Development and Validation of a Prognostic Nomogram to Predict Overall Survival of Elderly Patients with Gastric Cancer


 Background

Gastric cancer (GC) is one of the most common malignant tumors of digestive tract origin in China. The proportion of elderly patients with gastric cancer (GC) gradually increases as the population ages. We aimed to develop a prognostic nomogram for prediction of elderly (≥ 75 years old) GC patients in overall survival (OS).
Patients and Methods

Patients with GC from 2005 to 2014 were selected from the Surveillance, Epidemiology, and End Result (SEER) database and randomly assigned to development and validation sets. The variables for establishing nomogram were confirmed by univariate and multivariate Cox proportional hazard analysis based on the development set. The predictive accuracy and discriminative ability of the model was evaluated using the receiver operating characteristic (ROC) curve, the concordance index (C-index) and calibration curves, while its clinical utility was assessed using decision curve analysis (DCA) and Kaplan-Meier curve.
Results

A total of 1445 patients were included in this study. The nomogram was developed including histologic grade, AJCC stage T, N, M and surgery according to the univariate and multivariate cox regression analysis, the area under the time-dependent receiver operating characteristic curve (AUC) and Occam’s Law of Razor. The C-index of the nomogram was higher than the TNM system in the training cohort (0.710 vs 0.652, p < 0.001), which was also confirmed in the validation cohort (0.701 vs 0.643, p < 0.001); and high AUCs were noted in both development and validation sets. The nomogram showed good discrimination and calibration in both development and validation sets. The DCA curves showed that the nomogram had better clinical utility compared to the AJCC stage model. In addition, participants could be divided into three disparate risk groups (low, moderate, high) by the nomogram.
Conclusion

This study established a prognostic nomogram that improved the performance of the AJCC staging system with incorporation of risk factors to better predict the short-term survival in elderly GC patients.

Due to its peak prevalence in the 7th decade, the proportion of aged patients with GC will increase with the progression of aging populations [2,3]. Current treatment guidelines for GC are primarily based on evidence from randomized clinical trials (RCT) conducted in young patients, rather than in elderly patients, despite that these population have been corroborated to have poor prognosis [2].
The management of gastric cancer in elderly patients is undoubtedly very challenging for surgical oncologists. A wide range of progressive physiological changes appear with advancing age, including deterioration of physical function and body composition [4]; which lead to advanced aged patients commonly exhibit longer postoperative hospital stays, higher rates of wound infection, higher all-cause mortality, as well as higher all-cause mortality [5].
The American Joint Committee on Cancer (AJCC) built a staging standard for classi cation of GC by the depth of tumor invasion (T), number of metastatic lymph nodes (N), and presence or absence of distant metastases (M) [6]. This staging system has been generally applied to predict the clinical status and prognosis of patients with GC. However, it is common for patients of the same stage to present with very different survival outcomes. This may be caused by other clinical factors such as sex [7], tumor size [8], or in ammation-related factors [9]. Nomogram has been widely used in clinical tumor research to quantify risk through screening signi cant prognostic factors and models have also been set up based on these to calculate individual risk score to estimate the probability of a speci c result for an individual.
Compared with the traditional AJCC staging system, nomograms have better prediction accuracy and have been validated to predict the prognosis of young GC patients [10]. However, thus far there is no nomogram explicitly to predict the prognosis of elderly patients.
The main objective of this was to develop and validate a nomogram that can improve the performance of the AJCC staging system with incorporation of clinicopathological risk factors to predict the prognosis of elderly GC patients.

Study Population
The research data is extracted from the Surveillance, Epidemiology, and End Results (SEER) using SEER*Stat software (version 8.3.5, National Center Institute). We have obtained permission to visit them for research only (Reference number: 18942-Nov2018). The inclusion criteria were as follow: over 75 years old at time of diagnosis; known diagnosis time between 2005 and 2014; diagnosis staged with American Joint Committee on Cancer (AJCC) 7th edition; classi cation of tumor sites into upper, middle and lower; histopathologic types were limited to adenocarcinoma, mucinous adenocarcinoma, and signet ring cell carcinoma based on the International Classi cation of Oncology (Third Edition (ICD-O-3)) system. The exclusion criteria were incomplete clinical information, incomplete stage information, tumor site not included in classi cation, other histopathologic type and incomplete follow-up information ( Figure.1).
After initial screening, a total of 1445 patients were included. American Indian/Alaskan Native and Asian/Paci c Islanders were recorded as 'others' under race. Tumor size was translated into categorical variables based on traditionally accepted cut-off values. For further analysis, all included patients were randomly divided into development and veri cation sets (split ratio: 7:3).

