A landmark-based dynamic prediction model for anastomotic leakage after rectal cancer surgery: integrating perioperative and postoperative trajectories

Abstract

Background Anastomotic leakage (AL) remains a severe complication after radical surgery for rectal cancer. Current prediction models rely mainly on static variables and may not capture evolving perioperative physiological stress. This study developed and validated a twostage landmark-based dynamic prediction model that integrates baseline clinical features with postoperative inflammatory trajectories for individualized AL risk stratification. Methods We retrospectively analysed 815 patients who underwent radical rectal cancer resection. A consensus feature set was derived by majority voting across four feature-selection pipelines: LASSO, RF-MDA, SVM-RFE, and stepwise AIC. Model 1 used preoperative and intraoperative variables to estimate AL risk immediately after surgery. Model 2 sequentially incorporated early postoperative dynamic indicators. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration curves, integrated discrimination improvement (IDI), net reclassification improvement (NRI), and decision curve analysis (DCA). Results AL occurred in 82 of 815 patients (10.1%). The multi-algorithm consensus strategy identified 16 core baseline variables. Model 1 showed strong discrimination, with an AUC of 0.825 (95% CI, 0.774–0.875). After incorporating early postoperative dynamic parameters, including longitudinal neutrophil-to-lymphocyte ratio trajectory clusters and acute hypoalbuminemia, Model 2 improved the AUC to 0.889 (95% CI, 0.848–0.930; DeLong test, P < 0.001). Model 2 also improved risk reclassification, with an IDI of 0.116 (P < 0.001) and a net NRI of 0.129 (P = 0.006). The Brier score decreased to 0.057, and DCA showed consistently greater clinical net benefit across a broad range of threshold probabilities. Conclusions This two-stage landmark-based dynamic prediction framework improves early warning for AL by combining baseline risk factors with postoperative inflammatory dynamics. The model may support more precise perioperative risk stratification and earlier individualized intervention after radical rectal cancer surgery.