Confound Controlled Multimodal Neuroimaging Data Fusion and Its Application to Developmental Disorders

May combine two claims.

When confounding covariates are not addressed, neuroimaging analyses can yield inaccurate descriptions of brain-clinical relationships and spurious associations.

“A key challenge of brain data analysis is how to handle confounds, which, if unaddressed, can lead to an unrealistic description of the relationship between the brain and clinical measures.”

• Also stated in: § INTRODUCTION · hedged

  “Spurious associations between variables of interest, inaccurate conclusions, and reduced model performance may occur when confounding covariates are not dealt with correctly in neuroimaging studies [9].”

• cites [8] Pourhoseingholi, 2012, How to control confounding effects by statistical analysis, Gastroenterol. Hepatol. Bed BenchPartly supported

  Section 1 says an unaddressed confounder can distort an observed association so that results do not reflect the actual relationship between variables. The paper discusses epidemiological studies, however, and does not establish the claim specifically for neuroimaging or brain–clinical relationships. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [9] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

mCCAR+jICA can identify co-varying multimodal brain patterns significantly associated with cognition measures in psychiatry.

“In particular, the supervised fusion model of mCCA with reference + joint independent components analysis (mCCAR+jICA), can identify co-varying multimodal brain patterns significantly associated with cognition measures in psychiatry [37, 38].”

• cites [37] sourceSupported
• cites [38] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Sex is often a confounding factor in neuroimaging due to sexual dimorphisms in brain structural lifespan trajectories, and analyses may need to be conducted by sex.

“Due to potential sexual dimorphisms in brain structural lifespan trajectories, sex is often considered a confounding factor in neuroimaging studies and individual analyses need to be conducted by sexes (e.g., depression, autism) [42, 43].”

• cites [42] sourceNot checked — no full text of the source
• cites [43] sourcePartly supported

  The Methods (page unavailable) analyze autism-related gene expression separately in female- and male-brain samples, supporting the relevance of sex-specific analysis in autism. The paper does not establish that sex confounds neuroimaging because of differences in brain structural lifespan trajectories, or that neuroimaging analyses generally need to be conducted separately by sex. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

May combine two claims.

Existing fusion methods do not jointly optimize nuisance covariate removal and multimodal fusion; they usually regress covariates out before fusion.

“Nevertheless, the existing fusion methods do not jointly optimize for nuisance covariates and multimodal fusion. Rather, a linear regression model is typically used to remove the covariate-related information from each modality feature matrix prior to fusion analyses, aiming to remove the potential confounding effects [10].”

• cites [10] sourcePartly supported

  Section 1.2.5 describes usually linear regression to remove confound-related variance from neuroimaging features, and section 4.3 recommends doing this first in a decoding pipeline. The paper does not establish that existing multimodal fusion methods typically use this approach or that they fail to jointly optimize fusion and confound removal. Page numbers were not available in our copy. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

Age and head motion (FD) as confounding covariates are associated with worse model accuracy in functional connectivity-based neuroimaging classification.

“A recent Nature paper reported that the confounding covariates, especially age and head motion (FD), were associated with worse model accuracy [11] in functional connectivity based neuroimaging classification.”

• cites [11] sourceSupported

Classified as probably not a substantial claim. May combine two claims.

ADHD and ASD involve structural and functional brain abnormalities.

“Previous studies found both structural and functional brain abnormalities in ADHD and ASD [15, 16].”

• cites [15] sourcePartly supported

  The Results section reports structural brain differences in both ASD and ADHD, including gray-matter volume differences. The review focuses on structural imaging; the passages found do not establish the claimed functional brain abnormalities. No page numbers are available in our copy. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [16] sourcePartly supported

  The Abstract and Introduction report atypical inhibition-related brain activation in ADHD and ASD, supporting the functional part of the claim. The paper does not report structural brain abnormalities in both conditions; its reported comparisons concern brain activity during an inhibition task. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

The two-step pre-fusion regression approach does not guarantee a joint optimum and may worsen fusion or regression performance.

“The two-step process does not guarantee joint optimum and may result in worse performance in either fusion or regression procedures [10, 11].”

• cites [10] sourcePartly supported

  The Discussion reports that whole-dataset confound regression produced negatively biased, sometimes below-chance decoding performance. That supports a narrower concern about performance after regression, but the paper does not address pre-fusion regression, a joint optimum, or performance of fusion and regression procedures themselves. Page numbers were not available in our copy. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [11] sourceCould not verify

  The Methods sections on phenotype classification and investigating potential confounds describe classification and regression steps, but do not discuss pre-fusion regression, joint optimization, or a trade-off between fusion and regression performance. Searches for the claimed two-step process and for pre-fusion regression found no passage addressing that point. Page numbers were not available in our copy. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

May combine two claims.

