This research introduces a modality-agnostic auditing framework that can identify latent biases in medical AI datasets before they impact clinical decisions.

• Provides a systematic approach to generate targeted hypotheses about sources of bias in medical datasets
• Successfully identifies biases across multiple medical modalities including skin lesion images, EHR language, and ICU mortality prediction
• Addresses a critical gap in ensuring medical AI systems perform fairly and safely across diverse patient populations

Why it matters: As AI becomes central to evidence-based medicine, undetected dataset biases can lead to harmful clinical outcomes and exacerbate healthcare disparities. This framework enables proactive bias detection before AI systems are deployed in clinical settings.

Detecting Dataset Bias in Medical AI: A Generalized and Modality-Agnostic Auditing Framework