A consecutive cohort study of patients with stages II-III CRC who underwent radical surgery at the CNCC was retrospectively conducted between January 2009 and December 2019. The tumor location was determined according to the surgical records. Of the 982 stage II-III patients in the database, 903 patients met the inclusion criteria ( Figure 1 ). The tumor location was determined according to the surgical records and rectosigmoid carcinomas were classified as sigmoid carcinomas. The dominant (clinically more advanced) cancer data were collected if the patient presented simultaneously with multiple colon or rectal cancers. The inclusion criteria were as follows: (1) patients with stage II-III CRC who underwent radical surgery (R0 resection); (2) pathological type of adenocarcinoma (including mucinous adenocarcinoma and signet ring cell carcinoma); and (3) complete medical records and continuous follow-up records. The exclusion criteria were as follows: (1) familial hereditary rectal malignant tumor; (2) previous history of malignant tumor; and (3) severe complications (Clavien–Dindo grade 4) or local recurrence within 3 months after radical surgery. The CNCC Committee approved this study (20/355-2551).

One of the difficulties in assessing the outcomes of treatment for CRC in metachronous recurrences is that there is no consistent definition of the metachronous events in the published literature. We considered local recurrences as those diagnosed more than 3 months after radical surgery. In addition, locally recurrent CRC has been defined heterogeneously in the literature and between different institutions (14). Especially in China, surgeons often describe local recurrence as confined to the pelvic or abdominal cavity without distant metastasis, which makes it easier to analyze independent risk factors for long-term survival after local recurrence in patients. However, this study aimed to explore the factors that cause local recurrence so it defined local recurrence CRC as recurrence at the site around the tumor bed, the anastomotic site, or tissues around the intestine or mesenteric or regional lymph nodes after R0 surgery, with or without metastasis in the parenchymal organ. The diagnosis of local recurrence was confirmed by computed tomography (CT) scan, magnetic resonance imaging (MRI), positron emission tomography (PET-CT), or pathology (biopsy or specimen). According to different recurrence sites, locally recurrent colon cancer was categorized as peri-anastomotic, mesenteric/lymph nodal, peritoneal, retroperitoneal, abdominal wall, or pelvic (15). The location of the locally recurrent rectal cancer was categorized as central, sidewall, sacral, or mixed (16). More than one type of unifocal recurrence involved was categorized as multiple.

Outcomes of interest were factors influencing LRFS and OS. Starting with the first recurrence of surgical treatment, follow-ups will be conducted every 3 months for 2 years, then every 6 months, and annually after 5 years. Evaluation at follow-up included physical examination; colonoscopy; carcinoembryonic antigen (CEA) level; carbohydrate antigen 19-9 (CA199) level; chest, abdomen, and pelvis contrast-enhanced CT; and pelvic MRI or PET-CT as necessary. LRFS was defined as the time from radical surgery to the first local recurrence. OS was defined as the time from surgical treatment to death from any cause.

The patients from the CNCC were separated into training and test sets, which are shown in Table 1 . The χ2 test was performed to compare differences in categorical variables (e.g., sex) in the composition ratios of the two groups, and the Wilcoxon rank sum test was used to determine differences in ordered variables (e.g., stage) in the distribution of the two groups. Continuous variables were transformed into categorical variables for processing. Using univariate and multivariate Cox regression analyses, the independent predictors for LRFS and OS were identified. These statistically significant factors (P<0.05) identified by univariate analysis will be incorporated into the multivariate analysis. Subsequently, nomograms were created to predict LRFS and OS based on the independent predictors identified by multivariate Cox regression analyses. Additionally, the bootstrap method was used for internal and external validation to evaluate the performance of the prediction model. The bootstrap-corrected OS rates were calculated by averaging the Kaplan−Meier estimates based on 2,000 bootstrap samples. The discrimination and calibration of the nomogram were evaluated using calibration curves. Receiver operating characteristic (ROC) curves and calibration curves were constructed to assess the precision of the nomogram. R software version 4.0.1 was used for the Cox proportional hazard regression analysis and construction and evaluation of the nomogram. Two-sided statistical analysis was used to calculate all P values.

A total of 903 stage II-III CRC patients [351 (38.87%) women and 552 (61.13%) men; mean age, 60 years ± 11 (SD)] who underwent radical surgery were enrolled from the CNCC set. Among these patients, 136 suffered from locally recurrent CRC within 100 months after the first surgery. The main histological type of tumor was adenocarcinoma (78.96%). For tumor size, many patients (44.41%) had tumor sizes ranging from 3 cm to 5 cm, followed by 5-10 cm (42.52%). With regard to tumor differentiation grade, most patients (75.53%) had tumors in grade 1 or 2. The patients from the CNCC were divided into training and testing sets, as shown in Table 1 .

