Development of a Machine Learning-Based Predictive Model for Anastomotic Leakage Following Gastric Cancer Surgery

Abstract

Background

The incidence of anastomotic leakage (AL) following radical gastrectomy for gastric cancer ranges from 2.1% to 14.6%, with mortality rates up to 50%. Existing predictive methods lack both timeliness and accuracy. This study aimed to develop an early AL risk prediction model through machine learning-driven integration of multidimensional clinical data.

Methods

We retrospectively enrolled 1588 patients undergoing radical gastrectomy at two tertiary medical centers. AL diagnosis adhered to international consensus criteria. Thirty-six perioperative features were analyzed, including demographics, comorbidities, tumor pathology characteristics, surgical parameters, and dynamic laboratory indicators (assessed at preoperative days 4-6, ≤3 days preoperatively, and ≤3 days postoperatively). LASSO regression identified 11 core predictors; five machine learning models were optimized through 5-fold cross-validation with external validation in an independent cohort.

Results

The LASSO-Logistic model identified key predictors: CRP ≤3 days postoperatively (β = 0.54), age group (β = 0.43), history of abdominal surgery (β = 0.40), and albumin (β = −0.37). The model demonstrated optimal external validation performance (AUC = 0.871; sensitivity = 71.2%; specificity = 87.3%; negative predictive value [NPV] = 96.9%). In sensitivity-optimized mode (threshold = 0.250), NPV increased to 98.9% (sensitivity = 93.2%).

Conclusion

The LASSO-Logistic model, utilizing postoperative CRP (within 3 days) as its core predictor, provides precise early warning for anastomotic leakage risk. External validation demonstrated moderate generalizability with an AUC difference of 0.054 between training and validation sets. Multicenter validation remains essential for clinical translation.