Artificial Intelligence-Assisted Prediction of Immunotherapy Benefit and Recurrence Risk Stratification in Gastric Cancer: A Single-Center Real-World Study

Immune checkpoint inhibitors targeting the PD-1 pathway have improved outcomes in advanced gastric cancer; however, substantial heterogeneity in treatment response remains. Current prediction strategies mainly rely on pretreatment biomarkers, which provide static assessments and may not capture the evolving interactions between tumor progression and host immunity. This study aimed to develop a dynamic prediction framework integrating longitudinal clinical trajectories and immune remodeling for continuous risk assessment during PD-1 inhibitor therapy. Methods: This retrospective study included 171 patients with gastric cancer receiving PD-1 inhibitor-based treatment. A landmark analysis framework was established using sequential 28-day follow-up windows to predict subsequent progressive disease (PD) or recurrence based on available clinical and immune information. Three nested models were developed: a baseline model (M0), a dynamic clinical model (M1), and an immune-integrated model (M2) incorporating longitudinal lymphocyte subset features. Model performance was evaluated using discrimination, calibration, and internal validation with patient-level resampling strategies. Results: Predictive performance improved progressively with incorporation of longitudinal information. The ROC AUC increased from 0.559 (95% CI, 0.429–0.690) in M0 to 0.737 (95% CI, 0.613–0.852) in M1 and 0.786 (95% CI, 0.681–0.883) in M2. Dynamic clinical information significantly improved discrimination compared with baseline prediction (ΔAUC = 0.178, p = 0.025), while additional immune features further enhanced risk stratification (ΔAUC = 0.049, p = 0.040). Interpretation analysis identified tumor burden evolution, neutrophil-to-lymphocyte ratio trajectories, and immune functional axes involving T-cell, innate cytotoxicity, and humoral immunity as major contributors to risk estimation. Landmark analyses demonstrated the feasibility of continuously updating PD/recurrence risk during treatment. Conclusions: This study establishes a dynamic landmark prediction framework integrating longitudinal clinical trajectories and immune remodeling for continuous risk assessment during PD-1 inhibitor therapy. Dynamic clinical and immune features provided complementary predictive information beyond baseline characteristics, supporting an adaptive approach for precision immunotherapy monitoring in gastric cancer.