Nomograms predicting cancer-specific survival for stage IV colorectal cancer with synchronous lung metastases

This study aimed to establish a nomogram for the prediction of cancer-specific survival (CSS) of CRC patients with synchronous LM. The final prognostic nomogram based on prognostic factors was evaluated by concordance index (C-index), time-dependent receiver operating characteristic curves, and calibration curves. In the training and validation groups, the C-index for the nomogram was 0.648 and 0.638, and the AUC was 0.793 and 0.785, respectively. The high quality of the calibration curves in the nomogram models for CSS at 1-, 3-, and 5-year was observed. The nomogram model provided a conventional and useful tool to evaluate the 1-, 3-, and 5-year CSS of CRC patients with synchronous LM.

Patient population. The patient data entered in the database were considered to be representative of the overall population. SEER*Stat version 8.3.9 was used to generate a case list. Data from SEER was used to identify patients with CRC diagnosed between 2010 and 2015, and 2875 stage IV CRC patients with synchronous lung metastasis were extracted according to related indications in SEER ("SEER Combined Mets at DX-lung (2010 +): YES", "SEER other cause of death classification: Alive or dead due to cancer", "ICD-O-3: 8140, 8210, 8220, 8261, 8263, 8480, 8481 and 8490"). Then, we excluded the unknown CEA level(n = 845), unknown primary tumor site(n = 115), unknown regional nodes examined(n = 17), unknown race(n = 234), unknown pathologic type(n = 8), unknown AJCC N stage(n = 206). Finally, a total of 1450 patients were enrolled in this study, which was divided into the training cohort and the validation cohort in a 7: 3 ratio by EXCEL using the Rand function. A general description of our study design is presented in Fig. 1. All CRC patients included in this study were definitively diagnosed by pathological examination and LM was diagnosed by imaging examination www.nature.com/scientificreports/ or pathological examination. Several clinical and tumor-related variables were collected to analyze prognosis, including age, AJCC N stage, CEA level, extra-LM (defined as involving bone, brain, and liver), primary tumor site, primary tumor size, and regional nodes examined, sex, year of diagnosis, race, pathological type, and primary tumor resection (primary CRC surgery combined with lung metastasectomy) in the SEER database.
The construction of the CSS nomogram. The CSS was the endpoint in the present study, it was calculated from diagnosis to death of the patient or the date still alive at the last censored follow-up. CSS was assessed using the Kaplan-Meier method, with the log-rank tests used in univariate analysis. The final independent prognostic factors were identified by multivariate analysis using the Cox regression model. The nomogram was developed based on these prognostic factors (P < 0.05) to predict the CSS of CRC patients with synchronous LM.
Evaluation of nomogram performance. The discrimination ability of the nomogram was evaluated using the concordance index (C-index) and AUC value 18 . AUC-index of 0.5 indicated a random chance and 1.0 indicated a perfect ability to correctly discriminate the outcome with the model. AUC values of 0.5-0.7, 0.7-0.85, and 0.85-0.95 were defined as low, middle, and high credibility, respectively 18 . The calibration ability of the nomogram was evaluated with calibration curves for 1-, 3-, and 5-year CSS comparing the predicted survival with the observed survival. The 1-, 3-, and 5-year ROC curves were used to evaluate the predictive accuracy of the nomogram for different periods. Internal validation of the nomogram was achieved with the bootstrap resampling strategy (1000 resamples). External validation was conducted in the validation cohort. Briefly, the validation cohort was individually given a risk score calculated with the nomogram equation.
Statistical analysis. The Rand function of EXCEL was used for the randomization of patients.The Chisquare test was used to compare the differences between the training and validation cohorts for the categorical variables. The R statistical packages "rms", "survival", "Hmisc", "lattice", "Formula", "ggplot2", "foreign", "nomo-gramFormula", "survivalROC", "pROC" and "timeROC" were used to calculate the C-index, plot the calibration and ROC curves, build a nomogram, and draw the Kaplan-Meier curves. SPSS23.0 (SPSS Inc., Chicago, IL) and R (R version 4.1.1, http:// www.r-proje ct. org) were used for statistical analysis, and a P-value < 0.05 was considered statistically significant. This study has been reported in line with the TRIPOD statement 19 .
Ethics approval. The studies involving human participants were reviewed and approved by the Ethics Committee of the National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences. The written consent form was not required for data from the SEER database as all data were de-identified prior to release and did not contain personally identifiable information from patients.

