Evaluating and mitigating bias in AI-based medical text generation.

Area of Science:

• Medical artificial intelligence
• Natural Language Processing
• Algorithmic fairness

Background:

• Deep learning AI achieves expert performance in medicine.
• Concerns exist about AI amplifying human bias, especially in underserved populations.
• Fairness in medical AI is studied in imaging but less in text generation.

Purpose of the Study:

• Investigate fairness issues in medical AI text generation.
• Identify performance discrepancies across demographic groups.
• Develop and evaluate a bias mitigation algorithm.

Main Methods:

• Analyzed performance disparities in AI text generation across races, sexes, and age groups.
• Developed a novel algorithm for selective optimization of underserved groups.
• Evaluated the algorithm on multiple AI backbones, datasets, and modalities.

Main Results:

• Observed significant performance differences in medical AI text generation across demographic groups.
• The proposed algorithm effectively reduced bias in text generation.
• Fairness improvements were achieved without compromising overall AI performance.

Conclusions:

• Medical AI text generation systems exhibit fairness issues.
• The proposed selective optimization algorithm successfully mitigates bias.
• This approach enhances AI fairness in medicine for diverse populations.