Prognostic nomogram to predict the overall survival of patients with early-onset colorectal cancer: a population-based analysis

The SEER program of the National Cancer Institute (NCI) collects information on cancer incidence and survival from 17 population-based cancer registries and represents about 28% of the US population. In this study, a total of 8886 pathologically proven EO-CRC patients who were diagnosed from January 1, 2010, to December 31, 2017, were retrospectively extracted from the SEER database using the SEER*Stat program (v 8.3.6). Patients with EO-CRC were identified by the ICD-O-3 site code (C18.0, C18.2, C18.3, C18.4, C18.5, C18.6, C18.7, C19.9, C20.9) and the cancer staging scheme (version 0204). The inclusion criteria were as follows: (1) patients were 15–50 years old, (2) CRC was the only primary cancer, (3) complete survival information, and (4) follow-up > 1 month. Patients who had missing or incomplete clinicopathological and socio-economic information (primary site, histological type, grade, tumor size, regional nodes examined, metastatic situation, tumor stage, CEA level, perineural invasion, median household income) were excluded from this study. The detailed patient selection workflow is shown in Fig. 1. Eligible patients were randomly divided into a training cohort and a testing cohort (ratio, 70:30). The training cohort was used to explore the prognostic factors, and to construct a nomogram, the testing cohort was used for internal validation of the nomogram. This study was conducted under the SEER data use agreement, and patient informed consent was not required given the anonymized, de-identified data in the SEER database.

Eighteen factors, including sex, race, primary site, histology, grade, tumor size, number of examined regional nodes, LNR, liver metastasis, lung metastasis, bone metastasis, brain metastasis, TNM stage, T stage, N stage, CEA, perineural invasion, and median household income, were retrieved to predict prognosis of the training cohort. The primary site was defined as right-side (cecum, ascending colon, hepatic flexure of colon, transverse colon), left-side (splenic flexure of colon, descending colon, sigmoid colon, rectosigmoid), and rectum. The LNR was calculated by dividing the metastatic node number by the examined regional node number. Overall survival (OS), the primary endpoint, was defined as the interval from diagnosis until death or last follow-up.

Categorical variables were reported as whole numbers and proportions. The overall survivals of the study cohort were produced using the Kaplan–Meier method, and differences between overall survivals were examined using the log-rank test. The associations between clinicopathological, socio-economic variables and survival were evaluated using Cox proportional hazards regression models. Hazard ratios (HRs) were displayed with 95% CIs. Significant variables in the univariate analysis were subjected to multivariate Cox regression analysis by Backward stepwise selection under the Akaike information criterion (AIC). Variables statistically significant in the multivariate Cox regression analysis were determined as independent prognostic factors to predict the survival outcome. Then, these independent prognostic factors were used to establish a nomogram for predicting the 3- and 5-year OS of patients with EO-CRC. To allot points in the nomogram, the regression coefficients were used to define the linear predictor.

The performance of the nomogram was evaluated by the discriminatory ability and calibration [16]. The discriminatory ability refers to how well the model differentiates patients who will have an event from those who will not have an event. The concordance index (C-index) and the receiver operating characteristic (ROC) curve were applied to evaluate the discriminatory ability of our nomogram. A C-index or the area under the ROC curve (AUC) of 0.5 indicates the nomogram is devoid of discrimination, while a C-index or AUC of 1.0 suggests the perfect separation of patients with different results. A C-index or AUC more than 0.75 reflects useful discrimination [16]. The calibration refers to the consistency between the nomogram-predicted survival and the actual observed survival. Calibration plots were used to evaluate the calibration of our nomogram. In a calibration plot, the actual OS is plotted on the y-axis, and the nomogram-predicted OS is plotted on the x-axis. A perfect prediction would fall on a 45-degree diagonal line. All the statistical analyses were performed using SPSS version 25 and R software version 3.3.0 (Vienna, Austria; www.​r-project.​org). Only a two-tailed P value of < 0.05 was considered statistically significant. This study has been reported in line with the TRIPOD statement [17].

