This research introduces Fair-MoE, a fairness-oriented mixture of experts architecture that addresses bias in medical Vision-Language Models.

• Integrates fairness considerations directly into model architecture rather than as post-processing
• Creates specialized expert networks that handle different demographic groups more equitably
• Demonstrates improved performance while maintaining fairness across diverse patient populations
• Validated using the Harvard-FairVLMed dataset

Why it matters: As medical AI adoption grows, ensuring equitable treatment across all demographics becomes critical for ethical deployment. Fair-MoE provides a promising architectural solution to build fairness directly into clinical decision support systems.

Fair-MoE: Fairness-Oriented Mixture of Experts in Vision-Language Models