Pre-fusion exclusion of a biased subpopulation creates an artificial construct that fails to capture true population-level associations and reduces sample size and generalizability, especially in small samples.

“However, the pre-fusion exclusion of biased subpopulation creates an artificial construct that fails to capture the true population-level associations, while diminishing available sample size and data utility in neuroimaging research, thereby severely limiting the generalizability of findings, particularly in small-sample studies.”

• cites [13] sourcePartly supported

  The Discussion section ('A sampling view on confounds'; page not available) says matching and restriction can make it hard to collect large samples, supporting the sample-size concern. It does not discuss pre-fusion exclusion or say that excluding a biased subpopulation creates an artificial construct, loses true population-level associations, or severely limits generalizability in small studies. Instead, it describes sampling an ideal subpopulation as a strategy for controlling confounds. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

May combine two claims.

Salience network dysfunction in ASD may be related to abnormal sensory processing and may secondarily induce social deficits.

“SAN dysfunction in individuals with ASD might be related to abnormal sensory processing and may secondarily induce social deficits [59].”

• cites [59] sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Enlarged fronto-temporal GMV in ASD has been reported in previous studies.

“Enlarged fronto-temporal GMV in ASD was reported in previous studies [61], which is consistent with our results.”

• cites [61] sourceSupported

May combine two claims.

CR-mCCAR fusion of functional and structural neuroimaging data identifies ADHD-related co-varying patterns in striato-thalamo-cortical and salience areas that link with core symptoms and are uncorrelated with age and motion.

“Functional and structural neuroimaing data fusion reveals co-varying patterns in attention deficit/hyperactivity disorder (ADHD, striato-thalamo-cortical and salience areas) and in autism spectrum disorder (ASD, salience and fronto-temporal areas) that link with core symptoms but uncorrelate with age and motion.”

• AssumptionThe authors state this as an assumption: Each multimodal dataset can be modeled as a linear mixture of sources and a nonsingular mixing matrix.

May combine two claims.

In simulation, performance was sensitive to the λ2/λ1 ratio before reaching the optimal value; with λ1 fixed, larger λ2 gave better performance, and the optimum was at λ1=0.1 and λ2=0.5 for both modalities.

“Results show a clear sensitivity to the λ2λ1 before reaching the optimal value. Specifically, the larger the λ2 the better the performance when λ1 was fixed and close to optimal at λ1=0.1 and λ2=0.5λ2λ1=5 for both two modalities (Fig. 4).”

May combine two claims.

In ADHD-200, the thalamus showed higher fALFF and lower GMV in ADHD compared with healthy controls.

“Both fMRI_IC and sMRI_IC abnormalities are observed in the prefrontal cortex (PFC, lower fALFF and GMV in ADHD), thalamus (higher fALFF and lower GMV in ADHD) and middle occipital gyrus (MOG, lower fALFF and GMV in ADHD).”

May combine two claims.

In simulation, CR-mCCAR achieved nearly the best mixing coefficient and spatial map estimations, comparable to mCCAR, and was robust to noise.

“It is clear that CR-mCCAR achieves nearly the best mixing coefficient and spatial map estimations (comparable to mCCAR) while being quite robust to noise.”

May combine two claims.

In ADHD-200, the cortico-striato-thalamo-cortical and salience networks identified from multimodal MRI were associated with inattention and uncorrelated with age.

“The cortico-striato-thalamo-cortical (CSTC, including PFC, striatum and thalamus) and salience network (SAN, including ACC and insula) identified from multimodal MRI were associated with inattention while uncorrelated with the covariate age.”

May combine two claims.

CR-mCCAR fusion identifies ASD-related co-varying patterns in salience and fronto-temporal areas that link with core symptoms and are uncorrelated with age and motion.

“Functional and structural neuroimaing data fusion reveals co-varying patterns in attention deficit/hyperactivity disorder (ADHD, striato-thalamo-cortical and salience areas) and in autism spectrum disorder (ASD, salience and fronto-temporal areas) that link with core symptoms but uncorrelate with age and motion.”

• AssumptionThe authors state this as an assumption: Each multimodal dataset can be modeled as a linear mixture of sources and a nonsingular mixing matrix.

May combine two claims.

In ADHD-200, target components from mCCAR correlated with age, whereas those from CR-mCCAR did not correlate with age.

