Prospective multicenter validation of a machine learning model for predicting anastomotic leakage in patients with gastric adenocarcinoma undergoing total or proximal gastrectomy.

Objective: Predicting esophago-gastric and esophagojejunal anastomotic leakage (AL) is inherently challenging. The aim of the present study was to investigate the clinical utility of a real-time machine learning model for predicting AL.

Background: AL is one of the most serious postoperative complications following esophagogastric and esophagojejunal anastomoses. Traditional risk stratification methods have often struggled to accurately predict which patients are most at risk, owing to the multifactorial nature of AL and the variability in patient and operative factors.

Methods: In this prospective study, gastric adenocarcinoma patients who were scheduled for total or proximal gastrectomy from four medical centers were enrolled between January 2022 and January 2024. During operations, a developed machine learning model was used to assess the risk of AL. The primary outcome is the occurrence of AL.

Results: A total of 512 patients were included. AL was observed in 13 patients (2.54%). The model yielded an area under the operating characteristic curve of 0.780, a sensitivity of 0.769, a specificity of 0.577 and a negative predictive value of 0.990. Of the 512 patients, 221 were identified as high-risk and 291 as low-risk. Compared with the low-risk group, the AL rate was significantly higher in the high-risk group (10/221 vs. 3/291; P = 0.027). Post hoc analysis revealed ~ 35% (risk score<0.45)patients can safely avoid intensive monitoring.

Conclusions: By achieving high sensitivity while excluding nearly half of the non-AL subgroups, the model (https://gasal.21cloudbox.com/) provides effective risk stratification of AL in patients with gastric adenocarcinoma undergoing esophagogastrostomy or esophagojejunostomy.