Unveiling the Hidden Biases in Medical AI - Paving the Way for Fairer and More Accurate Imaging Diagnoses

ai-biasmedical-imagingfairnessmachine-learning

Abstraction: 29 identified bias sources across the medical imaging AI development pipeline

Key points:

Multi-institutional MIDRC team identified 29 bias sources across five pipeline stages: data collection, preparation/annotation, model development, evaluation, and deployment
Data collection bias: single-hospital or single-scanner datasets, differential treatment of social groups in research, temporal bias as medical practices evolve
Annotation bias: annotator personal biases and labeling oversights during data preparation
Model development: inherited bias (biased model output used to train next model), historical/societal biases, under-representation of subpopulations
Evaluation bias: biased benchmarking datasets and inappropriate statistical models; deployment bias includes off-label model use and automation over-reliance
Uncorrected biases can worsen existing healthcare access disparities by producing differential benefits across patient groups

Connections: Midrc · AI Bias · Medical AI · Algorithmic Fairness