Colorectal cancer (CRC) is a major cause of morbidity and mortality worldwide, its importance is expected to continue increasing over time (1, 2). In recent years, increased awareness and the introduction of population-based surveillance and screening programs have led to achieving higher rates of precancerous dysplastic lesions or early CRC detection (3, 4).

Early CRC is a tumor that is confined to the mucosa and/or submucosa regardless of the presence of regional lymph node metastasis (LNM). In certain cases of early CRC, endoscopic resection is a less invasive and cost-effective treatment compared to surgery (5–7). However, the CRC patients with LNM or distant metastasis cannot be adequately cured by local endoscopic treatment alone, and therefore subsequently require surgical resection for achieving a curative treatment.

LNM is found in approximately 6–16% of the patients with submucosal invasive CRC (8–10), however, the number might be underestimated, as clinicians make important treatment decisions based on limited examinations, such as computed tomography (CT) and ultrasonography.

Thus, an accurate and fast assessment of locoregional and/or distant metastases in patients with early CRC is essential to determine whether these patients should undergo additional surgical resections or be needed surveillance regularly. Currently, no universally accepted indications and criteria exist for additional surgical resection after endoscopic resection, even though a fast and accurate assessment of the risk of locoregional LNM after local endoscopic treatment of patients with early CRC is necessary.

Therefore, the aim of present study was to develop a novel prediction model for LNM by using simple histopathological and clinical parameters with high reliability, that can be used to improve patient risk stratification in early CRC.

In this study, we established a novel predictive model by combining eight clinicopathologic parameters to predict LNM in early CRC using seven ML models. To the best of our knowledge, this is the first large-scale study to develop a predictive model for LNM by combining easily available simple clinical and pathological data in patients with early CRC. Clinicians are often confronted with the difficulty of selecting candidates who will benefit from surgery after local endoscopic resection.

Currently, in clinical practice, risk stratification in these patients is usually performed by histopathologists carefully analyzing the specimen to determine the risk of LNM, caused by the limited capacity of CT to accurately identify LNM (23).

In previous studies, the pathological factors that showed the strongest independent predictive value for LNM in early CRC are tumor type, poor histological differentiation, and the depth of submucosal invasion (24–27). However, the high interobserver variability in the pathological assessment limits their clinical usefulness and should therefore be interpreted with caution as a univariate marker when deciding whether to proceed with surgery (28, 29). Therefore, the multivariable risk model combining the histopathological data with clinical data can reduce the inaccuracies associated with relying on individual subjective markers and to better define the optimal treatment strategy for early CRC.

With the recent rapid development of computer-aided technology, the application of ML model in cancer diagnosis has an important role; it is being widely used in the medical field with growing trend toward predictive medicine (30–32). We hereby developed an ML model by using the simple clinicopathological parameters in large data, which provided high predictive ability of LNM for patients with early CRC.

To date, a few ML models for prediction of metastasis in patients with early CRC have been developed and evaluated for prognosis and prediction in patients with early CRC (33–36). Ichimasa et al. developed the SVM model with 45 clinicopathologic factors for prediction of LNM in patients with early CRC. They reported that artificial intelligence significantly reduces unnecessary extra surgery after endoscopic resection of T1 CRC without LNM positive in comparison to the current guidelines (33). Another Japanese study showed a deep learning model for predicting LNM from pathology images with cytokeratin immunohistochemistry in early CRC (34). However, these studies were retrospective in nature with single center or small numbers of patients. Due to the low rate of metastasis in early CRC, only a limited number of events exist, leading to limited data. Furthermore, inadequate data could not provide sufficient satisfactory performance under ML algorithms and may have led to lower predictive performance ranging from 0.821 to 0.913, which is less than the result from our RF model. A recent Chinese study also presented a predicting model for LNM that incorporates both the radiomics signature, which combine multiple individual CT imaging features, and several clinical factors using the multivariable logistic regression analysis (35). Although this might be an interesting attempt, the model validity is not guaranteed considering the heterogeneity in the quality of CT image between facilities and its accuracy of approximately 78%, which is lower than the performance of the predictive model we constructed. Lastly, Kudo et al. also employed deep-learning-based modeling to predict LNM in T1 CRC (36). However, they only used NN model for nonlinear dynamic system with smaller sample size than our study and assessed LNM using only CT imaging in the cases treated by endoscopic resection, because pathologic confirmation was not available.

Meanwhile, the reason why the RF model outperforms the other ML algorithms is not easily explained. It might be attributed to that the RF models generally demonstrate the most substantial improvement over linear methods and, might be outperform kernel-based model and neural network model in many categorical variables and some outliers from the nature of large retrospective cohort data. However, to build robust prognostic models for LNM in early CRC, other variables, such as gene expression and histologic image data beyond clinical-pathological variables, should be needed.

In our study, we investigated the variable importance of the predictive models developed, as it could be useful for decision-making by clinicians. Our findings indicated that tumor size was the most important factor for predicting the presence of LNM in early CRC. The prognostic value of tumor size in CRC has long been studied, but no consensus has been reached. Zhang et al. and Kornprat et al. demonstrated a significant association between tumor size and metastasis in CRC (37, 38), whereas Miller et al. indicated no prognostic significance of tumor size in CRC (39). Furthermore, its potential prognostic role in patients with early CRC has not been well investigated. This is the first largest study to identify the prognostic value of tumor size for early CRC and provide statistical evidence for further prospective study. Despite the aforementioned, the current study has several limitations. First, since the SEER database is a nationwide program, several diagnostic criteria, such as histological grades and verification of tumor locations might be subjective, which could cause potential systematic bias. Second, detailed histopathological data, such as lymphovascular invasion, tumor budding, and precise depth of tumor invasion that have been associated with metastasis are insufficient. These data require further assessment to improve the performance of our ML algorithms. Third, our study comprised predominately of white patients; thus, the findings may not be generalized to other racial populations. Finally, the data have a class imbalance problem between the patients with and those without LNM, which means that the rate of LNM is low in early CRC. Therefore, during the tuning process, the parameters had to be further optimizing to avoid overfitting. To further improve the accuracy of the established model, it is necessary to collect more clinical data and further optimizing the parameters are necessary in subsequent studies.

In conclusion, we established and compared seven models to predict metastasis in early CRC by using easily available clinical and histopathological features in real practice. The RF model, a simplified reproducible predictive model, showed the highest predictive power compared with the other models. Tumor size the most important predictor of LNM in early CRC. Therefore, the patients with tumor larger than 3 cm, who were identified as high-risk through the model, may requires careful attention to selection and additional surgical treatment in early CRC. However, because of the limitations inherent in studies based on observational data, these findings should be confirmed in prospective clinical trials.

Publicly available datasets were analyzed in this study. This data can be found here: https://seer.cancer.gov.

MK designed the study. JA, JC and HS analyzed and interpreted the data and wrote the manuscript. JJ and JY supervised the project and revised the paper. All authors contributed to the article and approved the submitted version.

This research was supported by the Basic Science Research Program of the National Research Foundation of Korea (NRF), which is funded by the Korean Ministry of Science, ICT and Future Planning [grant number NRF- 2019R1C1C1003524].

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.