MAGE (Multimodal AI-Enhanced Gastrectomy Evaluation): Comparative Analysis of Machine Learning Models for Postoperative Complications in Central European Gastric Cancer Population

TL;DR

This study uses machine learning to predict postoperative complications in gastric cancer patients, aiming to improve surgical risk assessment and treatment planning.

Contribution

The study introduces a free online risk calculator and demonstrates superior performance of non-linear ML models in predicting complications.

Findings

XGBoost and Random Forest models achieved the highest accuracy in predicting postoperative complications.

Non-linear machine learning models outperformed traditional linear approaches in complication prediction.

A publicly accessible online risk calculator was developed for clinical use.

Abstract

Patients undergoing surgery for gastric cancer face a meaningful risk of postoperative complications, yet reliable tools to predict who is most at risk are still limited. In this study, we aimed to support preoperative decision-making by developing machine-learning models trained on data from gastric cancer patients receiving multimodal therapy at the Department of Surgical Oncology, Medical University of Lublin. If validated in future studies, this approach could help clinicians estimate surgical risk more accurately, individualize treatment planning, and improve perioperative safety. We also developed a free, user-friendly online risk calculator to facilitate clinical use.

Introduction: By leveraging dedicated datasets and predictive modeling, machine-learning (ML) algorithms can estimate the probability of both short- and long-term outcomes after surgery. The aim of this study was to evaluate the ability of ML-based models to predict postoperative complications in patients with gastric cancer (GC) undergoing multimodal therapy. In particular, we aimed to develop a free, publicly accessible online calculator based on preoperative variables. Materials and Methods: Patients with histologically confirmed locally advanced (cT2-4N0-3M0) GC who underwent multimodal treatment with curative intent between 2013 and 2023 were included in the study. ML models evaluation pipeline was used with Stratified 5-Fold Cross-Validation. Results: A total of 368 patients were included in the final analytic cohort. Among five algorithm classes under 5-fold cross-validation, Compute Area Under the Receiver Operating Characteristic Curve (ROC AUC) was 0.9719, 0.9652, 0.9796, 0.8339 and 0.7581 for XGBoost, Catboost, Random Forest, SVM and Logistic Regression, respectively. Macro F1 was 0.8714, 0.5094, 0.8820, 0.8714 and 0.4579 for XGBoost, SVM, Random Forest, CatBoost and Logistic Regression, respectively. Overall Accuracy was 0.8897, 0.5980, 0.8885, 0.8750 and 0.5466 for XGBoost, SVM, Random Forest, CatBoost and Logistic Regression models, respectively. Conclusions: In this Central and Eastern European cohort of patients with locally advanced GC, ML models using non-linear decision rules-particularly Random Forest and XGBoost- substantially outperformed conventional linear approaches in predicting the severity of postoperative complications. Prospective external validation is needed to clarify the model’s clinical utility and its potential role in perioperative decision support.

Figures