As a major global health challenge, colorectal cancer (CRC) ranks third in incidence and is the second leading cause of cancer deaths worldwide (1). While multimodal treatment strategies have reduced local recurrence rates, high rates of distant metastasis remain a significant challenge, leading to a need for improved overall survival.

Distant metastasis is the primary cause of death in CRC patients, with approximately 20% of patients presenting with distant disease at diagnosis and about one-third developing metastasis after curative surgery (2, 3). Recurrence typically occurs within three years, with distant metastases frequently observed in the liver, lungs, and peritoneum (4). The prognosis for metastatic colorectal cancer (mCRC) remains exceedingly poor, with a 5-year survival rate of less than 20% in the absence of medical intervention (5).

In colorectal cancer, the lung is the second most common site of distant metastasis after the liver, and the most frequent extra-abdominal site. Although timely radical resection combined with systemic chemotherapy improves survival, a substantial proportion of patients still develop pulmonary recurrence after curative surgery (6, 7). Data from the PulMiCC program reveal that the natural history of CRC lung metastases is less aggressive than previously thought, with 5-year survival rates exceeding the historically dismal rate of approximately 0% without oncological intervention (8, 9), These findings highlight the biological and prognostic differences between pulmonary and other visceral metastases (e.g., hepatic or peritoneal), underscoring the need for dedicated risk-stratification tools to estimate individual postoperative recurrence risk. Multimodal treatment initiated upon radiological detection—including perioperative chemotherapy, pulmonary metastatic lesion resection, stereotactic body radiotherapy (SBRT), or percutaneous thermal ablation—can achieve 5-year survival rates of 20–70% (10–13). Therefore, early identification of patients at high risk for pulmonary metastasis following radical resection could help improve survival outcomes and promote personalized approaches in cancer treatment (14–16).

While nomograms have proven valuable in predicting risk and prognosis for various cancers (17, 18), there is a lack of tools specifically designed to predict lung metastasis in CRC patients following radical surgery.

This study aims to develop a predictive model for identifying high-risk factors associated with lung metastasis in CRC patients following radical surgery. By forecasting the risk of lung metastasis, this model would assist clinicians in prognostication and the development of personalized follow-up and treatment strategies for CRC patients after surgery.

Colorectal cancer (CRC) is a significant global health issue, being the third most common cancer and the second leading cause of cancer deaths worldwide (1). Although advances in mesenteric-fascial plane surgery have significantly reduced locoregional recurrence, distant metastasis remains the principal cause of mortality in these patients. The lung is the most frequent site of extra-abdominal metastasis (7). According to relevant expert consensus guidelines, various treatment modalities, including surgical resection of metastatic lesions and systemic chemotherapy, can increase the 5-year survival rate to 35-70% in patients with pulmonary metastasis after radical surgery. Particularly for patients with isolated and resect able pulmonary metastasis, active medical intervention can significantly improve the prognosis of this population (21, 22).

Therefore, the prompt detection and aggressive treatment of lung metastases following curative-intent CRC resection are critically important. The construction of predictive models to identify high-risk patients prone to pulmonary metastasis following radical resection can assist surgeon in developing individualized follow-up strategies for different patients, thereby advancing precision medicine.

