Machine learning-based risk modeling for safety-focused learning curve assessment in robotic left-sided colorectal cancer surgery.

The transition from laparoscopic to robotic surgery for left-sided colorectal cancer raises safety concerns during the learning curve, particularly when complex cases are preferentially selected for the robotic platform. We evaluated a machine learning-based framework for risk-adjusted safety monitoring of robotic implementation, using outcomes from an established laparoscopic program as the reference. We retrospectively analyzed adult patients who underwent minimally invasive left-sided colorectal resection for malignancy between May 2023 and September 2025. A penalized logistic regression model predicting a composite adverse endpoint (anastomotic leak, reoperation, major complication, unplanned intensive care admission, or mortality) was developed in a laparoscopic training cohort (n = 211) using four preoperative variables (age, body mass index, American Society of Anesthesiologists physical status, and tumor location). Model-derived expected risks were applied to a robotic cohort (n = 93) to construct a risk-adjusted cumulative sum (RA-CUSUM) chart. The robotic cohort included a higher proportion of rectal tumors and more frequent neoadjuvant therapy than the laparoscopic cohort and had longer operative times, whereas the composite adverse event rate was similar (12.9% vs. 13.3%). The RA-CUSUM curve for the robotic series fluctuated around the expected risk baseline derived from the laparoscopic benchmark without a sustained upward drift. These findings suggest that, in this single-center experience, early robotic adoption did not show a clear signal of excess risk-adjusted short-term adverse events despite increased case complexity and demonstrate the feasibility of embedding a laparoscopic-derived risk model into RA-CUSUM analysis as a pragmatic tool for learning curve assessment.