Nomograms for predicting overall survival and cancer-specific survival in young patients with pancreatic cancer in the US based on the SEER database

Selection of patients

PC patients younger than 50 years old who were diagnosed from January 1, 2004 to December 31, 2015 were selected from the Incidence-SEER 18 Registries Custom Data (with additional treatment fields), released April 2019, based on the November 2018 submission. The patient meted the following conditions: (1) site recode ICD-0-3/WHO 2008: “pancreas”; (2) positive histology; (3) active follow-up; (4) one primary only. The exclusion criteria were as follows: (1) age, race, surgery unknown; (2) AJCC stage NOS, N/A or unknown; (3) age at diagnosed ≥50 years old; (4) survival time unknown and less than 1 month. The detailed flow chart was shown in Fig. 1. For this study, we signed the SEER research data agreement to access SEER information with the username10067-Nov2018. The SEER database is publicly available and the data for all patients are de-identified, so the approval and informed consent of the institutional review committee were not required in the current study.

Variable classification

Clinical variables including race, gender, primary site, grade, pathological types, AJCC stage, status of surgery, survival time, status of survival and cause for death were extracted from the SEER database. The degree of differentiation of tumors was divided into three groups: grade I (well differentiated) and grade II (moderately differentiated) were high differentiated, grade III (poorly differentiated) and grade IV (undifferentiated) were low differentiated, and unknown. The pathological types were divided into ductal adenocarcinoma and non-ductal adenocarcinoma. The International Classification of Diseases for Oncology, Third Edition (ICD-0-3) codes of ductal adenocarcinoma were 8500 and 8140 (Shi et al., 2019). The staging of cancer is based on the 6th edition of AJCC stage, which adapted to patients in the SEER database with a diagnosis time of 2004–2015.

Statistical analysis

In this study, the whole cohort was randomly divided into two groups, 2,904 (70%) were training set and 1,242 (30%) were validation set. Chi-square test was used to compare the clinicopathological characteristics between the training set and the validation set.

OS referred to the duration from diagnosis to any original death or last follow-up. variables associated with OS in univariable analysis (p < 0.05) were included into multivariable analysis. The method of forward stepwise selection in a multivariable regression model was applied to the training cohort to select variables. The independent prognostic factors on multivariable analysis were used to construct nomogram for OS.

Considering that death from other causes is a competitive event of pancreatic cancer death, we constructed a competing risk nomogram to predict CCS. CSS referred to the duration from diagnosis to death from PC, patients who were alive at the point of last follow-up were considered as censored events. Variables related with CSS (p < 0.05) in the univariable analysis or with important clinical value were included into a multivariable analysis based on proportional subdistribution hazard models and those independent variables were selected to build a nomogram for CSS.

Discrimination of the nomograms was measured through the concordance index (C-index) with its respective 95% confidence interval (CI), which quantifies the level of concordance between probabilities of prediction and the actual chance of having the event of interest (Iasonos et al., 2008). The larger the C-index is, the more accurate the nomogram is to predict the prognosis (Huitzil-Melendez et al., 2010). 10-fold cross-validation was applied to verify the stability of the model and calculate the average value of C index in the training group. In order to reduce the overfit bias, calibration was evaluated by comparing the actual probabilities and the plot of the nomogram using 1,000 bootstrap samples. In the calibration curve, the vertical axis is the actual value and the horizontal axis is the predicted value. If the actual/predicted value passes through the origin along the 45° line, it shows that the nomogram has been well calibrated (Iasonos et al., 2008).

For all analyses, p-value <0.05 was considered statistically significant. All date was obtained through SEER*Stat software version 8.3.5. Fine and Gray competing risk Statistical analyses were performed using R work, and the others performed by SPSS (IBM, NY). The nomograms and calibration curves were draw using R version 3.6.0 (http://www.r-project.org)

Patients characteristic

Finally, there were a total of 4,146 cases of young patients with PC diagnosed in the SEER database between January 1, 2004 and December 31, 2015. In the entire cohort, in terms of demography, mainly whites (74.1%) and males (53.7%); In terms of tumor characteristics, pancreatic head cancer (46.1%) was the most common, more highly differentiated tumors (30.7%), mainly ductal adenocarcinoma (68.0%), in addition, more advanced tumors in diagnosis (55.7%); In terms of treatment, the proportion of patients undergoing surgery was low (37.5%). Detailed patient clinical characteristics were summarized in Table 1.

