Joint Dynamic Trajectories of CEA and CA19-9 After Gastric Cancer Surgery: Prognostic Associations and a 12-Month Landmark Analysis

Background: Static single-time-point assessment of carcinoembryonic antigen and carbohydrate antigen 19-9 cannot capture postoperative temporal patterns. We evaluated whether their joint longitudinal trajectories were associated with survival after curative gastrectomy while explicitly addressing the time-dependent bias created when trajectory groups are defined using measurements obtained after surgery. Methods: We analyzed 208 patients with serial carcinoembryonic antigen and carbohydrate antigen 19-9 measurements. Five-time-point trajectories were retained as a descriptive and exploratory analysis. Primary prognostic inference used a 12-month landmark analysis restricted to patients alive and recurrence-free at 12 months with at least three observed paired measurements from the preoperative, 1-month, 6-month, and 12-month time points. Trajectories were re-derived using only pre-landmark information. Overall survival was defined by all-cause mortality. In leakage-controlled fivefold cross-validation, imputation, transformation, cluster selection, trajectory assignment, and model fitting were performed within each training fold. Results: The 12-month landmark cohort included 164 patients and yielded four joint trajectory groups: G1, n = 60; G2, n = 41; G3, n = 32; and G4, n = 31. After adjustment for age, sex, TNM stage, differentiation, tumor site, postoperative chemotherapy, and baseline log-transformed carcinoembryonic antigen and carbohydrate antigen 19-9, G4 was associated with higher point estimates for overall survival and disease-free survival but did not reach statistical significance, with hazard ratios of 1.81 and 1.90, respectively. Adding the landmark trajectory classification increased the apparent C-index from 0.689 to 0.710 for overall survival and from 0.711 to 0.723 for disease-free survival, while likelihood-ratio tests for incremental model fit were not significant. In leakage-controlled cross-validation, the Cox model with CTS4 achieved an out-of-fold C-index of 0.589 and a 24-month post-landmark AUC of 0.635, compared with 0.614 and 0.643 for logistic regression using the same inputs. The original five-time-point surgery-indexed analysis showed stronger associations but was retained only as exploratory because the exposure incorporated future follow-up information. Conclusions: Joint carcinoembryonic antigen and carbohydrate antigen 19-9 trajectories may contain prognostic information, but the association was materially attenuated after correction for exposure timing. These findings should therefore be considered exploratory and hypothesis-generating until prospective external validation is available.