Methods statement.
The authors confirm that all methods were carried out in accordance with relevant guidelines and regulations as publicly available data is used.

Results
Patient characteristics. From 2010 to 2015, 74,926 CRC patients were registered in the SEER database, a total of 1450 CRC patients with synchronous LM were included in our study, and patients were assigned in a nearly7:3 ratio to the training cohort and the validation cohort randomly. A total of 1021 patients with complete information were included in the final analysis for the training cohort. For the validation cohort, a total of 429 patients were included in the final analysis after the application of the same inclusion and exclusion criteria. The clinicopathological characteristics and demographics of the entire cohort (n = 1450), including the training (n = 1021), and validation (n = 429) subsets are described in (Table 1). In the training and the validation cohort, the majority of patients were ≥ 60 years old at diagnosis (62.1 and 64.3%, respectively). Most patients had an adenocarcinoma histological type (86.2 and 86.9%, respectively), and most patients' tumors were located in the descending colon (45.9 and 45.7%, respectively). Extra-LM were identified in 73.8% and 76.2% of the patients in the training and validation cohorts, respectively, and the CEA levels were positive in 84.6% and 88.1% of the patients. Whites accounted for 81.9% and 80.4% of all cases, respectively. Most people were AJCC N0 (39.3 and 41.7%, respectively) and N1 (40.5 and 40.1%, respectively). Across the study population of the two cohorts, only 0.8% and 0.7% of the patients underwent primary surgery, and 62.6% and 65.7% of the patients had no regional nodes examined. The primary tumor size of most patients was < 3 cm (59.6% and 62.0%, respectively). The range of CSS ranged from 0-106 months and 0-107 months in both cohorts, respectively. The mean CSS was 20.35 and 19.84 months. with a median CSS of 14 months and 15 months, respectively. Most of the variables including median survival month (P = 0.739), number of events (P = 0.102), age (P = 0.421), year of diagnosis (P = 0.182), race (P = 0.514), primary tumor site (P = 0.938), pathological type (P = 0.742), primary tumor resection (P = 1.000), Extra-lung metastasis (P = 0.343), CEA(P = 0.084), regional nodes examined (P = 0.723), AJCC N stage (P = 0.582), primary tumor size (P = 0.683) showed no significant differences between the training and the validation cohort, which indicated that patients in the training and validation cohorts had a balanced survival distribution and baseline clinical characteristics.

Survival analysis and independent prognostic factors in CRC patients with synchronous
LM. The survival curves for different variables were generated using the Kaplan-Meier method and were compared using the log-rank test. Seven variables, including age, AJCC N stage, CEA level, extra-lung metastasis, primary tumor site, primary tumor size, and regional nodes examined, were associated with CSS (P < 0.05) ( Table 2). The Kaplan-Meier curves for these factors are shown in Fig. 2 Development and assessment of predictive nomogram. We developed a predictive nomogram containing variables including age,primary site, extra-lung metastasis, CEA, primary tumor size and regional nodes examined, which were demonstrated to be statistically significant in multivariate analysis (Fig. 3). , which showed good discrimination (Fig. 4C). The AUC of the validation cohort was 0.785 (95% CI, 0.708-0.862) which demonstrated the nomogram was well fitted (Fig. 4D).

Risk stratification based on nomogram scores.
To further explore the predictive capacity of the nomogram, the total point of each patient was determined based on the nomogram in both the training and validation cohorts. The median point was 33.56 and 28.52 in the training and validation cohort, respectively. Then, we divided the patients into low-and high-risk groups according to the median points and performed a survival analysis using the Kaplan-Meier method. The mean CSS of the training cohort were 13.97 (95% CI = 12.59-15.36) months and 33.07(95% CI = 30.06-36.08) months in the high-and low-risk groups (P < 0.001) (Fig. 4E), and in the validation cohort, the mean CSS of the high-and low-risk groups were 14.36 (95% CI = 12.27-16.45) months and 29.32 (95% CI = 25.36-33.28) months respectively (P < 0.001) (Fig. 4F). The above results illustrated that patients in the high-risk group tended to have poorer outcomes than those in the low-risk group, and the nomogram had good distinguishing ability and generalizability.
Calibration curve analysis of the nomogram. The bootstrapping method (1000 repetitions) was used and a calibration curve was illustrated in Fig. 5. There were no obvious deviations between the model predicted risk and the actually observed risk curves (Fig. 5A-C), meaning good agreement between observed and predicted 1-, 3-, and 5-year CSS predicted by the nomogram in the training cohort. We further validated the model in the validation cohort using the same method ( Fig. 5D-F) and good calibration was observed.

