Unsupervised machine learning reveals prognostic value of dynamic carbohydrate antigen 125 trajectory in gastric cancer.

BACKGROUND: Carbohydrate Antigen 125 (CA125), a tumor-associated glycoprotein, is an established prognostic marker in gastric cancer. However, the trajectory of CA125 levels incorporating dynamic changes over time have not been investigated. METHODS: CA125 levels were collected during the first year of treatment to form individual trajectories and grouped using the unsupervised K-means algorithm. Cox proportional hazards model assessed the association between trajectory groups and overall survival (primary outcome). RESULTS: A total of 1,015 patients were included, covering stages I-IV. 7,347 CA125 tests were conducted, including 1,638 preoperatively and 5,709 postoperatively. The median CA125 trajectory exhibited a "rise-decline" pattern, with preoperative levels of 11.0 U/ml, followed by a temporary peak of 22.98 U/ml in the first quarter post-surgery, and subsequently declines to 12.21 U/ml in the second quarter and 10.90 U/ml at six months post-surgery. The unsupervised K-means machine learning identified 4 latent patterns of trajectories ranging at different levels. Trajectories in higher-level groups were significantly associated with worse survival outcomes: Low (n = 309, reference), Medium (n = 390, HR = 1.75, p = 0.006), High (n = 238, HR = 4.66, p < 0.001), and Ultra High (n = 78, HR = 17.9, p < 0.001). Subgroup analysis revealed that across all stages, the survival risks in the High and Ultra High trajectory groups were consistently higher compared to the Low group. Multivariable Cox proportional hazards model confirmed trajectory as an independent prognostic factor. In comparative Cox analysis, models with trajectory (c-index 0.74–0.81) showed superior efficacy than models with static values (c-index 0.59–0.66). CONCLUSION: We identified the distinct latent patterns of CA125 trajectory dynamics during the initial years of disease course. These patterns exhibited robust independent prognostic value.