“The target components identified by mCCAR are correlated with the covariate, age, but those from CR-mCCAR do not correlate with age (Fig. 5(d), red frame).”

May combine two claims.

In ADHD-200, target components identified by both mCCAR and CR-mCCAR preserved group differences between ADHD and healthy controls and correlations with inattention scores.

“The target components identified by mCCAR and CR-mCCAR both preserve group differences between ADHD and HC (Fig. 5(b)), as well as correlation with inattention scores (Fig. 5(c)).”

May combine two claims.

In simulated data, the associations between components and the reference and the independence between target components and the covariate found in the training set did not generalize to the test set.

“Results show that the associations between the components and the reference and the independence between the target components and the covariate identified in training set cannot be generalized to the test set (Supplementary Table 3).”

May combine two claims.

In ABIDE II, target components from mCCAR were significantly correlated with FD, whereas those from CR-mCCAR were not correlated with FD.

“The target components identified by mCCAR are significantly correlated with the covariate, FD, but those from CR-mCCAR do not correlate with FD (Fig. 6(d), red frame).”

May combine two claims.

Classification accuracy increased by about 2% from mCCAR to CR-mCCAR with comparable standard deviation: ADHD vs HC from 77.9% ± 0.010 to 80.3% ± 0.012, and ASD vs HC from 76.7% ± 0.012 to 78.2% ± 0.008.

“There is a 2% increase of the classification accuracy from mCCAR to CR-mCCAR, with comparable standard deviation (ADHD vs. HC: from 77.9% ± 0.010 to 80.3% ± 0.012; ASD vs. HC: from 76.7%±0.012 to 78.2%±0.008, accuracy ± standard deviation, Fig. 9).”

May combine two claims.

CR-mCCAR had a lower mean absolute precision deviation than mCCAR for distinguishing ADHD or ASD from healthy controls, though the group differences were not significant (ADHD p=0.42; ASD p=0.07).

“We also compared the group difference in absolute error across 100 folds between the two methods and found that CR-mCCAR exhibits a lower mean absolute precision deviation for distinguishing ADHD/ASD from HC compared to mCCAR, although no significant group differences are observed (ADHD: p=0.42 and ASD: p=0.07, Supplementary Fig. 6).”

May combine two claims.

Joint multimodal features achieved higher classification accuracy and stability than single-modality features for both ADHD vs HC (over 4.8% improvement) and ASD vs HC (over 3.7% improvement).

“Moreover, comparing with features of single-modality, the joint multimodal features achieved higher classification accuracy and stability for both ADHD vs. HC (with over 4.8% improvement) and ASD vs. HC (with over 3.7% improvement, Supplementary Fig. 7).”

May combine two claims.

In human data, regressing out covariates before fusion led to substantial information loss of about 25-50% for linear regression and about 28-54% for nonlinear regression, which was significantly higher than in simulation.

“Results show that covariate regression prior to fusion leads to substantial information loss, around 25–50% (Fig. 7(a), around 28–54% for nonlinear regression (Supplementary Fig. 3(b-c)), which is consistent with and significantly higher than the simulated results (Fig. 3(c)).”

May combine two claims.

In CR-mCCAR, loss loadings correlated with the covariate but not with the reference.

“The loss loadings (DmCCAR,CR−mCCAR) are correlated with the covariate and not correlated with the reference in CR-mCCAR (Fig. 7(c)).”

May combine two claims.

In independent validation on ABIDE I, projected loadings of target components were correlated with SRS scores and uncorrelated with FD.

“We observe that the resulting projected loadings of the target components in ABIDE I are indeed correlated with SRS scores while being uncorrelated with the covariate FD (Supplementary Fig. 5), demonstrating the robustness and generalization ability of CR-mCCAR, capturing features that preserve their relation to the reference and absence of covariate effects even in out of sample data.”

May combine two claims.

In multi-reference fusion on ABIDE II, identified target components correlated with SRS sub-domains and did not correlate with covariates.

“The identified target components are correlated with the sub-domains of SRS (Fig. 10(b)) and not correlated with covariates (Fig. 10(c)).”

May combine two claims.

In multi-reference fusion on ABIDE II, the identified spatial maps had 98.6% overlap for fALFF and 98.0% overlap for GMV with single-reference fusion.

“Fig. 10(a) displays the identified spatial maps of single-reference fusion (the total SRS scores as reference, Fig. 6(a)) and multi-references fusion (the SRS 5 sub-domains scores as references) with 98.6% and 98.0% spatial overlap for fALFF and GMV, respectively.”

