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

Area of Science:

• Neuroimaging
• Biostatistics
• Machine Learning

Background:

• Multimodal fusion leverages shared and complementary information from diverse data sources.
• Supervised fusion is valuable for identifying brain-based patterns linked to clinical measures.
• Handling confounds in brain data analysis is crucial to avoid spurious findings.

Purpose of the Study:

• To introduce CR-mCCAR, a novel method for joint optimization of multimodal fusion and confound removal.
• To capture reliable multimodal brain patterns associated with clinical domains while accounting for covariates.
• To enhance the detection of phenotype-linked multimodal biomarkers for neurological and psychiatric disorders.

Main Methods:

• CR-mCCAR employs a guided fusion model to simultaneously optimize for target components and discount covariate effects.
• Simulations were used to validate the accurate separation of reference and covariate factors.
• Functional and structural neuroimaging data from ADHD and ASD cohorts were analyzed.

Main Results:

• CR-mCCAR accurately separates target and covariate factors in simulations.
• The method identified distinct co-varying patterns in ADHD (striato-thalamo-cortical, salience) and ASD (salience, fronto-temporal) linked to core symptoms, independent of age and motion.
• These findings were replicated in an independent cohort.
• CR-mCCAR significantly improved classification accuracy between ADHD/ASD and controls compared to separate fusion or regression approaches.

Conclusions:

• CR-mCCAR offers a robust framework for jointly optimizing multimodal fusion and confound removal.
• The approach enhances the discovery of reliable, phenotype-linked multimodal biomarkers for brain disorders.
• CR-mCCAR demonstrates superior performance in identifying disease-specific neuroimaging patterns and improving diagnostic classification.