We included patients who underwent radical colorectal cancer resection at Hebei Provincial People’s Hospital and Shijiazhuang People’s Hospital between January 2010 and December 2020. The inclusion criteria were as follows: (1) post-operative pathological diagnosis of colon adenocarcinoma; (2) age greater than 18 years at the time of diagnosis. The exclusion criteria were as follows: (1) history of any previous cancer or distant metastasis; (2) failure to undergo radical resection for any reason; (3) missing essential clinical or follow-up information (e.g., gender, age at diagnosis, tumor size, TNM stage, or CEA levels); (4) other histological types of tumors; (5) prior neoadjuvant chemotherapy or radiotherapy; (6) post-operative survival time of less than one month; (7) tumors located at the rectosigmoid junction or the rectum.

Clinical and pathological data were extracted from the hospital’s electronic medical records, including various variables such as patient age, gender, marital status, Carcinoembryonic antigen, TNM stage, T stage, N stage, histological grade, maximum tumor diameter, tumor location, number of tumor deposits, survival time, date of death, and survival status. Postoperative pathology results were independently reviewed by two experienced pathologists. In cases of diagnostic discrepancies, a third senior pathologist reviewed the samples. A definitive diagnosis was established when at least two pathologists reached a consensus. Right-sided colon cancer was defined as tumors originating in the cecum, ascending colon, hepatic flexure, or transverse colon, while left-sided colon cancer was defined as tumors originating in the splenic flexure, descending colon, or sigmoid colon. Patient TNM staging was determined according to the eighth edition of the UICC (Union for International Cancer Control) TNM classification for malignant tumors. The primary endpoint of this study was overall survival (OS). Overall survival time was calculated from the date of surgery until the date of death or last follow-up. Death was treated as events in the analysis.

In this study, univariate Cox regression analysis was performed using the “survival” package in R software to assess the relationship between clinical-pathological variables and patient survival time in the training set. The clinical-pathological variables included age, gender, marital status, T stage, N stage, TNM stage, histological grade, primary tumor location, tumor size, TD, and CEA levels. Variables with statistical significance were then included in a multivariate Cox regression analysis to calculate the hazard ratios (HR) and their corresponding 95% confidence intervals (CI). The Cox multivariate analysis was used to select statistically significant predictors.

Tumor deposits refer to focal aggregates of tumor cells in the pericolic or perirectal mesenteric fat, which are distinct from the primary tumor and not associated with a lymph node. Histological slides and reports were reviewed to collect the following data. TD(-) indicates the absence of tumor deposits, while TD(+) indicates the presence of one or more tumor deposits. Patients were grouped by CEA status as follows: (1) patients with preoperative CEA ≤ 5.0 ng/mL (CEA(-)); (2) patients with preoperative CEA > 5.0 ng/mL (CEA(+)). A CEA-TD scoring system was developed based on the expression levels of CEA and TD. The scoring system is as follows: 0 points for CEA(-) and TD(-); 1 point for CEA(+) and TD(-); 2 points for CEA(-) and TD(+); and 3 points for CEA(+) and TD(+). Patients were classified into four categories based on the score.

A nomogram is a graphical representation of mathematical relationships, commonly used to estimate the results of a formula visually. In this study, a long-term survival prediction model for patients was developed based on the CEA-TD score combined with other clinical variables to assess the prognosis of colon cancer (CC) patients. Survival curves, calibration plots, concordance indices, and ROC curves were generated to evaluate the predictive ability of the nomogram for prognostic factors in both the training and validation cohorts.

Statistical analyses were performed using SPSS software (version 25.0, Chicago, IL, USA) and GraphPad Prism 9. Categorical data are presented as frequencies (percentages), while continuous data are expressed as medians (interquartile ranges). The comparison of two categorical variables was performed using the chi-square test, Fisher’s exact test, or the Mann-Whitney U test, as appropriate. For comparisons among multiple groups, the Kruskal-Wallis test was applied. The Kaplan-Meier survival curve was used to assess overall survival in colon cancer patients, and the log-rank test was employed to compare differences between groups. A p-value of <0.05 was considered statistically significant.

A total of 1, 029 patients with locally advanced colon cancer were included in this study, comprising 715 patients from Hebei Provincial People’s Hospital and 314 patients from Shijiazhuang People’s Hospital (Figure 1). Of the 715 patients from Hebei Provincial People’s Hospital, 429 were randomly assigned to the training set and 286 to the validation set using a 60:40 ratio, based on the “caret” package in R software. The 314 patients from Shijiazhuang People’s Hospital were allocated to the external validation set. The proportions of TD(+) in the three cohorts were as follows: 15.9% (68/429) in the training set, 14.6% (42/286) in the validation set, and 16.9% (53/314) in the external validation set. The proportions of CEA positivity were 39.6% (170/429) in the training set, 40.6% (116/286) in the validation set, and 39.2% (123/314) in the external validation set. In the training cohort, there were 125 KRAS mutant, 205 wild-type, and 99 unknown cases; in the internal validation cohort, 80 mutant, 146 wild-type, and 60 unknown; and in the external validation cohort, 90 mutant, 158 wild-type, and 66 unknown. No significant differences in clinical factors were observed between the two validation sets and the training set (Table 1 for further details).