Data on a total of 5585 eligible patients with early-onset colorectal cancer diagnosed from 2010 to 2017 were retrospectively collected from the SEER database. Patients were randomly divided into a training cohort (3910 patients) and a testing cohort (1675 patients). Clinicopathological and socio-economic characteristics of early-onset colorectal cancer patients are listed in Table 1.

Most patients were male (53.64%) and White (71.44%), with a median household income level of 50,000–75,000 dollars (50.67%). The majority of patients had the adenocarcinoma histological type (91.14%), were moderately differentiated (75.34%), examined ≥ 12 regional nodes (84.44%), with LNR ranged from 0 to 0.2 (79.52%), without perineural invasion (76.22%). The left-side colon (45.8%) was the most common primary tumor site, followed by the right-side colon (27.72%), and rectum (26.48%). 51.07% of the patients developed a smaller tumor size (< 5 cm), while 48.93% of patients developed a larger tumor size (≥ 5 cm). Liver metastasis, lung metastasis, bone metastasis, and brain metastasis were observed in 13.68%, 3.33%, 0.5%, and 0.14% of the patients, respectively. Patients with TNM stage I, II, III, and IV tumors accounted for 14.31%, 22.97%, 43.13%, and 19.59% of all cases, respectively. In total, 28.68% of the patients were tested with positive pretreatment CEA, with the remaining patients having negative CEA (39.7%) or unknown CEA (31.62%).

At a median follow-up of 42.0 months (range from 1.0 to 95.0 months), 19.7% (772 of 3910) of the patients had died in the training cohort, and 20.6% (346 of 1675) of the patients had died in the testing cohort. The 3-year and 5-year overall survival were 80.7% (95% CI, 79.3–82.1%), and 72.5% (95% CI, 70.7–74.3%), respectively.

Univariate Cox regression analysis indicated that race, primary site, histology, grade, tumor size, regional nodes examined, LNR, liver metastasis, lung metastasis, bone metastasis, brain metastasis, TNM stage, T stage, N stage, CEA, perineural invasion, and median household income were significantly associated with OS in the training cohort (Table 2).

After controlling confounding factors, the multivariate Cox regression analysis demonstrated that race, primary site, histology, grade, tumor size, regional nodes examined, LNR, liver metastasis, lung metastasis, bone metastasis, TNM stage, T stage, CEA, perineural invasion, and median household income were independent prognostic factors of EO-CRC patients as shown in Fig. 2.

A prognostic nomogram to predict 3- and 5-year OS was established, which contained the independent prognostic factors identified from the multivariable Cox regression analysis (Fig. 3). The corresponding score of each variable can be obtained by projecting to the top “points” axis according to the patient’s actual situation. In the same way, the total points are obtained by adding the corresponding scores of each variable. By projecting the total points to the bottom “3-year overall survival” and “5-year overall survival” axis, the 3- and 5-year OS can be estimated.

For instance, a 45-year-old White patient (3 points) with right-sided colon (16 points) adenocarcinoma (0 points), T4 (35 points), without lung, liver, or bone metastasis (0, 0, and 0 points), TNM stage III (48 points), poor differentiated (25 points), tumor size > 5 cm (8 points), examined 12 regional lymph nodes (0 points), LNR > 0.6 (50 points), CEA positive (24 points), without perineural invasion (0 points), median household income 70,000 dollars (10 points) would have a total of 219 points, which means a predicted 3-year OS of 40.0% and predicted 5-year OS of 20.0%.

Moreover, we compared the prediction ability of the nomogram and the TNM staging system. Compared with the C-index of the constructed nomogram (0.840, 95% CI 0.827–0.850), the C-index of the TNM staging system was lower (0.804, 95% CI 0.788–0.820, P < 0.001). More importantly, the constructed nomogram yielded a larger log-likelihood and a smaller AIC value than the TNM stage (Table 3). All the above results implied the stronger predictive power of the nomogram than the TNM staging system. And the same result was also observed in the testing cohort.