A Manifold Regularized Multi-Task Learning Model for IQ Prediction From Two fMRI Paradigms.

Area of Science:

• Neuroscience
• Machine Learning
• Cognitive Science

Background:

• Multi-modal brain functional connectivity (FC) data offer insights into individual differences in behavior and cognition.
• Joint learning of multi-modal data can enhance learning performance by leveraging intrinsic associations.
• Existing multi-task learning models often overlook structural information across modalities, limiting discriminative feature extraction.

Purpose of the Study:

• To propose a novel manifold regularized multi-task learning model for integrating multi-modal brain FC data.
• To simultaneously consider between-subject and between-modality relationships for improved feature learning.
• To identify significant biomarkers associated with intelligence quotient (IQ) variations.

Main Methods:

• Developed a manifold regularized multi-task learning model incorporating l2,1-norm for group-sparsity and a novel manifold regularizer.
• Enforced joint feature selection across modalities to preserve structural information within and between them.
• Validated the model on the Philadelphia Neurodevelopmental Cohort dataset using functional MRI (fMRI) data for IQ prediction.

Main Results:

• The proposed model achieved improved prediction performance for intelligence quotient (IQ).
• Identified a set of IQ-relevant biomarkers from functional connectivity network data.
• Demonstrated the model's adaptability to high-dimensional neuroimaging data with small sample sizes.

Conclusions:

• The developed multi-task learning model effectively integrates multi-modal brain data.
• The model facilitates the discovery of significant biomarkers related to human intelligence.
• This approach enhances the understanding of the neurobiological basis of cognitive traits.