Tensor dictionary-based heterogeneous transfer learning to study emotion-related gender differences in brain.

Area of Science:

• Neuroimaging analysis
• Machine learning
• Cognitive neuroscience

Background:

• Collecting auxiliary labeled neuroimaging data across domains is challenging, leading to insufficient sample sizes.
• Neuroimaging data, such as brain activation and functional connectivity, are high-order heterogeneous data.
• Existing methods struggle with distinct feature spaces in multi-domain neuroimaging data.

Purpose of the Study:

• To address insufficient sample sizes in neuroimaging using heterogeneous transfer learning.
• To leverage low-dimensional brain activation data to improve high-dimensional functional connectivity data classification.
• To identify emotion-related brain gender differences during adolescence using fMRI data.

Main Methods:

• Proposed a tensor dictionary-based heterogeneous transfer learning framework.
• Combined supervised tensor dictionary learning for feature encoding and heterogeneous transfer learning for knowledge sharing.
• Employed feature transformation based on mathematical relationships between data domains.

Main Results:

• The proposed framework significantly enhanced classification performance by incorporating prior knowledge.
• Experiments on simulated, fMRI, and EEG data demonstrated the method's generalizability.
• Identified temporal variability in brain activity as a key factor in gender differences in emotion regulation.

Conclusions:

• Heterogeneous knowledge sharing effectively captures multifaceted brain characteristics and improves model generalization.
• The framework reduces training costs and enhances classification performance in neuroimaging.
• Understanding gender-specific neural mechanisms of emotional cognition can inform gender-tailored treatments for neurological diseases.