Looking Beyond the Scale: Do Surgical Skill Models Learn Transferable Representations Across Assessment Rubrics?

This study investigates whether vision‑based models for surgical skill assessment learn representations that transfer across different scoring rubrics (GOALS and OSATS) using the LASANA and JIGSAWS datasets. By evaluating end‑to‑end training, Adaptive Sharpness‑Aware Minimization, and self‑supervised/contrastive pretraining, the authors find that models pretrained on JIGSAWS can transfer reasonably well to LASANA, but transfer to JIGSAWS fails, likely due to annotation inconsistencies. Control experiments with a Kinetics‑pretrained backbone show that task‑specific heads carry most of the skill prediction load, while the backbone provides general spatiotemporal features.