Background
The benefit of postoperative adjuvant chemotherapy (ACT) after first-line neoadjuvant chemotherapy (NACT) in gastric cancer (GC) is inconsistent in real-world practice. A practical risk tool is needed to identify patients most likely to benefit and to avoid overtreatment.

Methods
Patients from a four-center retrospective cohort (training n=725; validation-1 n=149; validation-2 n=276) were stratified using the GAM-derived risk score with a pre-specified threshold: 3-year predicted disease progression probability of 30% (high-risk vs low-risk). DFS separation between strata was assessed by Kaplan–Meier analysis in all cohorts. To evaluate ACT benefit while mitigating selection bias, inverse probability of treatment weighting (IPTW) was applied to balance baseline characteristics between ACT and non-ACT groups. Within each risk stratum, DFS was compared pre- and post-weighting. Clinical benefit was further quantified by 3-year ΔRMST (restricted mean survival time gain) and ΔRisk (absolute recurrence-risk reduction).

Results
Risk stratification consistently separated DFS, with significantly worse DFS in high-risk patients across training and both validation cohorts. After IPTW, covariates achieved good balance and overlap between ACT and non-ACT groups. A significant interaction between risk strata and ACT was observed: in the high-risk group, ACT was associated with significantly superior DFS across all cohorts; in the low-risk group, DFS did not differ significantly between ACT and non-ACT. High-risk absolute benefit was substantial: training cohort ΔRMST ~5 months and ΔRisk −14%; validation-1 ΔRMST 7 months and ΔRisk −20%; validation-2 ΔRMST 6 months and ΔRisk −20%. In low-risk patients, 95% CIs for both ΔRMST and ΔRisk crossed the null across cohorts.

Conclusions
An ML-guided risk score can identify GC patients who meaningfully benefit from ACT after first-line NACT, while supporting omission of chemotherapy in low-risk patients.