Answers are grounded in this paper and cite their sources.
Aim: Whereas computer vision has advanced the capabilities of Fundamentals of Laparoscopic Surgery (FLS) trainers, real-time coaching using machine learning (ML) remains limited. This retrospective proof-of-concept study evaluated whether an ML model could provide automated FLS pass/fail assessment comparable to faculty, using pre-recorded videos. Methods: Residents performed video-recorded FLS peg transfers scored pass/fail. A training set of 60 videos (51 standard cases, 9 edge-case videos) was used to develop a multi-stage pipeline combining a “you only look once” (YOLOv8n) object detector, and a deterministic rule engine encoding FLS pass/fail criteria. The model was evaluated on 24 resident-disjoint, held-out videos. A faculty majority-vote reference standard was set by five evaluators per video. Model predictions were compared against this reference standard. Results: The ML model classified outcomes accurately in 92% of cases [95%CI: 74%-98%; F1 score = 0.93 (pass positive); Cohen’s κ = 0.83]. Total analysis time was 6 min vs. approximately 56 min of video playback. The model concurrently identified errors in 70% (31/44) of error events, with 97% (30/31) concordance in error classification. Among the 13 remaining events, the model preceded the faculty-annotated timestamp in 8/13 (62%). Real-time instructional prompts were implemented in a heads-up display but not evaluated here. Conclusion: The system provided accurate FLS peg-transfer assessment comparable to faculty while reducing evaluation time. As a retrospective proof-of-concept, it supports automated assessment and, pending prospective validation, could deliver scalable real-time feedback in skills training. External validation in larger, multi-institutional cohorts should precede adoption.
No takes yet. Share an insight, caveat, or question.
Kewalramani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: