Background
Recurrence risk after first-line neoadjuvant chemotherapy (NACT) and curative gastrectomy remains highly heterogeneous in gastric cancer. However, postoperative risk assessment still relies mainly on ypT/ypN staging, which may not adequately capture residual disease burden or support individualized postoperative management.

Methods
In this four-center study, we developed and externally validated an interpretable machine-learning model to predict disease-free survival (DFS) after first-line NACT for gastric cancer using routinely available clinicopathologic variables. Model performance was assessed by discrimination, calibration, and decision-curve analysis, and compared with a conventional ypT/ypN-based Cox model. SHAP was applied to improve interpretability and to quantify the contribution of individual predictors to patient-level risk estimates.

Results
The model demonstrated robust performance across multicenter cohorts and maintained predictive ability in independent external validation sets. It effectively stratified patients according to postoperative recurrence risk, yielded greater clinical net benefit than treat-all/treat-none strategies, and outperformed a conventional ypT/ypN-based Cox model. Its interpretable framework also enabled transparent identification of key variables driving individual predictions, thereby improving clinical plausibility and facilitating bedside interpretation. These findings support its transportability across heterogeneous real-world multicenter settings.

Conclusions
An interpretable ML model provided robust, externally validated prediction of DFS after first-line NACT for gastric cancer across heterogeneous multicenter cohorts. Beyond conventional postoperative staging, it may support risk-adapted surveillance, MDT discussion, and identification of patients who may benefit from closer follow-up, refined postoperative counseling, or further management strategies.