We developed and externally validated a clinicopathological nomogram to predict the probability of postoperative pulmonary metastasis (PPM) following curative-intent resection of CRC. These variables span surgical, serological, and histological domains, collectively forming a comprehensive risk profile. Each factor captures a distinct aspect of tumor biology, encompassing intrinsic aggressiveness, the systemic host response, and the impact of the surgical approach. Consequently, this predictive model provides greater discriminatory accuracy than algorithms based on a single variable class, such as pathological stage alone. The nomogram translates the multivariable equation into a graphical calculator, facilitating individualized risk quantification at the bedside. In the training cohort (n = 252), the model demonstrated good discrimination and calibration, with a sensitivity of 71.9% and specificity of 73.3%, and a Hosmer-Lemeshow test P-value of 0.475. External validation using an independent cohort (n = 147) confirmed the model’s robust performance, with an area under the curve (AUC) of 0.779, and adequate calibration, with a Hosmer-Lemeshow P-value of 0.331. These results indicate that the nomogram maintains predictive accuracy across diverse patient populations and is suitable for implementation in clinical practice. DCA showed that the nomogram yields a higher net clinical benefit than both the “treat-all” and “treat-none” strategies across a wide range of threshold probabilities. Furthermore, using the median nomogram-derived risk score as the cutoff, patients were divided into high-risk and low-risk cohorts with significantly different lung-metastasis-free survival (LMFS) in both the training and validation cohorts. The median time to pulmonary metastasis was significantly shorter in the high-risk group, highlighting the model’s utility for prognostic stratification and for identifying patients who might benefit from intensified surveillance or adjuvant therapy. By quantifying individual PPM risk, the nomogram facilitates personalized postoperative follow-up strategies. High-risk patients can be enrolled in CT-intensive surveillance programs to facilitate early treatment of oligometastatic disease. In contrast, low-risk patients can be spared unnecessary imaging, thereby optimizing resource allocation and improving cost-effectiveness. Additionally, the nomogram may guide decisions regarding adjuvant therapy. Although adjuvant chemotherapy is standard for stage III CRC, its benefit in stage II disease remains controversial (23). Integrating the nomogram into multidisciplinary discussions could help identify stage II patients with a high cumulative risk who warrant adjuvant therapy, while sparing low-risk patients from unnecessary cytotoxic exposure. In summary, this nomogram provides a practical, evidence-based tool for optimizing both surveillance protocols and adjuvant treatment strategies in patients who have undergone curative resection for CRC.

This study unequivocally demonstrates that preoperative CEA levels serve as an independent risk factor for lung metastasis following R0 resection of CRC. Serum CEA level is involved in intracellular recognition and facilitates adhesion between tumor cells and host cells, which may explain the observed increase in lung metastasis after colorectal cancer surgery (24). It is widely accepted that elevated preoperative CEA levels can facilitate tumor metastasis and postoperative dissemination (25). In this study, CEA measurement was limited to the final preoperative assessment; postoperative levels were not obtained. Consequently, the association between preoperative CEA levels and the development of pulmonary metastasis may not be fully characterized. Future studies should incorporate serial measurements of postoperative CEA to better elucidate the relationship between its dynamics and metastasis. Currently, the tumor-node-metastasis (TNM) classification is a widely endorsed staging system for colorectal cancer. The extent of lymph node involvement has been closely associated with the occurrence of distant metastases (26, 27). During tumor progression, neoplastic cells can disseminate via blood vessels, lymphatic vessels, or nerves (28). Previous research has demonstrated a significant association between perineural invasion and both the malignant pathological phenotype and poor prognosis in CRC. Additionally, perineural invasion has been identified as a prognostic indicator for the efficacy of adjuvant chemotherapy in CRC (29, 30). Given the heterogeneity of molecular subtypes in colorectal cancer patients, it remains challenging to definitively ascertain whether N staging and perineural invasion independently contribute to the development of pulmonary metastasis. In our study, utilizing both univariate and multivariate logistic regression analyses, preoperative CEA level (P = 0.045), N-stage (P<0.001) and perineural invasion (P = 0.047) were identified as independent risk factors for postoperative lung metastasis in colorectal cancer and assigns weighted scores to each. This approach enhances the model’s reliability and credibility.

The protective effect of laparoscopic surgery may stem from several factors, including minimized surgical trauma and reduced postoperative immunosuppression (31). This minimally invasive approach helps preserve host immune function, a key mechanism in controlling occult micro-metastases that may be present at resection. Conversely, open surgery, characterized by larger incisions and more extensive tissue manipulation, can promote tumor cell shedding into the circulation, thereby elevating the risk of distant metastasis. These findings emphasize that the surgical approach is a modifiable risk factor in colorectal cancer management, highlighting the potential of minimally invasive techniques to reduce postoperative recurrence.

The application of machine learning (ML) models for predicting cancer prognosis has garnered growing interest in recent years. Guo et al. recently developed an ML model using the SEER database to predict lung metastasis in patients with CRC (32). The random forest model achieved notably high AUC values: 0.980 in the internal test set and 0.927 in the external validation set. Although the AUC values exceed those of our nomogram, differences in study design must be considered. Specifically, Guo et al. employed a substantially larger dataset (n=39,674) and included a wider array of predictor variables. However, the clinical relevance and interpretability of predictor variables are also critical considerations. The variables in our nomogram were selected via logistic regression and represent commonly used clinical indicators that are readily interpretable to clinicians, enhancing the model’s practical utility. Furthermore, using a more parsimonious set of predictors mitigates the risk of overfitting.

