What happened
A new way to pick which medical data points a human should label — using interactive 2D charts instead of random selection — produced better results for training AI models across multiple medical tasks. When doctors can visually explore patterns in high-dimensional data before labeling, they catch rare cases more effectively and the overall model performs better, though with some trade-offs in consistency when annotators work in isolation.
Why it matters
Medical AI models are only as good as the labeled data they train on, and labeling medical time-series data (like infant movement patterns or voice recordings) is expensive and tedious work. This research shows that letting annotators use visualization to strategically pick which data to label — rather than picking randomly or by algorithm — improves both what gets labeled and how well the resulting AI performs, which could reduce the human annotation burden and cost in building biomedical AI systems.