Perspective on Bias in Biomedical AI: Preventing Downstream Healthcare Disparities
A new perspective paper highlights significant biases in biomedical AI, stemming from data collection and research prioritization rather than just clinical implementation. Analysis of omics publications and large datasets shows limited reporting of demographic information, with European-ancestry data dominating. The paper warns that as biomedical foundation models become more prevalent, they risk perpetuating these early-stage biases, leading to cascading healthcare inequities. The authors propose a focus on provenance, openness, and reliability through evaluation transparency to mitigate these issues. AI
IMPACT Biomedical AI models risk perpetuating healthcare disparities due to biased data, necessitating a focus on data provenance and transparency.