Development and Validation of a Machine Learning Model to Predict Postoperative Complications Following Radical Gastrectomy for Gastric Cancer

Zhenmeng Lin¹Yan Mingfang²Shenghong Wei²Yangming Li¹Jinliang Jian^1*Haiyan Peng^3*

• ¹Department of Gastrointestinal Surgery, Clinical Oncology School of Fujian Medical University & Fujian Cancer Hospital, Fuzhou, China
• ²Department of Anesthesiology, Clinical Oncology School of Fujian Medical University & Fujian Cancer Hospital, Fuzhou, China
• ³College of Animal Science, Fujian Agriculture and Forestry University, Putian, China

Postoperative complications significantly adversely affect recovery and prognosis following radical gastrectomy for gastric cancer. We developed and validated machine learning (ML) models to predict these complications and constructed a clinically applicable dynamic nomogram. Methods: Using a prospectively maintained database, we conducted a retrospective analysis of 1,486 patients from Fujian Cancer Hospital (training cohort) and 498 from the First Hospital of Putian City (validation cohort). Feature selection integrated Lasso regression, the Boruta algorithm, and Recursive Feature Elimination (RFE). Six ML models were developed and evaluated: TreeBagger (TB), Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Gaussian Naïve Bayes (GNB), and Artificial Neural Network (ANN). The significant predictors identified were incorporated into a logistic regression model to determine independent risk factors, which then formed the basis of a dynamic nomogram deployed as an interactive web application for clinical use. Results: RF demonstrated numerically superior performance among the evaluated models in both cohorts. Independent risk factors included age, BMI, diabetes mellitus, ASA grade, operative time, and surgical approach. The dynamic nomogram achieved AUCs of 0.805 (training) and 0.856 (validation), with calibration curves and decision curve analysis confirming its reliability. DeLong's test revealed no significant difference in AUC between the RF model and nomogram in either cohort (training: Z = -0.385, p = 0.701; validation: Z = -1.756, p = 0.058). Conclusion: While the RF model provided optimal predictive accuracy among ML algorithms, the interpretable nomogram offers comparable discrimination and clinical accessibility. Both tools facilitate the early identification of high-risk patients, enabling personalized interventions to optimize postoperative recovery.