Time-sequential prediction of postoperative complications after gastric cancer surgery using machine learning: a multicenter cohort study.

Background: Although many studies have developed logistic regression models for predicting complications using preoperative and intraoperative data, none have applied comprehensive perioperative information with machine learning (ML) to enable time-sequential predictions.

Methods: This study included patients undergoing gastric cancer surgery between 2013 and 2019 at two hospitals. Comprehensive perioperative data were collected. Four ML models were developed: the postoperative day (POD) 1 and POD 3 models predicted complications occurring from POD 2 and POD 4, while the 24-h and 8-h models predicted complications within the 24 and 8 h, respectively, after collection of the most recent biochemical data and vital signs. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) with repeated validation for generalizability.

Results: Among 4139 patients, 782 (18.9%) experienced complications (Clavien-Dindo grade ≥ II). The 8-h model achieved the highest AUC (0.737) for overall complications. The POD 3 model outperformed the POD 1 model, with AUCs exceeding 0.8 for pancreatic fistula (0.869) and intra-abdominal abscess (0.821). The 8-h and the 24-h model both achieved AUCs above 0.8 for specific infectious complications. The 8-h model demonstrated the following AUCs: 0.889 for pancreatic fistula, 0.842 for intra-abdominal abscess, 0.826 for pneumonia, and 0.824 for anastomotic leakage, surpassing all POD-based models. In each 8-h model, C-reactive protein, pulse rate, and intraoperative blood loss consistently emerged as significant variables.

Conclusion: Hour-based ML models incorporating comprehensive perioperative data predict post-gastric cancer surgery complications with high accuracy and time-sequential capability, potentially aiding clinical decision-making and improving outcomes.