Globally, colorectal cancer (CRC) ranks third in the frequency of digestive system malignancies (1). In approximately 20% of cases, distant metastases are present at the time of initial diagnosis (2, 3), resulting in a 5-year survival rate of <14%, which is further reduced in rectal cancer patients. Current treatments for metastatic CRC (mCRC) focus on improving both cancer-specific survival (CSS) and overall survival (OS), leading to the widespread utilization of systemic and palliative interventions (4). The National Comprehensive Cancer Network guidelines suggest primary tumor resection (PTR) as a surgical option for stage IV CRC patients (5). PTR, with chemotherapy, improves OS and CSS in certain patient populations (6). One study demonstrated a median OS of 18.3 months in patients that received PTR with chemotherapy, compared to 8.4 months for those who underwent chemotherapy alone (7). Despite notable advancements in the efficacy of chemotherapies, approximately 70% of patients choose PTR (8, 9). Nevertheless, determining treatment strategies for unresectable stage IV CRC depends on the expertise, discernment, and individual preferences of the attending clinician. Our knowledge about which patients derive the greatest benefit from surgery remains incomplete. Hence, a predictive model of effectively recognizing potential candidates for PTR is urgently needed.

Recently, notable advancements in machine learning algorithms in medicine have been made, emphasizing oncological applications (10). Nevertheless, the efficacies of these machine-learning models depend on the availability of substantial datasets for training, which is challenging in certain medical settings. Additionally, the opaque nature of these models can impede the comprehension of decision-making mechanisms by medical practitioners and patients. Therefore, using machine learning to address specific clinical issues is not always optimal. Traditional logistic regression models remain viable alternatives (11), providing clear and interpretable visual representations that enhance comprehensibility regarding outcomes (12, 13).

Herein, we constructed a predictive model employing data from SEER database to identify candidates suitable for PTR among mCRC patients. Additionally, the predictive performance of traditional logistic regression and machine learning models are compared.

Our findings support the initial hypothesis by evaluating the predictive capabilities of a model for selecting candidates for primary tumor surgery in mCRC. The positive outcomes highlight the efficacy of PTR for mCRC. The traditional logistic model exhibited superior performance compared with machine learning models, providing clinicians with a reliable tool to estimate the potential benefits of surgery for patients.

Stage IV CRC patients are typically managed with systemic, palliative, or end-of-life care, and local treatments, including PTR, are avoided (5). However, considerable heterogeneity among mCRC patients, including variations in age, histological subtypes, and chemotherapy protocols, can affect prognosis (18). Some studies have questioned the benefits of PTR for mCRC patients. A Japanese randomized controlled trial demonstrated no notable difference in overall survival between the surgery and non-surgery cohorts (median OS of 25.9 vs. 26.4 months, P<0.05), suggesting that PTR might not improve survival in CRC patients (19). However, Lam-Boer et al. (20) and Doah et al. (21)used PSM to reduce selection bias and reported benefits of PTR for advanced CRC. Furthermore, studies indicate that 7-22% of patients without an initial PTR require emergency surgeries (22–24). Wang et al. observed that PTR improved quality-of-life as well as reduced the risk of severe problems including bleeding and perforation (25).

Consistent with previous reports, our study demonstrated that PTR was associated with improved survival in patients with stage IV mCRC, as indicated by the Cox regression analysis (20, 26, 27). However, after applying the competing risk model, PTR did not significantly improve CSS when non-cancer-related death was considered as a competing event. Notably, some patients who underwent surgery did not reach a median CSS of 12 months, suggesting that surgical intervention may not benefit all individuals. These findings highlight the limitations of current surgical recommendations and underscore the need for more selective patient stratification. For patients unlikely to benefit from PTR, non-surgical management or palliative care strategies should be considered as alternative approaches.

Surgical intervention may improve survival due to selection bias; however, this is not the sole factor. Herein, chemotherapy and age were key predictors of surgical benefits. Previous studies have observed that stage IV CRC patients receiving both PTR and chemotherapy had a median OS of 23 months than those receiving only chemotherapy (13 months), sorely surgery (6 months), or without intervention (2 months) (28). In some studies, the age varied from 60 to 75 years, with no significant differences across treatment groups. Our data ( Supplementary Figure S9A ) suggests that PTR with systemic chemotherapy provides greater benefits than PTR alone. Additionally, younger patients benefited more from PTR ( Supplementary Figure S9B ), underscoring the importance of individual characteristics in surgical decision-making. This finding suggests that healthier patients with longer life expectancies are more likely to choose aggressive treatments, including surgery.

