Development and validation of a prognostic nomogram for early-onset colon cancer

Colorectal cancer (CRC) is one of the common malignant death-caused diseases worldwide [1]. In the United States, CRC patients were newly registered in approximately 130000 cases with over 50000 death reports [1]. In Europe, CRC is both the second common cause of death in the European Union with 215000 cases and second common cancer sites with 447000 cases [2]. In Singapore, CRC ranks top in incidence and second in cause of cancer death [3]. Meanwhile, the incidence and death rates of CRC have been increasing in China [4]. Although the incidence and death rates have been reduced in CRC patients older than 50, the incidence of early-onset CRC (age < 50) increases by 22% and the death rate increases by 13% in the United States during the last decade [1].

Radical surgical intervention remains the primary treatment for CRC [5]. Nevertheless, approximately 25% of CRC patients develop recurrence or distant metastasis [6]. Intriguingly, combinational therapies of chemotherapy and targeted drugs have significantly improved the therapeutic benefits in CRC [7,8]. However, the intrinsic complexity of early-onset CRC remains largely unknown. Generally, early-onset tumors are more likely to be associated with germline genotypes. Hong et al. discovered seven genes (CYR61, UCHL1, FOS, FOSB, EGR1, VIP, and KRT24) as a susceptibility gene set associated with early-onset CRC patients [9]. Ågesen et al. indicated that CLC and IFNAR1 were differentially expressed between young and elderly CRC patients with respect to somatic gene expression, highlighting the genomic complexity associated with age [10]. Nevertheless, the overall prognosis of early-onset CRC remains largely constrained by clinical heterogeneity. Therefore, a refined nomogram system is needed to contribute to the prognosis evaluation of early-onset CRC. In fact, given the genomic-features and clinic-management variances between the colon cancer (COCA) and rectal cancer [11,12], the present study exclusively focused on COCA.

Previously, the tumor-node-metastasis (TNM) cancer staging system of American Joint Committee on Cancer (AJCC) has been periodically updated for effective cancer management [13]. However, increasing studies indicated that other factors, including age, race, and tumor site have also been in association with tumor prognosis in individual case [10,14,15]. Therefore, it is needed to establish a prognostic indicator system specified for early-onset COCA patients.

The nomogram-based statistical method has been widely implemented in prognosis-associated clinical studies with comparable results [16,17]. In fact, nomograms enable specifically individual survival scores by dynamically incorporating clinical variables with technical feasibility and reproducibility. However, nomograms for the prognosis of early-onset COCA have not been fully characterized.

Because of this need, a prognostic nomogram based on the large population of COCA data retrieved from the surveillance, epidemiology, and end results (SEER) program (2004–2015) was developed to predict individualized survival in early-onset COCA.

The clinicopathological data of all COCA patients were retrieved from the SEER program of the United States National Cancer Institute (NCI). SEER program is established to comprehensively collect clinical information on various cancer types for associated incidence, prevalence, and prognostic studies [18,19].

Patients diagnosed with COCA from 2004 to 2015 were retrieved. The histological code [International Classification of Diseases for Oncology, Third edition (ICD-O-3)] and the cancer staging scheme (version 0204) were used to identify all the patients. The inclusion criteria were (1) tumor site was colon excluding rectum; (2) diagnosis had been microscopically confirmed; (3) age <50; (4) complete TNM stage information; (5) COCA was the first and the only cancer primary; (6) surgery had been performed. All the included patients were randomly assigned to a training set and validation set. The present study was performed with the data from SEER program and reference number 14622-Nov2017. No informed consent or institutional review board approval was required in the present study due to the public-available data of SEER.

Clinical variables of COCA patients were extracted, including age, sex, marital status, histological grade, tumor site, TNM stage, tumor size, SEER stage follow-up information, and corresponding death causes. Overall survival (OS) was the primary endpoint, defined as the time period from the diagnosis to the death or last follow-up. Cancer-specific survival (CSS) was the second endpoint, defined as the time period from the diagnosis to the death caused by COCA or censoring. Age and tumor size were transformed into categorical variables.

