The widespread occurrence of CRC has led to substantial research focused on improving patient outcomes and their overall quality of life. We developed a competing risk model to forecast cancer-specific mortality in patients. It has demonstrated exceptional accuracy in predicting mortality risk for colon cancer. In the training set, the overall C-statistic was 0.837, and the areas under the ROC curve for predicting 1-, 3-, and 5-year CSS were 0.826, 0.843, and 0.836, respectively. The model predicted an overall C-statistic of 0.885 for the validation cohort and areas under the ROC curve of 0.833, 0.843, and 0.835 for 1-, 3-, and 5-year CSS. The competing risk model, constructed based on these independent predictors, exhibited highly satisfactory predictive performance.

The competing risk model is employed to assess scenarios with multiple potential outcomes (events), where the occurrence of one event prevents others from occurring [13]. This model describes the event of interest’s probability occurring within a specified time using the cumulative incidence function (CIF), while accounting for the effects of other competing events. Typically, it employs non-parametric methods (like variants of the Kaplan-Meier estimator) or semi-parametric methods (such as the Fine and Gray model) to estimate the cumulative incidence function and the influence of covariates on event risk [14]. Moreover, when competing events are present, it is advisable to develop a tumor-specific mortality prediction model based on the competing risk model.

Recent studies have constructed several predictive nomograms for all-cause mortality risk using SEER big data, which are also applicable to CRC [15,16,17,18,19,20]. Although the C-statistic of their predictive nomograms for OS ranged from 0.7 to 0.8, indicating reasonably desirable predictive accuracy, there is still room for improvement. Additionally, we observed that the proportion of deaths due to other causes was significant among all colon cancer patients who died during the follow-up period. In our training cohort, deaths from other causes made up approximately 35.92% of the total mortality. While the SEER data is invaluable for building prognostic score models due to its extensive coverage, it lacks factors that can predict deaths from other causes. For all-cause mortality, deaths from other causes may be seen as a “noisy” factor. We are also concerned that the cancer-specific mortality prognostic models constructed by these studies had C-statistics between 0.7 and 0.8, which could be attributed to using traditional survival analysis models (Cox regression models). In contexts with numerous competing events, the traditional survival analysis approach may overestimate cancer-specific mortality, and the influencing factors and weights can be highly biased. In our study, the C-statistic was significantly improved, indicating that the competing risk model offers more appropriate conditions for high levels of competing events.

Additionally, compared to other predictive tools, Perineural and CEA have been incorporated into our predictive nomogram. In cases of colon cancer without lymph node metastasis (i.e., stage III pN1c tumors), TD need to be considered for prognosis and staging decisions. The presence of TD correlates with poorer DFS and OS. A negative impact of TD was noted in both pN1a/b and pN2 groups. Our predictive model integrates the number of TD and lymph node metastases to enhance the accuracy of TNM staging prognosis prediction [19]. Elevated CEA levels above 5 µg/L are linked with poor prognosis in newly diagnosed CRC patients [21]. Serial measurements of CEA demonstrate sensitivity and specificity rates of approximately 80% and 70%, respectively, for identifying recurrent CRC. Monitoring CEA levels in CRC patients following primary treatment has proven effective in detecting cancer recurrence in follow-up after colorectal surgery (FACS) trials, and these recurrences can be treated successfully. National guidelines in North America and Europe also support CEA measurement during postoperative follow-up of CRC patients [22]. Another study [23] suggested that vascular and perineural invasion, combined with postoperative CEA levels, might serve as significant indicators for early relapse of colon cancer after surgery. Importantly, perineural infiltration was highlighted as a particularly notable risk factor in rectal cancer cases. The incidence of PNI in colorectal cancer varies, with estimates ranging from 9 to 30%, and it is more commonly found in advanced stages of the disease. Research indicates that PNI occurs in about 10% of cases in stages I-II, may affect up to 30% of patients in stage III, and can reach as high as 40% in stage IV CRC [24,25,26]. It is well-documented that PNI serves as an independent indicator of an unfavorable prognosis and is associated with reduced survival rates in cancer patients [24, 25, 27, 28]. Identifying CRC patients at high risk for early recurrence is crucial, as these individuals often exhibit a significantly reduced overall survival rate compared to those without early recurrence, regardless of whether they have colon or rectal cancer. Early recognition of this risk group can guide more targeted and aggressive treatment strategies to enhance clinical outcomes.

Our study presents the following limitations. First, the SEER database lacks detailed information on radiotherapy and chemotherapy methods, including whether treatments like neoadjuvant or immunotherapy were administered, which appears to limit the analysis of survival status predictors. Secondly, the SEER database does not include patients’ medical history and detailed laboratory indicators, potentially restricting the model’s predictive performance. Thirdly, our constructed model requires additional external validation, which we plan to address in future research.

Our research discovered that colon cancer is associated with numerous alternative causes of death, which are frequently challenging to forecast. The mortality risk model we constructed, grounded on a comprehensive dataset from the SEER database, demonstrates optimal accuracy, albeit with limited predictors. Consequently, future studies should investigate more effective predictors, such as patients’ responses to chemotherapy and immunological biomarkers, to improve predictive accuracy. Furthermore, efficient model types apt for scenarios with multiple competing events should also be developed.