A nomogram for predicting cause-specific mortality among patients with cecal carcinoma: a study based on SEER database

SEER * Stat software (version 8.4.1) was used to extract the data in the SEER database (https://​seer.​cancer.​gov/​) about patients who were diagnosed with CC between 2010 and 2015. The diagnosis of CC is based on the third edition of the International Classification of Diseases Oncology (ICD-O-3). We collected demographic information such as age, marital status, gender, and race of patients, as well as clinical information related to tumor pathology, tumor differentiation level, AJCC staging (7th edition), T staging, N staging, surgical procedures, lymph node dissection, distant metastasis site surgery, radiotherapy, chemotherapy, tumor deposits, lymph node metastasis, bone metastasis, brain metastasis, liver metastasis, lung metastasis, tumor size, and patient survival time and status. Data were excluded if (1) patients had a survival time of less than one month were excluded, (2) patients were younger than 18 or older than 100 years old and information patients were involved in missing information; (3) data involving unknown causes of death (COD). As for continuous variables, patients were split into three groups by age: < 40 years old, 40 to 60 years old, and > 60 years old. They were divided into three groups by tumor size: < 30 mm, 30 to 50 mm, and > 50 mm. In the present study, CC-specific death and death from factors except CC compete with each other to deliver the event of interest.

The demographic and clinical characteristics of patients in the training and validation dataset were analyzed. The Chi-square test was utilized for describing distributional differences between two datasets. Multiple variables were screened by univariate and multivariate analysis using the Fine and Gray regression model. Variables with P < 0.05 in the univariate analysis were then included in multivariate analysis, in which factors with P < 0.05 were then used to construct the final competing risk model and nomogram. We calculated the concordance index (C-index) of the final model to evaluate its performance. Calibration plots and receiver operating characteristic (ROC) curves were used to compare predicted and observed probabilities to analyze the performance of the model. We performed statistical analyses using software R version 4.1.2 (https://​www.​r-project.​org/​). The tableone package (version 0.13.2) was employed for data description and riskRegression (2021.10.10) for conducting Fine and Gray regression analysis and establishing the competing risk model. Two-tail P value less than 0.05 was the threshold of statistical significance.

Data on 23,794 patients were extracted. These patients were split into a training set (N = 16,655; 70%) and a validation set (N = 7,139; 30%). Among them, 10,992 patients died during the follow-up period (7,013 dying from CC and 3,979 dying from other causes). Patients dying from non-cancer-related causes accounted for 36.20% of the total deaths. Among all patients, those over 60 years old accounted for 95.85% of the total (n = 22,807); 13,083 patients (54.98%) were female, and 10,711 (45.02%) were male. Most patients (n = 14,173; 59.57%) had tumors with a size of fewer than three centimeters, and 22.98% (n = 5469), 30.25% (n = 30.25), 32.2% (n = 7662), and 14.57% (n = 3466) have stage I, II, III, and IV tumors, respectively. No significant difference was observed regarding the follow-up data in the training and the validation data set (Table 1).