All the categorical variables were presented with frequencies and proportions, and analyzed by a chi-square test. The Kaplan–Meier method and log-rank test were used to analyze each potential prognostic variable. All significant variables from univariate analysis were subject to a multivariate Cox proportional hazards analysis. The construction of nomogram was based on the multivariate cox regression model by the R statistical package rms (R Foundation for Statistical Computing, Vienna, Austria).

The validation set was used for the validation of the nomogram by the discrimination, calibration, and bootstrap resampling. The concordance index (C-index) was used to measure the difference between the observed and predicted results from the constructed nomogram. Receiver operating characteristics curve (ROC) analysis was performed for sensitivity and specificity. We further compared the nomogram, the TNM stage, and the SEER stage using the C-index. Calibration plot was used to visualize the variance between nomogram-predicted prognosis and actual prognosis. The 45-degree line in a calibration plot was used as a perfect model to compare the actual outcomes. Furthermore, decision curve analysis (DCA) was used to depict the threshold probabilities ranges in comparison with TNM stage and SEER stage. All the statistical analyses were performed using R software version 3.3.0 (Vienna, Austria; www.r-project.org). P-value <0.05 was considered as statistically significant.

A total of 11220 eligible cases were included in the present study with 7856 cases randomly assigned to the training set and 3364 into the validation set (Figure 1). 52.2% of all patients were married and 41.7% were unmarried (also including widowed, single, and divorced). For all the early-onset patients, 7780 were between 40 and 50 years old (69.3%), whereas 2383 were between 30 and 40 years old (21.2%). 1057 were younger than 30 years old (9.4%). The majority of race was white (73.1%). The most common tumor site for COCA in the present study was the sigmoid colon (35.3%), followed by appendix (14.7%) and cecum (13.8%) and ascending colon (12.4%). For tumor size, 2–5 cm was the most common type (38.7%), followed by 5–10 cm (30.7%). 52.5% of all the patients were N0 whereas 80.1% were M0 in AJCC stage system. 42.5% of all the patients were regional in SEER stage (Table 1).

Marital status, age, race, grade, site, TNM stage, tumor size, and SEER stage were significantly associated with OS by univariate analysis in the training set (Table 2). Further multivariate analysis indicated that marital status, race, grade, TNM stage, tumor size, and SEER stage were significantly associated with OS. Therefore, a nomogram of 3- and 5-year OS was established with the independent variables (Figure 2A). In addition, the prognostic values and clinicopathological characterization of patients with different marital status were displayed (Supplementary Figure S1 and Table S1). Moreover, each variable was also examined for CSS and therefore used to build a CSS nomogram as well (Table 3 and Figure 2B).

The nomograms were both internally and externally validated. The internal validation was performed via the training cohort with the C-index as 0.835 (95%CI, 0.823–0.847) in OS and 0.851 (95%CI, 0.840–0.862) in CSS, respectively (Table 4). The external validation was performed via the validation cohort with the C-index as 0.848 (95%CI, 0.831–0.865) in OS and 0.863 (95%CI, 0.847–0.879) in CSS, respectively. Calibration plots of the validations of OS and CSS nomograms indicated highly correlations between the predicted and observed results (Figures 3 and 4).

The area under ROC curve (AUC) was analyzed for both the training and validation set, respectively (Figure 5A–H). Furthermore, the comparisons between the nomograms and TNM stage and SEER stage were performed. The OS and CSS nomograms showed comparable results to TNM stage and SEER stage in both training and validation sets (Table 4). Moreover, the DCA results of nomograms also strengthened the clinical applicability of nomograms from OS and CSS with superiority over TNM stage and SEER stage (Figure 6).

