The study, designed as a retrospective cohort analysis, received ethical approval from the Institutional Review Board of the Second Affiliated Hospital of Zhejiang University School of Medicine (2020LSYD829). Patients with stage I to IIIA NSCLC who received curative surgery at the Second Affiliated Hospital between June 2015 and September 2023 were consecutively recruited. The following were the criteria for participants to be included in this research: 1) Histologically confirmed diagnosis of primary NSCLC. 2) No history of other cancers. 3) No local surgical treatments other than radical resection. 4) Negative surgical margins with no residual lesions. 5) No neoadjuvant therapy.

Serum CEA concentrations were measured via chemiluminescent immunoassay using the COBAS 8000 e602 analyzer (Roche Diagnostics, Tokyo, Japan), in accordance with the World Health Organization (WHO) international standard 73/601 (17). The pre-surgical CEA level was defined as the concentration measured within 30 days prior to resection. Patients with preoperative measurements were included in preoperative -related studies ( Supplementary Figure S1a ). The postoperative CEA level referred to the measurement closest to the surgical date, taken within seven months after resection. Patients with postoperative measurements were included in postoperative-related studies ( Supplementary Figure S1b ). For trajectory analysis, patients who had CEA measurements taken before the operation and whose CEA levels were measured at least three times during the 36-month period following the surgery were enrolled. All included CEA measurements after surgery were taken before any clinical outcomes occurred.

The study’s follow-up was completed on November 2, 2023. Recurrence-free survival served as the primary outcome, capturing both local and metastatic disease events. The secondary endpoint focused on overall survival (OS). Recurrence-free duration was calculated from the date of surgery to the time of relapse detection. which was established through histological examination of biopsy specimens or positive findings on imaging studies. Survival information was obtained through telephone follow-ups.

Covariates included age, sex, smoking status (yes or no), surgical methods (wedge resection, segmentectomy, or lobectomy), pathological stage (I-IIIA), histological subtype (adenocarcinoma or squamous cell carcinoma), and tumor differentiation grade (highly - matured, intermediately - matured, or poorly - matured). Preoperative CEA levels were also accounted for in the analysis of the relationships between trajectories and NSCLC outcomes. Data were retrieved from internal departmental records and digital medical archives. The pathological staging was assigned based on the 8th edition of the American Joint Committee on Cancer (AJCC) classification criteria for lung malignancies.

The study workflow is depicted in the diagram shown in Figure 1 . All the statistical analyses were conducted in R (v4.4.2), with missing values addressed through multiple imputation using the “mice” package (v2.1.0). We looked into the differences between groups. For continuous variables (stated as median [interquartile range]), the Kruskal - Wallis test was replaced with a non - parametric bootstrap - based test to assess group differences. Categorical data (presented as n [%]) were compared using the G-test, which serves as a likelihood-ratio alternative to the classical chi-square test.

CEA was classified as elevated when exceeding 5 ng/mL, and normal when at or below this threshold, according to standard reference values. Combining preoperative and postoperative levels further categorized into normal preoperative, normalization, and persistently elevated groups. Given the substantial variability within CEA measurements, values were capped at tenfold the upper reference limit to improve trajectory fitting. The logarithms of these truncated values were subsequently employed to model trajectories, due to the fact that they exhibited a non-normal distribution.

The application of a latent class growth mixed model (LCGMM) enabled the identification of distinct trajectory patterns of perioperative CEA levels. This method segments the population into latent classes based on estimated heterogeneity and constructs individual trajectories using a linear mixed model framework (18, 19). We modeled the long - term variations in the measured values as either linear or non - linear relationships with respect to the passage of time. These relationships included time, its second - order or third - order terms. We explored 2 to 5 potential trajectory groups. The number of optimal clusters and the most suitable curve form were selected based on the Bayesian Information Criterion (BIC), with the constraint that each class represented ≥ep of the cohort and exhibited an average posterior probability above 0.7. The “he.eb package (version 1.9.2) was used to perform the latent class growth mixture modeling. Kaplan-Meier curves were generated for overall survival (OS) and recurrence-free survival (RFS), with intergroup differences evaluated using the log-rank test. Associations between trajectory classes and outcomes were further analyzed using Cox regression models. The proportional hazards assumption was assessed using log-log plot. We made use of three models: the first was unadjusted; the second controlled for age, sex, and smoking status; and the third further accounted for surgical type, pathological stage, histological subtype, and tumor differentiation. In analyses involving CEA trajectory groups, preoperative CEA levels were also taken into account. We also constructed three more statistical models incorporating preoperative CEA, postoperative CEA, and both, respectively. Model comparisons were conducted via likelihood ratio tests (LRT). To flexibly characterize nonlinear effects, Restricted cubic splines (RCS) modeling was applied to assess how continuous pre- and postoperative CEA levels relate to recurrence risk in NSCLC. In survival analysis, RCS can model nonlinear relationships in Cox proportional hazards regression models (20, 21). Three knots were selected when modeling. The model took into account factors such as patient’s age, gender, smoking status, types of surgical procedures, disease stage based on pathology reports, pathological characteristics, and the extent of cellular differentiation.

The study included 5733 individuals diagnosed with NSCLC. To elaborate, 5309 were incorporated into the preoperative analysis, 2027 into the postoperative investigation, 1860 into the study examining dynamic changes in the perioperative, periods and 988 into the CEA trajectory analysis. A summary of participant evaluation and exclusion rationale can be found in Figure 2 .