According to the results of univariate analysis, risk factors with P < 0.05 included pathological type of tumor, bone metastasis, brain metastasis, liver metastasis, lung metastasis, tumor grade (degree of differentiation), lymph node dissection, T stage, N stage, M stage, lymph nodes-positive, tumor size, AJCC stage, etc. These variables were then included in the multivariate analysis and the results showed that the following factors affected the survival of patients: age (< 40 years as a reference, 40 to 60: HR = 1.638, 95% CI [0.803, 3.342]; > 60: HR = 2.189, 95% CI [1.089, 4.399]), race (White as a reference, Black: HR = 1.162, 95% CI [1.068, 1.264]), marital (being married as a reference; divorced: HR = 1.014, 95% CI [0.93, 1.106]; single: HR = 1.304, 95% CI [1.201, 1.415]), pathological type of tumor (Behav1 as a reference, Behav2: HR = 0.945, 95% CI [0.841, 1.063], Behav3: HR = 1.009, 95% CI [0.921, 1.106]; Behav4: HR = 0.773, 95% CI [0.652, 0.916]), tumor stage (Stage I as a reference, Grade 1 as a reference; Grade 2: HR = 1.094, 95% CI [0.958, 1.249]; Grade 3: HR = 1.338, 95% CI [1.161, 1.542]; Grade 4: HR = 1.283, 95% CI [1.071, 1.537]), AJJC Stage (Stage I as a reference, Stage II: HR = 1.419, 95% CI [1.161, 1.734]; Stage III: HR = 3.68, 95% CI [2.914, 4.64]; Stage IV: HR = 9.704, 95% CI [7.634, 12.335]), T stage (T1 as a reference, T2: HR = 1.14, 95% CI [0.942, 1.378; T3: HR = 1.759, 95% CI [1.445, 2.141]; T4: HR = 3.029, 95% CI [2.48, 3.7]), type of surgeries(Surg1 as a reference, Surg2: HR = 1.047, 95% CI [0.964, 1.136]; Surg3: HR = 2.252, 95% CI [1.56, 3.251]; Surg4: HR = 0.942, 95% CI [0.789, 1.124]), lymph node dissection (> 4 as a reference;1 ~ 3: HR = 0.927, 95% CI [0.705, 1.219]), chemotherapy(None as a reference, Yes:HR = 1.554, 95% CI [1.444, 1.672]), deposits(None as a reference, Yes: HR = 1.179, 95%CI [1.093, 1.272]), lymph nodes-positive (No as a reference, 1 ~ 3: HR = 1.311, 95% CI [1.044, 1.646]; 4 ~ 6:HR = 1.714, 95% CI [1.276, 2.3]; > 7: HR = 2.234, 95% CI [1.681, 2.971]), liver metastasis (None as a reference; Yes: HR = 1.446, 95% CI [1.302, 1.605), lung metastasis (None as a reference; Yes: HR = 1.206, 95%CI [1.064, 1.367]) (Table 2). These variables were used to construct a competing risk model to estimate the probability of cause-specific mortality in CC patients at 1, 3, and 5 years (Fig. 1).

The predictive accuracy of the nomogram was evaluated using the C-index and the area under the RCO curve (AUC). The calibration performance of the nomogram was assessed using the calibration curve. In the training set, the C-index of the model was 0.848 (se:0.0009). The AUC of 1-, 3-, and 5-year survival was 0.852 (95% CI [0.842, 0.861], 0.861 (95% CI [0.853, 0.868]), and 0.856 (95% CI [0.848, 0.864]), respectively (Figure 2), indicating that the model performed well regarding risk prediction and the calibration curve showed that the predicted probability was in good agreement with the observed one (Figure 3). In the validation set, the C-index of the model was 0.847 (se:0.0015), and the AUC at 1-, 3-, and 5-year survival was 0.841 (95% CI [0.825, 0.856]), 0.862 (95% CI [0.851, 0.874]), and 0.852 (95CI [0.839, 0.864], respectively (Figure 4), and the calibration curve showed that the predicted probability was in good agreement with the observed one (Figure 5). Taken together, the nomogram performed well regarding prediction and calibration.

Traditional approaches to survival analysis such as Cox proportional hazard model are used to estimate the probability of one event over time, in which the occurrence of one type of death will prevent the occurrence of the death from other factors. However, in our study, CC-specific death and death from other factors are competing events. Patients dying from non-cancer-related causes accounted for 36.20% of the total deaths. In this scenario, the use of Cox model to analyze competing event data tends to overestimate the mortality in CC patients. In the present study, the predicted probability of mortality at 1 year, 3 years, and 5 years among CC patients estimated using the classical Cox model was 12.98%, 28.37%, and 35.06%, respectively, and that using the competing risk model was 10.8%, 23.59%, and 29.03%, respectively, demonstrating a relatively significant difference between these two models in risk prediction. Specifically, the mortality rate estimated by traditional COX survival analysis was higher than that estimated by the competitive risk model (Table3).