Representation Is All We Need: Performance and Fairness of Google X-ray Foundation Model Representations - A Preliminary Study

AI has shown remarkable potential in healthcare, but faces accessibility challenges due to high computational and expertise demands, especially in medical image analysis. Vector embeddings (Emb) offer a solution by converting large medical image datasets into compact representations via foundation models in zero-shot inference, reducing GPU and storage needs. We evaluate AI models trained on Emb versus medical images for chest X-ray diagnosis and findings show that Emb-based models maintain classification performance while improving fairness across demographics(e.g., race, sex, and age), making AI more accessible to low-resource communities.