May combine two claims.

In ABIDE II, target components from both mCCAR and CR-mCCAR preserved group differences between ASD and healthy controls and correlations with SRS scores.

“The target components of each modality identified by mCCAR and CR-mCCAR both preserve group differences between ASD and HC (Fig. 6(b)), as well as correlation with the SRS scores (Fig. 6(c)).”

May combine two claims.

CR-mCCAR's computation time and memory usage increase linearly with the number of components and dataset size, and it can run on a conventional desktop.

“Finally, despite the fact that adding constraints to the covariate may increase computational complexity compared to mCCAR, the proposed method exhibits a linear increase in computation time and memory usage with the larger number of components and bigger dataset size, and can be run on a conventional desktop (see Supplementary Table 14-15).”

May combine two claims.

In R+mCCAR, loss loadings for each modality correlated with the reference but not with the covariate.

“The loss loadings (DmCCAR,R+mCCAR, the product of feature matrix Xi and the generalized inverse matrix of the loss spatial map Closs=CmCCAR−CR+mCCAR) for each modality are correlated with the reference in R+mCCAR, but not correlated with the covariate (Fig. 7(b)).”

May combine two claims.

In ABIDE II, the salience network and fronto-temporal regions identified from multimodal MRI were associated with symptom severity and uncorrelated with FD.

“SAN and fronto-temporal identified from multimodal MRI were associated with symptom severity while uncorrelated with the covariate FD.”

May combine two claims.

In ABIDE II, both fMRI and sMRI components identified by the proposed method showed abnormalities in the middle occipital gyrus: lower fALFF and higher GMV in ASD.

“Both fMRI_IC and sMRI_IC identified by our proposed method find abnormalities in the MOG (lower fALFF value and higher GMV in ASD).”

May combine two claims.

Existing fusion methods are unsupervised and fail to capture brain patterns related to a specific measure of interest.

“Although the above fusion methods can optimize joint source separation among multiple modalities in different aspects, they are unsupervised, without introducing any prior reference, which fails in capturing brain patterns related to a specific measure of interest.”

May combine two claims.

CR-mCCAR identifies one or more target components significantly correlated with the reference while uncorrelated with the covariate.

“Eventually, one or more components are identified as target components (here the 2nd component), which is significantly correlated with the reference while uncorrelated with the covariate.”

• AssumptionThe authors state this as an assumption: Each multimodal dataset can be modeled as a linear mixture of sources and a nonsingular mixing matrix.

Probably not a substantial claim: may not pass S3 (load-bearing).

CR-mCCAR can be extended to include multiple targets and multiple covariates.

“CR-mCCAR can be extended to include multiple targets and multiple covariates.”

• Also stated in: § Extension into Several References with Multiple Covariates · plain

  “Moreover, the proposed CR-mCCAR can be extended to multiple references with multiple covariates.”

• AssumptionThe authors state this as an assumption: Each multimodal dataset can be modeled as a linear mixture of sources and a nonsingular mixing matrix.

May combine two claims.

fALFF and GMV are representative measures of fMRI and sMRI, reflecting local spontaneous neuronal activity and morphological changes, and have been widely used to study neurodevelopmental disorders.

“Fraction amplitude of low frequency fluctuation (fALFF) and gray matter volume (GMV) are the two representative measures of fMRI and sMRI, reflecting local spontaneous neuronal activity and morphological changes of brain, respectively, and has been widely used to study various neurodevelopmental disorders [50, 51].”

• cites [50] sourceNot checked — no full text of the source
• cites [51] sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Covariates are a major concern in neuroimaging-based brain disorder studies and need careful handling in multimodal fusion.

“Thus, covariate is a major concern in the study of neuroimaging-based brain disorders that needs to be carefully addressed in multimodal fusion.”

Classified as probably not a substantial claim. May combine two claims.

Brain structural atrophy related to aging may confound biomarker identification, so identified biomarkers may reflect age-related atrophy rather than disease.

“Previous studies suggest that brain structural atrophy related to aging may result in confounding effects, especially in brain disorder biomarker identification. This can cast doubt on the results since the identified biomarkers may be descriptive of just age-related structural atrophy instead of the brain disease itself [41].”

• cites [41] Smith, 2018, Statistical challenges in, Big Data Human Neuroimaging, NeuronNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

None of the above fusion methods controlled for confounding covariates during fusion, which may lead to spurious associations between identified brain patterns and variables of interest, especially in supervised fusion.

