Dynamic Machine Learning Prediction of Colorectal Anastomotic Leak Using Preoperative Factors and Serial Postoperative Biomarker Trajectories

This is the authors' abstract. We don't add key points for this paper.

Aim Anastomotic leak remains a life-threatening complication after colorectal surgery. Static preoperative models show only modest discrimination, whereas clinical signs often appear late. We aimed to develop and internally validate dynamic machine-learning models integrating preoperative factors with serial postoperative biomarker and physiological trajectories. Method In this retrospective single-centre cohort, XGBoost models were developed using information available preoperatively and through POD1, POD3, or POD5. For the POD3 and POD5 models, patients with AL diagnosed by the respective time point were excluded so that each model estimated the risk of a subsequent leak. Internal validation used repeated stratified 5-fold cross-validation with 5 repeats, with preprocessing and threshold selection performed within the training data. Results Among 409 patients, 31 (7.6%) developed AL. AUROC was 0.696 ± 0.085 preoperatively and 0.714 ± 0.083 at POD1. At POD3, 405 patients remained at risk and 27 subsequently developed AL; AUROC was 0.781 ± 0.086. At POD5, 398 patients remained at risk and 20 subsequently developed AL; AUROC was 0.908 ± 0.047 and AUPRC 0.460 ± 0.152, with sensitivity 0.850 ± 0.177, specificity 0.792 ± 0.074, PPV 0.194 ± 0.081, and NPV 0.990 ± 0.011. Conclusion Serial postoperative trajectories improved internal risk stratification for AL, particularly by POD5. The high POD5 NPV should be interpreted in the context of low event prevalence and internal validation; external validation is required before the model is used to support clinical decisions.