Construction and validation of nomogram for predicting OS of young patients with PC

The results of univariable and multivariable Cox regression models for OS were shown in Table 2. Univariable analysis showed that race, gender, primary site, grade, pathological types, AJCC stage and surgery were correlated with OS (p < 0.05). Multivariable analysis suggested that gender, grade, pathological types, AJCC stage and surgery were independent risk factors for OS. In details, female (HR:0.91, 95% CI [0.83–0.98]; p = 0.019), non-ductal adenocarcinoma (HR:0.40,95% CI [0.36–0.45]; p < 0.001) and receiving surgery (HR:0.43,95% CI [0.38–0.50]; p < 0.001) had a better OS. Black (HR:1.15, 95% CI [1.03–1.28]; p = 0.016), low differentiated tumor (HR:1.59, 95% CI [1.40–1.80]; p < 0.001) and advanced stage (HR:4.63, 95% CI [3.53–6.07]; p < 0.001) had a worse OS. The Nomogram was constructed of the above six variables in training set (Fig. 2A). It can be seen from the nomogram that AJCC stage, which range of risk score is from 0 to 100, had the greatest contribution on OS (including 1-, 3- and 5-years), followed by pathological subtypes, surgery, grade, race and gender. The detailed steps for the application of the nomogram are as follows: Draw a vertical line to the horizontal axis marked ”points” at the top of the nomogram according to the classification (e.g., gender is divided into male and female) of each prognostic variable (gender, grade, pathological type, AJCC stage, and surgery). At the position where the vertical line passes through the ”Points” axis, each prognostic variable can obtain a score. Add the scores of the five variables to get the total score, find the position of the total score on the horizontal axis marked as ”total points”, and draw a vertical line from the total score position marked on the horizontal axis of ”Total Points” to the 1-, 3- and 5-years OS axis. Where the vertical line intersects the 1-year OS axis is the % probability of the 1-year overall survival.

Compared with C-index based on AJCC stage in training set (0.677, 95% CI [0.666–0.688]) and validation set (0.672, 95% CI [0.656–0.688]), The C-index of our model in the training set and validation set was 0.733 (average = 0.731, 95% CI [0.724–0.738]) and 0.742 (95% CI [0.725–0.759]) respectively, showing a better degree of discrimination. The calibration curve of training set and validation set showed good consistency between prediction and observation in the probability of 1 -, 3 -, and 5-year OS, respectively (Fig. 3).

Construction and validation of nomogram for predicting CSS of young patients with PC

The results of univariable and multivariable competing risks models for CSS were shown in Table 3. Fine and Gray analysis showed that primary site, grade, pathological types, AJCC stage and surgery were correlated with CSS. Multivariable analysis suggested that primary site, pathological types, AJCC stage and surgery were independent risk factors for CSS. In details, non-ductal adenocarcinoma (SHR:0.57, 95% CI [0.50–0.65]; p < 0.001), receiving surgery (SHR:0.77, 95% CI [0.68–0.87]; p < 0.001) and tumors in others site (SHR:0.87, 95% CI [0.77–0.99]; p = 0.037) had a better CSS. Advanced stage (SHR:1.53, 95% CI [1.22–1.93]; p < 0.001) had a worse CSS. The Nomogram was constructed of the above four variables in training set (Fig. 2.B). It also can be seen from the nomogram that AJCC stage had the greatest effect on CSS (including 1- ,3- and 5-years), followed by pathological subtypes, surgery and primary site. Finally, similar to the above, we could also predict 1 -, 3 -, and 5-year CSS in patients with PC.

Compared with C-index based on AJCC stage in training set (0.706, 95% CI [0.704–0.707]) and validation set (0.692, 95% CI [0.695–0.689]), The C-index of our model in the training set and validation set was 0.792 (average = 0.765, 95% CI [0.742–0.788]) and 0.776 (95% CI [0.773–0.779]), respectively, showing a better degree of discrimination. The calibration curve of training set and validation set also showed good consistency between prediction and observation in the probability of 1 -, 3 -, and 5-year CSS, respectively (Fig. 4).