“However, none of the above methods have controlled for confounding covariates during the fusion process, which may lead to spurious associations between the identified brain patterns and variables of interest, especially in supervised fusion.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The salience network plays a critical role in attentional capture of biological and cognitively relevant events.

“SAN plays a critical role in attentional capture of biological and cognitively relevant events [56], and the atypical activation patterns (primarily hypoactivation) of these regions have been extensively reported in ADHD [57, 58].”

• cites [56] sourceNot checked — no full text of the source

May combine two claims.

The classification results demonstrate the importance of multimodal fusion, the necessity of removing covariate effects during fusion, and the superior performance and reliability of CR-mCCAR compared with TMC in brain disorder diagnosis.

“The above results further demonstrate the importance of multimodal fusion, the necessity to remove the covariate effects during fusion and the superior performance and reliability of CR-mCCAR compared with TMC in brain disorder diagnosis.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study only included fMRI and sMRI data, so other modalities such as DTI were not combined.

“One possible limitation of this work is that only fMRI and sMRI data were included in this study. Other neuroimaging modalities, such as diffusion tensor imaging (DTI), can also be combined to provide an even richer understanding of brain patterns in ADHD and ASD.”

May combine two claims.

The cortico-striato-thalamo-cortical and salience networks are two major brain networks underlying the mechanisms of symptom severity in ADHD.

“These suggest that CSTC and SAN are two major brain networks underlying the mechanisms of symptom severity for ADHD.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Fronto-temporal dysconnectivity is considered a mechanism contributing to core behavioral impairments in ASD, especially social communication.

“Also, fronto-temporal dysconnectivity has been considered a mechanism contributing to the core behavioral impairments in ASD, especially social communication [60].”

• cites [60] sourceNot checked — no full text of the source

May combine two claims.

CR-mCCAR is robust and generalizes to out-of-sample data, capturing features that preserve their relation to the reference and lack covariate effects.

“We observe that the resulting projected loadings of the target components in ABIDE I are indeed correlated with SRS scores while being uncorrelated with the covariate FD (Supplementary Fig. 5), demonstrating the robustness and generalization ability of CR-mCCAR, capturing features that preserve their relation to the reference and absence of covariate effects even in out of sample data.”

Probably not a substantial claim: may not pass S4 (based).

CR-mCCAR can be extended to accommodate several target references with multiple covariates.

“The advantages of the proposed method include: 1) effectively removing covariates in supervised fusion while avoiding information loss; 2) robustness to noise; 3) generalizability to independent datasets; 4) improved performance in downstream tasks such as classification; and 5) CR-mCCAR can be extended to accommodate several target references with multiple covariates.”

May combine two claims.

CR-mCCAR preserves information for each modality as much as possible while improving downstream task performance such as classification.

“Comparing to traditional fusion models, this method can preserve the information for each modality as much as possible, while improving performance in downstream tasks, such as classification.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The proposed idea can be directly generalized to similar fusion methods such as parallel ICA, SCCA, and mCCA.

“The idea proposed in this study can directly be generalized to similar fusion methods (parallel ICA, SCCA, mCCA).”

May combine two claims.

Co-varying multimodal brain patterns in ADHD and ASD that correlate with symptom scores but not with covariates were identified and replicated in an independent cohort.

“We identified co-varying multimodal brain patterns in ADHD and ASD that are correlated with symptom scores, but not correlated with the covariates, and showed the identified patterns can be replicated in an independent cohort.”

May combine two claims.

CR-mCCAR effectively removes covariates in supervised fusion while avoiding information loss.

“The advantages of the proposed method include: 1) effectively removing covariates in supervised fusion while avoiding information loss; 2) robustness to noise; 3) generalizability to independent datasets; 4) improved performance in downstream tasks such as classification; and 5) CR-mCCAR can be extended to accommodate several target references with multiple covariates.”

May combine two claims.

Covariate regression before fusion causes substantial loss of reference information, while CR-mCCAR removes covariate effects and retains reference information in supervised fusion.

“This demonstrates that covariate regression before fusion leads to substantial reference information loss. However, the proposed method CR-mCCAR is able to remove the covariate effects while retaining information about the reference in supervised fusion, which demonstrates its key advantages.”

Classified as probably not a substantial claim. May combine two claims.

CR-mCCAR can be extended into several target references with multiple covariates within one framework, promising wide utility for multimodal biomarker detection across brain disorders.

“Furthermore, CR-mCCAR can be extended into several target references with multiple covariates within one framework, promising wide utility for multimodal biomarker detection across brain disorders.”