Basic Science and Pathogenesis.

Area of Science:

• Biomedical data science
• Machine learning for healthcare
• Neuroscience

Background:

• Alzheimer's disease (AD) diagnosis increasingly uses multimodal data (neuroimaging, biomarkers).
• Demographic biases in data can lead to disparities in diagnostic accuracy across groups.
• Traditional methods like CCA can amplify these inequities.

Purpose of the Study:

• To develop a novel framework, Fair Representation CCA (FR-CCA), for equitable Alzheimer's disease biomarker analysis.
• To harmonize fairness and diagnostic utility in multimodal data integration for AD.
• To address demographic biases in learned representations for improved clinical decision-making.

Main Methods:

• Proposed FR-CCA, a framework maximizing cross-modal correlations while ensuring statistical independence from sensitive attributes (age, sex).
• Learned low-dimensional representations from multimodal data.
• Validated FR-CCA on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, comparing against conventional CCA and fair variants.

Main Results:

• FR-CCA demonstrated superior fairness-performance trade-offs, improving fairness by 26-105% on ADNI data with minimal correlation loss (<2%).
• Significantly reduced demographic parity (62-76%) and equality of opportunity gaps (81-88%).
• Reduced sex-related bias in MRI and PET modalities by 84-90% while maintaining competitive classification performance.

Conclusions:

• FR-CCA offers a clinically actionable solution for equitable Alzheimer's disease diagnosis.
• Balances fairness and accuracy, ensuring robust generalization across diverse populations.
• Provides a scalable framework for other high-stakes medical applications requiring equitable AI.