The study included 988 patients (435 [44.03%] male; median age [interquartile range, IQR]: 60 years [52-67]). The median follow - up duration was 29.54 months (IQR: 18.33-38.21). Recurrence occurred in 98 patients (9.92%) during follow-up, yielding an incidence of 5.44 per 1,000 person-years. Meanwhile, 48 patients (4.86%) died, with a corresponding rate of 2.50 per 1,000 person-years. A total of 6298 individual CEA measurements were assessed, averaging a median of 6 measurements per patient (range: 4-36) ( Supplementary Figure S2 ). The LCGMM model fitting results are presented in Supplementary Table S1 . In line with the predefined guidelines, non - linear curves representing three possible clusters were considered the best fit for CEA modeling. Figure 3A depicts the perioperative trajectory of CEA concentration. Three separate trajectory groups were pinpointed for perioperative CEA: low-stable (n = 926, 93.72%), early-rising (n = 34, 3.44%), and later-rising (n = 38, 3.85%). Among the members of the low-stable group, CEA levels consistently remained below 5.0 ng/mL from the preoperative period through 36 months after surgery. In the early -rising group, CEA decreased to its lowest value at 5 months post-surgery, followed by a rapid increase. In the later - rising group, CEA dropped to the standard reference interval within 6 months post - surgery and gradually ascended after 10 months.

Relative to the group with consistently low CEA levels, the groups demonstrating an early increase and a late increase in CEA levels were characterized by lower tumor stages and differentiation levels. Additionally, fewer patients in these rising groups underwent wedge resection or segmentectomy. Both rising groups also demonstrated elevated risks of recurrence and mortality ( Table 1 ). The 5-year RFS rates were 23.85% (95% CI: 12.10%–47.02%) for the early-rising group, 32.00% (95% CI: 12.00%–85.30%) for the later-rising group, and significantly higher at 86.15% (95% CI: 81.99%–90.52%) in the low-stable group (P < 0.001; Figure 3B , left). A similar trend was observed for overall survival: the 5-year OS rate reached 94.55% (95% CI: 92.00%–97.18%) in the low-stable group, compared to 33.93% (95% CI: 19.47%–59.13%) in the early-rising group and 80.50% (95% CI: 61.60%–100.00%) in the later-rising group (P < 0.001; Figure 3B , right). Sub-analysis stratified by histology shows that in adenocarcinoma patients, the patterns in 5-year RFS and OS across the three CEA trajectory groups are highly consistent with those observed in the overall population ( Supplementary Figures S5a, b ).

Tables 2 , 3 summarize the relationships between perioperative CEA trajectories and both RFS and OS. Relative to the low-stable group, the risk of recurrence was significantly higher in the early- and late-rising groups, with hazard ratios (HRs) of 10.84 (95% CI: 5.57–21.10, P < 0.001) and 3.56 (95% CI: 1.62–7.81, P = 0.002), respectively. After adjustment for demographic characteristics and preoperative CEA, the associations were slightly attenuated but remained significant. The early-rising group showed a markedly higher risk of death (HR = 13.37, 95% CI: 5.45–32.81, P < 0.001) relative to the low-stable group, whereas the mortality risk for the later-rising group was not statistically significant (HR = 2.00, 95% CI: 0.58–6.93, P = 0.276) following full adjustment.

Figure 4 displays Kaplan-Meier survival curves stratified by preoperative, postoperative, and perioperative CEA profiles. The 5Cox model revealed that varying CEA levels were significantly associated with NSCLC prognosis, as shown in Tables 2 , 3 . Higher preoperative log2-transformed CEA was associated with increased risk of recurrence (HR = 1.31, 95% CI: 1.19-1.46) and mortality (HR = 1.24, 95% CI: 1.10-1.40). Individuals with elevated preoperative CEA levels demonstrated significantly higher risks of recurrence (HR = 2.49, 95% CI:1.80-3.44) and mortality (HR = 2.26, (95% CI: 1.54-3.33) compared to those with normal preoperative CEA levels. Likewise, higher postoperative log2-transformed CEA was associated with increased risk of recurrence (HR = 1.53, 95% CI: 1.29-1.81) and mortality (HR = 1.72, 95% CI: 1.42-2.09). Individuals with elevated postoperative CEA levels demonstrated significantly higher risks of recurrence (HR = 2.64, 95% CI: 1.65-4.23) and mortality (HR = 3.34, 95% CI: 2.09-5.80) compared to those with normal postoperative CEA levels. After adjusting for preoperative CEA, postoperative CEA remained an independent predictor of mortality (HR = 2.13, 95% CI: 1.42–2.09), as shown in Supplementary Table S2 .

Individuals with consistently elevated CEA levels both preoperatively and postoperatively exhibited markedly higher risks of recurrence and mortality, with HRs of 3.00 (95% CI: 1.77-5.10) for recurrence and 3.30 (95% CI: 1.79-6.07) for death. Patients whose CEA levels normalized after surgery still faced elevated risks of recurrence and death—HRs of 1.82 (95% CI: 1.17–2.85) and 1.80 (95% CI: 1.01–3.22), respectively—although these risks remained lower than in those with sustained CEA elevation.

Through multivariable-adjusted polynomial regression analysis ( Figure 5 ), a positive nonlinear relationship was found between NSCLC recurrence risk and preoperative CEA levels (P for non-linearity=0.007). As preoperative CEA levels increased, so did the recurrence risk. The points where the hazard ratio for preoperative CEA levels equals 1.0 is approximately at 2.1 ng/mL. Below this threshold, the recurrence risk for NSCLC is less than 1.0, whereas above it, the recurrence risk exceeds 1.0. However, no non-linear association was detected between the levels of postoperitive CEA and the occurrence of recurrence (P for non-linearity =0.381).