A comparison of clinical and pathological features between TD(+) and TD(-) patients revealed that TD(+) patients had higher tumor stages, T stages, and a higher proportion of CEA positivity compared to TD(-) patients, with statistically significant differences (P<0.05). However, no significant differences were observed in relation to age, gender, tumor size, tumor location, or histological grade (Table 2).

Patients were categorized into four groups based on the CEA-TD score: 0 points (CEA(-), TD(-)), 1 point (CEA(+), TD(-)), 2 points (CEA(-), TD(+)), and 3 points (CEA(+), TD(+)). Univariate Cox regression analysis showed a significant association between survival time and the CEA-TD score. Additionally, the multivariate Cox regression analysis show that T stage, N stage, TNM stage, histological grade, and CEA-TD score were all significantly related to patient prognosis. Table 3 provides the estimated regression coefficients and hazard ratios for each variable.

TD(+) patients were significantly associated with poor OS compared to TD(-) patients (HR = 4.264, 95% CI: 2.911-6.246, P<0.001). Similarly, CEA(+) patients had significantly worse OS than CEA(-) patients (HR = 2.437, 95% CI: 1.608-3.692, P<0.001). In the N stage, patients with lymph node metastasis had worse prognoses compared to N0 patients (N1: HR = 2.588, 95% CI: 1.699-3.941, P<0.001; N2: HR = 4.670, 95% CI: 2.970-7.343, P<0.001) (Figure 2). When patients were grouped based on both TD and CEA, significant survival differences were observed among the four groups. Further, survival times also differed significantly among the four groups in Stage III patients (Figure 3, Supplementary Table 1). To further explore the prognostic ability of the CEA-TD score, the area under the curve (AUC) was plotted, as shown in Figure 4.

A new nomogram was developed to estimate the prognosis of colon cancer patients using the CEA-TD score combined with four additional clinical variables: T stage, N stage, histological grade, and tumor location (Figure 5A). In the nomogram, each variable is represented by a vertical scale indicating its value range. By aligning the variable values and observing the intersections on the nomogram, survival probabilities at 1, 3, and 5 years can be estimated. Subsequently, calibration curves, ROC curves, and time-dependent concordance index (C-index) curves were generated to assess the predictive performance of the nomogram, which demonstrated good applicability and accuracy. The AUC of the nomogram in the training set was 0.796 at 1 year, 0.834 at 3 years, and 0.807 at 5 years, outperforming other variables in the internal validation set, where the AUCs were 0.801 at 1 year, 0.831 at 3 years, and 0.807 at 5 years (Figures 5B, 6). The C-index of the nomogram was 0.800 (95% CI: 0.763-30.838), also higher than other clinical variables (Figure 7). Moreover, patients were stratified into high-risk and low-risk groups based on the median risk score calculated from the nomogram. Patients in the high-risk group had shorter survival times compared to those in the low-risk group (Figure 7). The nomogram risk scores for patients with different survival statuses are shown in Figure 7, indicating that an increase in the risk score corresponds to a higher mortality rate in CC patients. Finally, risk scores derived from the same formula used in the nomogram were calculated for both the internal and external validation sets. In both validation cohorts, patients with low-risk scores had better prognoses than those with high-risk scores. These results suggest that the nomogram can accurately and conveniently predict the prognosis of LOCC patients.

Patients were classified into three groups based on preoperative and postoperative (4–6 weeks) serum CEA levels: normal both pre- and postoperatively, elevated preoperatively but normalized postoperatively, and persistently elevated. In the training cohort, the groups included 259, 92, and 78 patients, respectively; in the internal validation cohort, 170, 65, and 51; and in the external validation cohort, where postoperative CEA data were missing for 40 patients, the groups included 162, 63, and 49, respectively. Kaplan–Meier analyses showed significantly different survival among the three groups (p<0.001), with the best outcomes in patients with consistently normal CEA, intermediate outcomes in those with normalization postoperatively, and the worst outcomes in those with persistently elevated CEA (Supplementary Figure 1). These findings suggest that dynamic changes in CEA may more accurately reflect prognostic risk.

Across all three cohorts, the model effectively stratified overall survival within both KRAS mutant and wild-type subgroups. High-risk patients consistently showed significantly worse outcomes than low-risk patients (all p<0.001), confirming the robustness of the CEA-TD based model irrespective of KRAS status (Figure 8).