Machine learning model-guided selective use of temporary diverting ileostomy in rectal cancer surgery: a randomized controlled trial.

Abstract

The appropriate use of temporary diverting ileostomy (TDI) in rectal cancer surgery lacks standardized criteria. A randomized controlled trial was performed to evaluate whether the Risk-Guided Temporary Ileostomy Decision (RTID) system, a machine learning-based anastomotic leakage (AL) prediction model, could improve the suitability of TDI utilization. A total of 872 patients with stage I-III rectal cancer undergoing anterior resection were randomized 1:1 to surgeon discretion versus RTID-guided decisions. The final analysis included 750 patients (control, n = 368; RTID, n = 382). The RTID group showed lower overall TDI rate (18.6% vs. 40.5%; P < 0.001) and unnecessary stoma formation (17.7% vs. 41.3%; P < 0.001). Although a numerical increase in necessary TDI use was observed (55.6% vs. 10.0%), this difference was not significant (P = 0.057). Critically, the incidence of the co-primary safety outcome, AL, was comparable between the RTID and control groups (2.4% vs. 2.7%; P = 0.753), indicating that the pre-specified TDI reduction endpoint was met, while the study was underpowered to formally test non-inferiority for AL. RTID offers an objective tool to support TDI decision-making without a compromise in safety. The present study was registered on ClinicalTrials.gov (no. NCT04999007) on August 8, 2021.