Models May Learn Who Generated the Data: Robotic Surgery Study Shows AUC Drops 0.89 to 0.57

bravo_abad · x · 2026-09-30

Ueki et al. show models can pick up habits of the data generator: analyzing 98 robotic-surgery procedures by 16 surgeons via 25 features from 3D hand trajectories, a random forest hit AUC 0.89 with random splits but only 0.57 when each test surgeon's procedures were excluded from training — a warning that random train-test splits can overstate model usefulness wherever the data source correlates with the label.

Related event: Study Finds Models Can Learn the Habits of Data Generators(2 posts)→

Original post →

More from Embodied

Embodied channel →