The present study established OS and CSS prognostic nomograms for COCA patients derived from the SEER program with favorable discrimination and calibration and comparable predictive capability. In fact, the nomogram highlighted the clinical significance of marital status, race, grade, TNM stage, tumor size, and SEER stage in early-onset COCA patients.

The role of marital status in cancer had been previously investigated in SEER program [20]. Married patients were featured by less metastatic diseases, reduced cancer-specific deaths, and more likely to receive definitive therapy [20]. For colon cancer, married patients were more likely to be diagnosed at earlier stage and to receive surgical treatment [21]. Moreover, married patients had significantly lower risk in CSS [21]. Our study also indicated the consistent results.

Noteworthy, race, sex, and tumor site were not an independent prognostic variable for early-onset COCA patients in the present study. In fact, race and sex has been viewed as one of the essence variables for cancer treatment [22–24]. Tumor site in CRC had been intensively studied with large population [25,26]. Right-sided colon cancers exhibited increased mortality risk compared with left-sided colon cancers. However, no specific tumor site has been under investigation. Intriguingly, although age was associated with significant prognosis in univariate analysis, it remained insignificant in multivariate analysis, indicating that the subtle stratification between age <30, 30–40, and 40–50 required further investigation.

Of note, tumor size was an independent prognostic variable in the nomogram in the present study. In fact, only the tumor >10 cm exhibited significant higher prognostic risk than tumor <2cm whereas the rest stratification remained insignificant. It was possible that the tumor size could be one of the insightful variables for the prognostic risk prediction.

Up to now, nomogram statistical tool provided reasonable, reproducible, and rigid algorithms for individualized prognostic assessment. It has been implemented as prognostic indicators for several cancer types including pancreas, gastrointestinal stromal tumor (GIST), and gastric cancer [16,27,28]. In pancreatic cancer, a nomogram constructed by CSS data of 53028 patients diagnosed as pancreatic cancer from the SEER program and eight independent clinical variables (C-index = 0.734) [16]. For a total of 5622 GIST patients, similar nomograms were also established by seven independent clinical variables in both CSS and OS data. Noteworthy, these nomogram exhibited better discrimination power than TNM stage and SEER stage systems [27]. Furthermore, Liu et al. identified a prognostic nomogram for disease specific survival (DSS) of gastric cancer patients using the SEER program and external validation set [28]. Collectively, incorporation of numerous prognostic-associated clinical variables into nomogram algorithms could display comparable staging system and more disease-specific features.

The limitations of the present study were the lack of external clinical data from independent sources and the vacant data on chemotherapy and radiotherapy, as well as the genomic phenotypes in SEER program. Moreover, given the potential confounding factors within the surgical patterns, surgical styles, postoperative complications, and some surgical items remaining contradictory, the COCA patients without surgery or the complete TNM stage data were excluded from the present study.

The nomograms established in the present study provided an alternative tool to both OS and CSS prognostic prediction compared with TNM and SEER stages.

The datasets supporting the conclusion of this article are included within the article.

This article does not contain any studies with human participants or animals performed by any of the authors.

The authors declare that there are no sources of funding to be acknowledged.

Y.Z. and C.Y. carried out data analysis. Y.Z. and C.Y. drafted the manuscript, and participated in study design and data collection. All authors read and approved the final manuscript.

The authors declare that there are no competing interests associated with the manuscript.

AUC

area under receiver operating characteristics curve

AJCC

American Joint Committee on Cancer

95%CI

95% confidence intervals

C-index

concordance index

COCA

colon cancer

CRC

colorectal cancer

CSS

cancer-specific survival

DCA

decision curve analysis

GIST

gastrointestinal stromal tumor

HR

hazard ratio

ICD-O-3

International Classification of Diseases for Oncology, Third edition

NCI

National Cancer Institute

OS

overall survival

ROC

receiver operating characteristics curve

SEER

surveillance, epidemiology, and end results

TNM

tumor-node-metastasis stage