Bridging Prediction and Reality: AI-based External Validation of Predictive Models for Lymph Node Metastasis in Gastric Cancer
Menée à l'aide de modélisations utilisant l'intelligence artificielle et des données portant sur 601 patients atteints d'un adénocarcinome gastrique traité par gastrectomie, cette étude détermine le nombre minimum de ganglions lymphatiques à prélever durant l'opération pour stadifier la maladie et établir un pronostic avec précision
Introduction : The optimal extent of lymphadenectomy in gastric cancer surgery remains a subject of ongoing debate. Our previous modelling work indicated that identifying at least 26 lymph nodes may reveal occult nodal metastases in patients undergoing more limited dissections, thereby reducing the risk of under-staging. However, those analyses were population-based and lacked individual-level validation of clinical outcomes, including disease-free and overall survival. This study sought to externally validate, using artificial intelligence (AI)–driven modelling, the predicted risk of pN upstaging associated with simulated increases in retrieved lymph node counts.
Materials and Methods : The PT cohort comprised 209 patients with gastric adenocarcinoma who underwent curative-intent gastrectomy between 2004 and 2022, all with fewer than 26 examined lymph nodes, pN+ status, and M0 disease. External validation was conducted using the European GASTRODATA (GD) cohort, which included 392 patients from a multinational registry (2019–2022). Patients in both cohorts met identical eligibility criteria. Cohort comparability was assessed using a combination of AI-trained regression models, Random Forest algorithms with cross-validation, and multidimensional projection analysis, ensuring robust evaluation of dataset equivalence. Proportional and exponential simulation models were subsequently applied to the combined dataset (n=601) to explore different assumptions regarding nodal yield scaling.
Results : The two cohorts were statistically comparable across key clinicopathological variables, supporting their suitability for external validation. Simulations predicted increases of 21% (proportional model) and 16% (exponential model) in the number of metastatic lymph nodes, indicating a clinically meaningful risk of under-staging when lymphadenectomy yield falls below recommended thresholds.
Conclusion : These findings externally validate, in a larger multicentre cohort, the previously developed simulation models and reinforce the recommendation that harvesting at least 26 lymph nodes during gastrectomy is essential to optimise staging accuracy and prognostic assessment in gastric cancer, regardless of institutional setting.
European Journal of Surgical Oncology , article en libre accès, 2026