Discussion
This study focused on the prediction of the 1-, 3-, and 5-year CSS of CRC patients with synchronous LM. From the perspective of predicting indicators, age, CEA levels, extra-LM, primary tumor site, primary tumor size, and regional nodes examined were defined as independent prognostic factors of stage IV CRC patients with synchronous LM. A nomogram based on the aforementioned variables was constructed to forecast the 1-, 3-, and 5-year CSS, and the discrimination ability was estimated by calibration and discrimination in both training and validation cohorts. The calibration curve of 1-, 3-, and 5-year CSS in the training cohort showed favorable agreement between the predicted and actual observed probabilities, similarly, the validation cohort also showed optimal calibration in the 1-, 3-, and 5-year CSS, which supported the repeatability as well as reliability of the constructed model. Then, we stratified CRC patients with LM into high-risk and low-risk groups according to median individual points. This stratification indicated that patients in the high-risk group had poorer CSS, therefore, more intensive follow-up and more comprehensive treatment should be considered. Both the C-index and AUC values revealed the good discriminatory capacity of the nomogram: the C-index of the training and www.nature.com/scientificreports/ validation groups were 0.648 and 0.638, respectively, and the AUC values were 0.793 and 0.785, respectively. Furthermore, the clinical parameters to be input into the nomogram are easy to obtain from the patient's clinical records, making it a simple tool to use. Overall, our nomogram had a good ability to predict CSS and could be applied as a convenient tool to predict 1-, 3-, and 5-year CSS of CRC with synchronous LM. The present study revealed that the prognosis of CRC patients with synchronous LM was better among individuals < 60 years of age, which was consistent with previous studies 20, 21 reporting that age was an independent prognostic factor for CRC patients with synchronous LM and suggested that in an increasingly aging population, the functional status of patients should be taken into account. Furthermore, many studies have reported that the primary tumor site could be a prognostic factor for metastasis CRC [22][23][24] . A proposed explanation is that the colon and rectum present differences in the microbiome composition, chromosomal, and molecular characteristics 25 . Therefore, colorectal tumors with different sites have different capacities for metastasis. This study shows that the primary tumor site could be prognostic in CRC with LM. We showed that different CSS could be predicted for patients with tumors at the right colon, transverse colon, descending colon, and rectum. The primary tumor in the transverse colon for CRC with LM had worse CSS. A high level of CEA was associated with CRC. Preoperative CEA levels were important in determining diagnosis and prognosis and were widely used in clinical practice 20,26,27 , and during the post-resection follow-up period, the CEA level was an important indicator to detect local recurrence and distant metastases after surgery in CRC patients. However, it remains unknown whether CEA was an independent factor in CRC survival with LM. A study evaluating synchronous CRC reported 26 that patients with multiple colorectal tumors were more likely to express high levels of CEA, Unfortunately, it did not explore the relationship between CEA and the prognosis of synchronous colorectal carcinoma. Although, our study demonstrated that elevation of the CEA level could predict the prognosis of patients with CRC with synchronous LM.In this study, tumor size < 3 cm was shown to be associated with poor CSS. This result was www.nature.com/scientificreports/ inconsistent with previous research 28,29 , which showed that tumor size ≥ 5 cm was an independent risk factor for a worse prognosis. However, in the clinic, there may be cases of locally advanced CRC with small primary tumors and small colorectal tumors do not always correlate with early-stage disease. Furthermore, CRC patients with synchronous LM were not the main target investigated in previous studies. Our research indicated that for CRC patients with synchronous LM, smaller tumors tended to have stronger invasiveness and metastatic capacity. Patients with a tumor size < 3 cm may have a poorer prognosis, but further studies are needed. In our study, extra-LM lesions were also identified as a prognostic factor. The presence of extra-LM is associated with a more aggressive primary tumor, on the other hand, extra-LM could influence a patient's organ function, aggravating symptoms, therefore, the extra-LM may represent a negative factor in the CSS of the patients. Cancer