Скачать или смотреть When AI is trained mostly on men, it gets worse at diagnosing everyone

A 2025 PLOS One study tested how sex bias in training data affects medical AI. Researchers used 5,000 knee ultrasound images (equal numbers of men/women, with and without joint recess distension) and trained three AI models: one on male images only, one on female images only, and one on a balanced mix.

The male-only model was consistently the worst performer – about 3.5 percentage points less accurate than both the female-only and balanced models, a gap that was statistically significant in 5,000 bootstrap tests. In contrast, the female-only model performed as well as the balanced one.

The same dataset also trained a separate AI to predict patient sex from knee ultrasounds alone — it hit 95% accuracy (AUC 0.967), proving that “neutral” medical images still carry strong sex signals AI can latch onto.

This study shows that sex bias in training data can quietly bend AI decision-making and that diverse, balanced datasets are not optional if we want fair diagnostics.

Source: Mendez et al. (2025). Evaluating the impact of sex bias on AI models in musculoskeletal ultrasound of joint recess distension. PLOS ONE, 20(11): e0332716.