, we use this resting-state-based prediction model to generate noxious responses. We identify a significant correlation between these predicted noxious responses and infants’ white matter mean diffusivity, and this relationship is subsequently confirmed within our nociception-paradigm dataset. These findings reveal that a newborn infant’s pain-related brain activity is tightly coupled to both their spontaneous resting-state activity and underlying white matter microstructure. This work provides proof-of-concept that knowledge of an infant’s functional and structural brain architecture could be used to predict pain responses, informing infant pain management strategies and facilitating evidence-based personalisation of care.

show abstract

How to control for confounds in decoding analyses of neuroimaging data

Cited by 113 publications

References 62 publications

Controlling a confound in predictive models with a test set minimizing its effect

Controlling a confound in predictive models with a test set minimizing its effect

Connectivity patterns of task-specific brain networks allow individual prediction of cognitive symptom dimension of schizophrenia and link to molecular architecture

Functional and diffusion MRI reveal the functional and structural basis of infants’ noxious-evoked brain activity

Contact Info

Product

Resources

About