The Cox regression analysis was applied in the CNCC set to identify the predictors of LRFS and OS. Univariate analysis indicated that vascular invasion [hazard ratio (HR): 2.99; 95% confidence interval (CI): 1.99-4.48, p<0.001], perineural invasion (2.22, 1.49-3.33, p<0.001), tumor grade (1.85, 1.30-2.63, p<0.001), T stage (3.81, 2.34-6.20, p<0.001), N stage (3.36, 2.40-4.72, p<0.001), and harvested lymph nodes (1.74, 1.05-2.87, p=0.031) were correlated with LRFS in the patients with CRC (all p<0.05) ( Table 2 ). Furthermore, the multivariate Cox analysis demonstrated that vascular invasion (1.67, 1.04-2.68, p=0.033), perineural invasion (1.65, 1.06-2.58, p=0.027), T stage (5.72, 2.86-11.41, p<0.001), N stage (2.43, 1.51-3.89, p<0.001), and harvested lymph nodes (2.78, 1.50-5.13, p=0.001) were independent risk factors for LRFS. Thus, these variables were included in the predictive model ( Table 2 ). Age at diagnosis (2.59, 1.81-3.69, p<0.001), tumor pathological type (1.50, 1.07-2.12, p=0.019), vascular invasion (2.58, 1.71-3.88, p<0.001), perineural invasion (2.54, 1.70-3.79, p<0.001), tumor grade (1.51, 1.08-2.12, p=0.017), T stage (2.94, 1.94-4.45, p<0.001), N stage (3.83, 2.79-5.26, p<0.0001) and harvested lymph nodes (2.45, 1.61-3.74, p<0.001) were correlated with OS in the patients with CRC (all p < 0.05) ( Table 2 ). In addition, the multivariate Cox analysis demonstrated that age at diagnosis (2.58, 1.63-4.09, p<0.001), tumor pathological type (1.67, 1.01-2.77, p=0.047), vascular invasion (1.99, 1.23-3.22, p=0.005), perineural invasion (1.64, 1.05-2.57, p=0.030), T stage (6.09, 3.04-12.19, p<0.001), N stage (2.48, 1.55-3.96, p<0.001), and harvested lymph nodes (2.91, 1.56-5.44, p=0.001) were independent risk factors for OS. Thus, these variables were included in the predictive model ( Table 2 ).

To evaluate the efficacy of adjuvant therapy for influencing recurrence and survival, we analyzed the prognosis before and after surgery for LRFS and OS. The Cox regression analysis suggested that adjuvant therapy did not influence the prognosis ( Table 2 ). The subgroup analysis showed that adjuvant therapy before surgery improved the LRFS and OS for patients with N1- or 2-stage CRC and adjuvant therapy after surgery improved the OS for patients with vascular invasion, perineural invasion, or grade 3 or 4 ( Figures 2A–D ). Baseline characteristics comparisons of patients with CRC between the adjuvant and no adjuvant therapy groups were performed before and after propensity score matching (PSM) using the CNCC set ( Supplementary Tables S1 , S2 ). To compare the effectiveness between the no adjuvant therapy and adjuvant therapy before surgery groups, we identified matching patients with CRC after PSM. Notably, there was no significant difference in LRFS and OS between the no adjuvant and adjuvant therapy before surgery groups, both before and after PSM, with all P values >0.05 ( Supplementary Figure S1 ). Similarly, there was no significant difference in LRFS and OS between the no adjuvant and adjuvant therapy groups after surgery, with all P values >0.05 ( Supplementary Figure S2 ).

Our results showed that adjuvant therapy could not predict the LRFS and OS for CRC patients in the CNCC set. We constructed nomogram models to predict LRFS and OS between the two sets (training sets and test sets). Vascular invasion, perineural invasion, T stage, N stage, and number of harvested lymph nodes were identified as independent risk factors for LRFS by the multivariate analysis. The risk factors that were correlated with LRFS were incorporated into the nomogram development ( Figure 3A ). In addition, age at diagnosis, tumor pathological type, vascular invasion, perineural invasion, T stage, N stage, and number of harvested lymph nodes were identified as independent risk factors for OS by the multivariate analysis. These risk factors correlated with OS were incorporated into the development of the nomogram ( Figure 3B , Supplementary Figure S3 ). The prognostic nomogram for patients with CRC was created in the training set, which was then validated in the testing set. To add clinical convenience, a user-friendly online application for the two nomograms was developed and uploaded to our website (https://oldcoloncancer.shinyapps.io/dynnomapp/).

To facilitate the use in clinical practice of these nomogram models, we divided the patients with CRC into two risk groups according to the risk scores of the nomogram models: a high-risk group and a low-risk group. We identified the cutoff values using the median risk scores in the training set and verified them in the test set. This pragmatic visualization of the risk level could help decide the optimal treatment strategy for patients with CRC. According to the risk scores from both nomogram models, the LRFS and OS of the patients in the low-risk group were better than those of the patients in the high-risk group in all sets (log-rank test, P < 0.05). Subgroup analysis was performed to evaluate the performance of these risk groups in different subgroups. The LRFS and OS of the patients in the low-risk group were better than those in the high-risk group ( Supplementary Figure S4 ).

To evaluate the discrimination ability of the two nomograms, we evaluated the AUCs ( Figures 4A–D ). We evaluated the relationship between the OS rates and the predicted probabilities in all models and staging systems. The bootstrapped calibration curves plotted with 1-,2-, 3-,4- and 5-year LRFS and OS were well matched with the idealized 45° line for both the nomograms ( Figures 4E–H ).

Among the patients in all risk groups, 250 (51.87%) in the high-risk group and 244 (57.96%) in the low-risk group were older than 60 years (Chi-squared test; p=0.077) according to the model for LRFS. There were more patients with locally recurrent colon cancer categorized as peri-anastomotic in the low-risk group than in the high-risk group ( Figure 5A ). A higher percentage of patients with locally recurrent colon cancer received R0 surgery in the low-risk group than in the high-risk group ( Figure 5B ). In addition, there were no significant differences in 3-year LRFS or recurrence sites between the two groups ( Figures 5C, D ).