Dynamic CEA trajectories predict survival for non-metastatic colon cancer: development and validation of a combined prognostic model

Technical Summary: Dynamic CEA Trajectories for Prognosis in Non-Metastatic Colon Cancer

Problem Statement
Accurate postoperative risk stratification is critical for managing stage I–III colon cancer, as 30%–50% of patients experience recurrence after curative resection. While Carcinoembryonic Antigen (CEA) is the standard biomarker for monitoring, conventional clinical practice relies on static thresholds (typically 5 ng/mL). This approach fails to capture the temporal dynamics of tumor behavior. Furthermore, the prognostic utility of postoperative inflammatory markers (e.g., NLR, PLR, SII) remains controversial regarding optimal measurement timing and independent value. There is a need for models that integrate longitudinal biomarker data to improve survival prediction and guide surveillance strategies.

Methodology
This single-center retrospective cohort study analyzed 1,126 patients with stage I–III colon cancer who underwent R0 resection at Zhejiang Cancer Hospital between January 2010 and December 2025.

• CEA Trajectory Classification: The authors developed a four-category grouping method based on temporal slope parameters rather than latent class modeling (which showed instability). The groups were defined as:
  1. Low-stable: Consistently stable CEA levels (reference group).
  2. Early-rising: Significant increase in CEA within 12 months post-surgery (slope difference >75th percentile).
  3. Late-rising: Significant increase in CEA after 12 months (slope difference <25th percentile).
  4. Sustained-high: Mean CEA level ≥5 ng/mL.
• Inflammatory Markers: Inflammatory indices (NLR, PLR, LMR, SII) were calculated using blood counts from the first postoperative follow-up (days 21–56), a window chosen to avoid acute surgical trauma effects.
• Model Construction: Three Cox proportional hazards models were constructed for Overall Survival (OS) and Disease-Free Survival (DFS):
  1. Clinicopathological Model: Included age, sex, T/N stage, differentiation, lymphovascular/perineural invasion, and adjuvant chemotherapy.
  2. CEA Trajectory Model: Included only the four trajectory groups.
  3. Combined Model: Integrated both clinicopathological variables and CEA trajectories.
• Evaluation Metrics: Performance was assessed using Harrell's C-index, time-dependent AUC (at 12–60 months), calibration curves (Integrated Absolute Error), and Decision Curve Analysis (DCA) with 500 bootstrap resamples.

Key Results

• Survival Stratification: The four CEA trajectory groups exhibited significantly different survival outcomes (Log-rank P < 0.001 for both OS and DFS). The Sustained-high group had the worst prognosis (5-year OS: 34.9%; 5-year DFS: 36.6%). The Late-rising group also showed significantly worse DFS compared to the Low-stable group (HR = 2.03). Surprisingly, the Early-rising group showed survival outcomes comparable to the Low-stable group (5-year OS: 98.0%), with no statistically significant difference in DFS risk (HR = 1.28, P = 0.368).
• Model Performance:
  • The CEA Trajectory Model alone demonstrated non-inferior performance to the Clinicopathological Model for both OS (C-index: 0.773 vs. 0.736, P = 0.314) and DFS (C-index: 0.738 vs. 0.735, P = 0.802).
  • The Combined Model significantly outperformed both individual models, achieving a C-index of 0.841 for OS and 0.819 for DFS (both P < 0.001). Time-dependent AUCs for the combined model ranged from 0.811 to 0.865.
  • All models showed good calibration (Integrated Absolute Error < 0.03).
• Inflammatory Markers: In multivariable analyses, inflammatory markers (NLR, PLR, LMR, SII) showed no significant independent prognostic value for OS or DFS.
• Decision Curve Analysis (DCA):
  • For OS prediction, the CEA trajectory model provided the highest net benefit across most clinically relevant threshold probabilities (>0.1).
  • For DFS prediction, the optimal model was threshold-dependent: the clinicopathological model was superior at low thresholds (<0.2), while the CEA trajectory model was superior at high thresholds (>0.2). The combined model showed intermediate net benefit across most ranges.

Key Contributions

1. Novel Classification: The study proposes a robust, clinically interpretable four-category CEA trajectory classification based on slope parameters, avoiding the convergence issues of latent class growth mixture models.
2. Prognostic Equivalence: It demonstrates that dynamic CEA trajectory data alone can predict survival outcomes with accuracy comparable to a complex model containing eight traditional clinicopathological variables.
3. Synergistic Modeling: It validates that combining longitudinal CEA data with clinicopathological factors yields superior predictive accuracy (C-index > 0.84) compared to either approach alone.
4. Negative Finding on Inflammation: The study reports that postoperative inflammatory markers measured at 21–56 days did not provide independent prognostic value in this cohort, suggesting potential informational redundancy with CEA trajectories or suboptimal timing of measurement.

Significance and Claims
The authors conclude that dynamic CEA trajectory monitoring provides complementary prognostic information to traditional clinicopathological factors. The study advocates for integrating longitudinal CEA monitoring into routine postoperative risk assessment.

• Clinical Utility: The choice of the optimal model depends on the clinical decision threshold. The CEA trajectory model is particularly valuable for predicting Overall Survival and high-risk Disease-Free Survival scenarios, potentially simplifying follow-up protocols by reducing reliance on extensive blood panels if the trajectory is clear.
• Risk Stratification: Patients in the "Sustained-high" group require intensified surveillance and potentially stronger adjuvant therapies, while "Early-rising" patients may not require excessive concern if their trajectory stabilizes, as their outcomes were comparable to low-risk patients.
• Limitations: The authors acknowledge the study's retrospective, single-center nature and the heterogeneity in CEA sampling frequency. They suggest that the negative findings regarding inflammatory markers may be due to the specific timing of blood collection (postoperative recovery phase) or the fact that CEA trajectories may already encapsulate systemic inflammatory signals. Future work should focus on multicenter validation and the integration of circulating tumor DNA (ctDNA).