Abstract / Summary
Automated surgical phase recognition may convert routinely recorded robotic surgery videos into objective workflow metrics, but its clinical interpretability in complex robotic pediatric and hepatobiliary procedures remains unclear. In this retrospective proof-of-concept study, we developed and internally evaluated a video-based phase-recognition model for robotic choledochal cyst surgery using nine predefined operative phases. The model was trained on 20 robotic cases and evaluated on an independent internal dataset of 10 cases. AI-derived phase durations were applied to a separate clinical interpretation cohort of 20 consecutive single-surgeon robotic cases, including 12 pediatric and 8 adult patients. Phase-specific workflow differences, case-order trends, contributors to operative-time variability, correlations with estimated blood loss, and exploratory early-phase prediction were analyzed. After temporal smoothing, the model achieved an overall accuracy of 80.1% (95% CI 77.2–82.9%). Adult cases had significantly longer total operative times than pediatric cases (median 516.2 vs. 372.8 min; p = 0.009), with longer durations in preparation, transverse colon takedown, post-anastomosis-to-completion, and Phase 7 (Others). The main phases associated with operative-time variability differed by age group: post-anastomosis-to-completion and lower bile-duct dissection in pediatric cases, and proximal bile-duct dissection in adults. Estimated blood loss correlated positively with AI-derived durations of lower bile-duct dissection (rho = 0.684, p = 0.001) and proximal bile-duct dissection (rho = 0.475, p = 0.034). No clear association was observed with postoperative bile leakage. AI-based phase recognition provided clinically interpretable workflow metrics in robotic choledochal cyst excision, identifying phase-specific workflow variation not captured by total operative time. These preliminary findings support further validation of phase-based robotic workflow analysis.