The 2018 TNM staging system for mCRC was updated to classify metastases as M1a, M1b, and M1c. Some M1c cases were not distinguishable in the SEER database; therefore, staging was redefined using the 7th edition of the AJCC system to ensure sample representativeness. This staging strategy may affect prognostic interpretation following PTR in mCRC patients. Research indicates that complete removal of the primary neoplasm and metastatic masses through PTR can prolong survival and quality-of-life in M1b cases. However, M1c cases with more widespread disease distribution may experience limited treatment effectiveness and survival (29, 30). Combining M1b and M1c stages may complicate the assessment of the metastatic burden on prognosis. Accurate determination of the M1c stage in clinical practice requires expensive imaging studies. Despite our model not differentiating between M1b and M1c, it still exhibited strong predictive capabilities.

In the surgical benefit prediction model developed in this study, patients with stage III/IV disease and T4 tumors exhibited a lower probability of deriving benefit from PTR, as reflected by lower odds ratios. This finding aligns with the intrinsic relationship between tumor biology and surgical feasibility. T4 tumors are typically characterized by aggressive local invasion and a higher likelihood of involving adjacent organs or structures, which complicates surgical resection and reduces the likelihood of achieving curative outcomes. Similarly, stage III/IV disease indicates more extensive regional or distant metastasis, suggesting a higher degree of systemic tumor progression. In such patients, even if PTR is technically feasible, the overall survival benefit may be limited and could be accompanied by an increased risk of postoperative complications. Therefore, the decision for surgical intervention should not rely solely on anatomical resectability but must also incorporate a thorough evaluation of tumor biology and progression patterns. The findings from our prediction model underscore this principle, highlighting the central role of tumor biology in guiding surgical decision-making.

Although this study is based on the SEER database, which provides a large sample size and high data completeness for model development, several inherent structural limitations may affect the interpretation of our results. First, the SEER database does not systematically record the specific sites of distant metastases (e.g., liver or lung) or the number of metastatic lesions. This limitation prevents us from distinguishing oligometastatic disease from widely metastatic cases. In clinical practice, such distinctions are critical for surgical decision-making, particularly when assessing the suitability of PTR in patients with mCRC. Second, patient performance status data, such as Eastern Cooperative Oncology Group (ECOG) scores or Karnofsky Performance Status (KPS), are not available in SEER. As a result, we could not directly evaluate physical condition or surgical tolerance in our model, which may introduce risk stratification bias. Although we used propensity score matching to balance available covariates such as age and comorbidities, the lack of functional status indicators remains a limitation to the model’s generalizability. In addition, SEER does not capture information on disease recurrence or postoperative complications. This prevents a comprehensive assessment of long-term recurrence risks and non-cancer-related postoperative mortality, potentially leading to an underestimation of long-term outcomes. Although SEER ensures high survival data accuracy through linkage to sources such as the National Death Index, follow-up time may still vary across patients. To address this, we used the reverse Kaplan-Meier method to estimate the median follow-up time, which was 65 months. This indicates an overall adequate follow-up period. However, right-censoring bias may still occur in long-term survivors due to unobserved late events. These limitations—particularly the absence of data on metastasis burden and performance status—may reduce the accuracy and clinical applicability of the predictive model in guiding PTR decisions for stage IV disease. The lack of treatment-specific variables also limits our understanding of how PTR and systemic therapy interact to affect outcomes. Future studies should consider integrating multicenter clinical datasets or real-world electronic health records (EHRs). Such data sources can provide a more comprehensive set of preoperative variables, including metastatic load, functional status, and postoperative complications. This would help improve the accuracy, clinical relevance, and decision-support capability of predictive models for mCRC.

We present an approach to identify suitable candidates for surgical intervention among stage IV CRC patients. Notwithstanding the widespread adoption of machine learning, traditional logistic regression models still demonstrate competitive predictive capabilities. Our findings revealed that PTR can positively impact mCRC patients. However, this is limited to specific patient groups, and the extent of the benefits is influenced by the features of primary tumor. Specifically, younger patients and those with cystic/mucinous and serous tumors, Grade II, T2 stage, M1a stage, undergoing chemotherapy, and surgery at distant sites are likely to benefit more from PTR.