Machine learning-based dynamic CEA trajectory and prognosis in gastric cancer.

Abstract

Background: Static carcinoembryonic antigen (CEA) levels are well‑established prognostic markers in patients with gastric cancer, but the significance of their dynamic trajectories over time has rarely been reported.

Methods: We analysed the perioperative CEA levels (presurgery, early postsurgery, and late postsurgery) of 578 gastric cancer patients who underwent curative resection, with a median follow-up of 29 months. We used the entire cohort for k-means clustering. Survival differences between clusters were assessed using Kaplan-Meier analysis and Cox regression.

Results: Of the 578 patients, 15.57% exhibited elevated CEA levels before surgery (median 2.07 ng/mL), which then decreased to 3.29% (median 1.74 ng/mL) after surgery. However, after six months, a slight rebound was observed (18.51% elevated, median 2.98 ng/mL). K-means clustering identified three CEA trajectories: high, medium, and low (Calinski-Harabasz index: 358). Survival analysis demonstrated that higher CEA trajectories were associated with worse disease-free survival (DFS) and overall survival (OS). With the low cluster as a reference, multivariate Cox regression analysis revealed that a higher CEA trajectory was an independent prognostic factor, with an elevated risk in the high cluster (HR 2.64, 95% CI: 1.37-5.0), indicating that the high cluster had more than twice the mortality risk of the low cluster and that the medium cluster had a moderately increased mortality risk (HR 1.69, 95% CI: 1.0-2.85).

Conclusion: Higher CEA trajectories are associated with a worse prognosis, highlighting the importance of enhanced monitoring for this group of patients.