Stacking Machine Learning for Predicting Postoperative Abnormal Coagulation After Surgery

Background Postoperative abnormal coagulation (PAC) is a frequent and clinically consequential complication after hepatectomy. Although patients may appear stable on postoperative day 1 (POD1), PAC emerges during POD2–POD5. We try to develop and explain a model that predicts PAC using perioperative data up to POD1. Methods We performed a single-centre retrospective study of adults undergoing elective hep- atic surgery (2018–2019). Preoperative, intraoperative, and POD1 clinical/laboratory variables were extracted. Eight base models were trained; their predicted probabilities were concatenated into a feature matrix and fed to a neural stacking model. Discrimination (AUC, AP), classification (F1, precision, recall, accuracy), and probabilistic performance (calibration curves, Brier score) were eval- uated on held-out data. Model interpretability was assessed with SHAP, and all models were retrained using the 11 variables selected by SHAP. Results Among 2,722 patients, 557 (20.5%) developed PAC on POD2–POD5. The best stacking model achieved AUC 0.912, F1 0.724, precision 0.769, recall 0.690, and accuracy 0.893; AP reached 0.836. Calibration improved overall (best Brier score 0.0803) compared with single models. SHAP analysis consistently highlighted 11 key variables. Retraining models with only these 11 variables preserved comparable performance, indicating parsimony with clinical interpretability. Conclusions Using perioperative data, stacking machine learning provides accurate, calibrated, and explainable prediction of PAC, enabling early identification of high-risk patients before clinical deterio- ration. The reduced-feature model facilitates practical integration into clinical workflows and supports targeted postoperative monitoring.