Interpretable video-derived markers of surgical experience during robotic suturing: a proof-of-concept study in an ex vivo porcine model

BACKGROUND: We present LUMICS (Multi-Level Understanding of Minimally Invasive Colon Suturing) and examine whether routine surgical video yields interpretable markers of general surgical experience in a standardized ex vivo porcine model. METHODS: This observational proof-of-concept study analysed 15 endoscopic recordings from 14 operators. Predictions were converted into action distribution, step duration, and adapted log dimensionless jerk (ALDLJ). Experience classification used 14 procedures from 13 surgeons in four Random Forest configurations with leave-one-surgeon-out evaluation. Because the workflow extractor was trained on procedures represented in the classifier cohort, AUCs are internal estimates and may be optimistic. RESULTS: Experienced surgeons had higher mean ALDLJ (Cliff's delta 0.67, 95% CI 0.14-1.00) and AUC 0.83 (0.55-1.00). Adding ALDLJ to actions increased grouped AUC from 0.52 to 0.62, reproduced by nested cross-validation (0.48 to 0.64) and procedure-level analysis (0.49 to 0.67). Actions plus duration yielded AUC 0.64 and all features 0.54; duration and ALDLJ were strongly inversely associated (Spearman rho = -0.98). CONCLUSIONS: Routine endoscopic video yielded clinically interpretable, sensor-free markers of efficiency and motion smoothness. ALDLJ showed a consistent experience-related signal beyond action distribution, while its overlap with duration suggests convergent information. These findings support ALDLJ's further development as a formative feedback marker pending external validation.