This retrospective study has several inherent limitations. Data extraction from electronic medical records inevitably resulted in missing information and introduced potential selection bias. To mitigate these biases, we implemented stringent inclusion and exclusion criteria, which included the exclusion of patients with incomplete clinical data. Of the 692 patients initially screened, only 399 ultimately met all eligibility criteria and were enrolled in the final analysis. In this process, patients with incomplete follow-up data were likewise excluded, which may have selected for a cohort more adherent to surveillance protocols, potentially biasing the outcomes. Therefore, future prospective controlled trials are needed to validate these findings with greater scientific rigor. As only routinely documented variables were available, we could not incorporate emerging biomarkers such as circulating tumor DNA (ctDNA). Future studies should incorporate a broader range of tumor-related variables, including novel tumor markers, into the model, thereby enhancing both the innovativeness and objectivity of the predictive model. These inherent shortcomings of retrospective analyses preclude definitive causal inferences. The study can only identify associations, and observed relations hips may be influenced by unmeasured confounders. For instance, the higher incidence of pulmonary metastasis following open resection may reflect more advanced disease or unreported surgical comorbidities (e.g., intestinal obstruction or tumor perforation) rather than a direct effect of the surgical approach. These unmeasured confounding factors could bias the statistical outcomes. Furthermore, given the rising adoption of robotic surgery, future studies should analyze it as a distinct cohort to clarify its specific impact.

The two center design with both institutions located in the same region of China, further limits the generalizability of our findings. Before this nomogram can be recommended for widespread clinical use, it is essential to conduct external validation across populations from different regions, as this serves as an effective means of demonstrating the clinical utility of predictive models. Since a number of patients continued their care at other institutions following initial cancer resection, detailed chemotherapy regimens were not available. Consequently, patients could only be stratified as having received “adjuvant chemotherapy” or “no adjuvant chemotherapy.” This lack of granular data on systemic therapy represents a potential unmeasured confounder that may affect the observed incidence of pulmonary metastasis. At the same time, we could not assess whether specific chemotherapeutic regimens differ in their efficacy for preventing pulmonary metastasis. To improve the clinical utility of this model, future validation should include more regional populations.

In our study, the occurrence of pulmonary metastasis was defined as the endpoint event, and patient follow-up was consequently terminated at that point. Consequently, no data were collected on subsequent survival outcomes or treatments received after metastasis. We observed that some patients transferred their care to other institutions following the diagnosis of pulmonary metastasis. Consequently, we were unable to obtain information regarding their treatment regimens following the onset of pulmonary metastasis. Therefore, assessing overall survival alone was deemed methodologically inadequate. We acknowledge that the lack of long-term follow-up data is a major study limitation. Addressing this issue is an explicit goal of our future research. Furthermore, the absence of molecular data constitutes another major limitation. Incorporating biomarkers would likely improve the model’s accuracy. Future predictive algorithms should therefore integrate comprehensive histopathological, molecular, and radiomic data. It should be noted that although some studies suggest that non-contrast chest CT scans alone can diagnose pulmonary metastases from colorectal cancer (21), in our study, the diagnosis of lung metastases was based on contrast-enhanced chest CT or PET-CT to minimize diagnostic errors. However, despite consensus readings by multiple experienced radiologists, diagnostic inaccuracies may still occur.

To address these limitations, future studies should employ a prospective, multicenter design. Prospective data collection would minimize selection bias, ensure more complete follow-up, and facilitate the integration of novel variables such as CT radiomics and liquid biopsy biomarkers. Large, ethnically diverse validation cohorts would provide the statistical power needed to confirm the nomogram’s generalizability across different patient subgroups. Ultimately, broader and richer datasets will facilitate the development of precision medicine tools for risk stratification, enabling tailored surveillance and therapy to improve outcomes for all patients with colorectal cancer.

We have developed a novel predictive model for lung metastasis in CRC patients following radical surgery. This model integrates clinical data and pathological features, including carcinoembryonic antigen (CEA) levels, surgical approach, N staging, and perineural invasion, into a comprehensive scoring system. This model can assist clinicians in assessing the risk of pulmonary metastasis after radical colorectal cancer surgery, providing a reference for personalized treatment and follow-up. Its clinical intervention effect needs to be further verified by prospective studies.