cells in regional lymph nodes 30 are often associated with reduced survival. Radical surgery including lymphadenectomy of CRC is a decisive factor for prognosis and is necessary for the therapeutic staging of the patient, the minimum number of 12 examined lymph nodes has been accepted in clinics since 1990 31 . We found that the identification of fewer than 12 regional nodes was associated with a worse CSS, which was consistent with many studies 20,32 . In particular, the role of primary tumor resection in CRC patients with distant metastases remains controversial. A previous study 33 reported better outcomes in patients with CRC LM undergoing lung metastasectomy, and also achieved remarkable improvement in prognosis. But several studies 34,35 have also suggested that surgical resection should not be performed if all known tumors could not be completely removed (R0 resection), because it could not provide survival benefits for CRC patients with metastatic diseases. A recent study 36 also found that prognosis was similar for those who underwent lung metastasectomy and those who did not. Our study showed that primary surgery cannot comprehensively improve CSS for patients with CRC with synchronous LM. The explanation may be due to the advancement of palliative treatment including chemotherapy and molecular targeting treatment (i.e. FOLFOX or FOLFIRI combined with molecularly targeted drugs), which has improved the www.nature.com/scientificreports/ prognosis of CRC patients with LM and will continue to improve. However, this needs to be further studied as only 0.8% of the patients in the study had completed primary surgery, which means that a selection bias cannot be avoided, furthermore, detailed treatment information including chemotherapy was not recorded in the SEER database; thus, further studies should include more cases with lung metastasectomy and treatment information. This large-scale study revealed the clinical characteristics, risk, and prognostic factors for CRC patients with synchronous LM. However, this study still has some limitations. First, this was a retrospective study that included patients from the United States. A future external validation involving patients from other countries is needed to evaluate the generalizability of this nomogram. Moreover, as lung metastasis was the aim of this study, extra-lung metastasis including bone, brain, and liver metastases was not discussed individually while they could influence the evaluation for the CSS. Therefore, Cox regression model might not be the best method to analysis the CSS and a more suitable competing risk model would be applied in the next study, however, the result of this study was still helpful in our future study to further research the CSS influence of other coexisted metastasis for CRC patients with lung metastasis. Also, the 1-year AUC value of the nomogram based on both the training and validation group was 0.68, which suggests the reliability of the 1-year CSS prediction needs to be enhanced. In addition, some patients who received surgery solely targeting their colorectal tumors were not further extracted from this study. Further studies were needed to clarify the necessity of palliative surgery without lung metastasectomy. Finally, this study did not investigate specific treatment options, because detailed treatment information, including radiation therapy, was not recorded in the SEER cohort. In summary, this nomogram was based on the SEER database and made full use of its indicators. The incorporation of other factors, such as additional biomarkers, will improve this model. Despite these limitations, this nomogram remains a good risk model and could be applied to predict the prognosis of stage IV CRC patients with synchronous LM.
In summary, currently, patients with stage I to III tumors who undergo primary surgery are predicted to have a good prognosis; whereas, the prognosis for stage IV CRC patients with LM remains poor. Although these patients only account for a small proportion of CRC patients, greater attention should be placed on their prognosis. In this study, we constructed and validated a predictive nomogram for the CSS of CRC patients with synchronous LM, which could be used to accurately evaluate the 1-, 3-, and 5-year CSS of stage IV CRC patients with synchronous LM and help distinguish high-risk patients who may require more aggressive treatment and follow-up strategies.

Figure 3.
Nomogram for predicting the 1-, 3-, and 5-year CSS. The total score was obtained according to the value of each indicator (age, primary site, extra-lung metastasis, CEA, primary tumor size and regional nodes examined), and the 1-, 3-, and 5-year CSS corresponding to the total score was the predicted rate by nomogram.