Demographic Comparison between Development and Validation Set
Categorical variables expressed as percentages were analyzed by the chi-squared test. Univariate cox analysis was conducted to analyze the association between each variable with OS.

Development of a Prognostic Nomogram
Multivariate cox proportional hazard analysis was applied to evaluate the independent factor selected by univariate cox analysis. This is to determine which variables to be used to establish the nomogram. We used the area under the time-dependent receiver operating characteristic curve (AUC) to comprehensively assess the combined performance of ltered variables. Furthermore, we eliminated unnecessary variables in order to make the model more concise based on Occam's Law of Razor [11].
The prognostic nomogram was developed based on the ve most signi cant factors to predict the shortterm survival rate of elderly GC patients.

Validation of the Prognostic Nomogram
Nomogram performance was assessed in the development and veri cation sets by discrimination and calibration [12]. The C-index and the AUC-ROC were used to assess the discriminative ability of the nomogram. The calibration curves were used for comparison between the actual outcomes and the nomogram-predicted probabilities. Moreover, nomogram was subjected to bootstrapping with 1,000 resamples to evaluate prediction accuracy.

DCA Curve Analysis
Decision curve analyses (DCA) were conducted to the development set to determine the clinical bene ts and usefulness of the prognostic nomogram by comparing the net bene ts at various threshold probabilities with the traditional AJCC stage model.

Risk Strati cation
Each included variable was assigned a corresponding risk score in the nomogram. By computing the total risk score through the nomogram and making a vertical line, we can know the short-term survival rate of each individual. In terms of the best cut-off value generated by the X-tile analysis, the total risk score was assigned into high, medium and low risk groups. The Kaplan-Meier method analysis further demonstrated the performance of nomogram in risk strati cation

Result
Characteristics of Development and Validation sets A total of 1,445 elderly GC patients were recruited in the study, of which 1013 were randomly assigned to development set and 432 were assigned to veri cation set ( Figure. Table 1.  Eventually, a nomogram incorporated the above independent predictors from the development set was developed for the prediction of survival of elderly GC patients at 3 and 5 years (Figure.3). The total risk scores were classi ed into three risk groups by the X-tile analysis: low-risk (< 68), moderate-risk (68 ~ 105), high-risk (> 105) ( Figure.6 ).

Validation of the Prognostic Nomogram
To con rm the validity of the nomogram in predicting prognosis of elderly GC patients, we conducted exhaustive validations of the nomogram between the development and validation sets. We veri ed from the following three heading: Firstly, the C-index for the nomogram of development set was 0.710 (95% CI, 0.689 to 0.732) and of the validation set was 0.701(95% CI, 0.668 to 0.734) which performed better than those from the AJCC stage model: development set was 0.652 (95% CI, 0.628 to 0.676), veri cation set was 0.643 (95% CI, 0.608 to 0.678), there was a signi cant difference in two sets (p < 0.01); Secondly, high AUCs were noted in both development and validation sets ( Figure.2); Thirdly, the nomogram calibration plot demonstrated high consistency between nomogram-predicted probabilities and actual outcomes for 3-and 5-year survival both in the development and validation sets (Fig. 4).

Clinical Bene ts
The results of DCA for the prognostic nomogram and that for the traditional AJCC stage model were presented in Figure.5. The nomogram' s 3-and 5-year DCA curves showed larger net bene ts at all threshold probabilities in the development set compared to the AJCC stage model.

Performance in Risk Strati cation of Patients
The total risk scores which were computed by the nomogram in the development set were classi ed into three risk groups based on the cut-off value reckoned by the X-tile analysis: low-risk (< 68), moderate-risk (68 ~ 105), high-risk (> 105), in which the distribution of patients were also shown in Figure.6. The validation set was also divided into three risk groups through the same classi cation criteria. The Kaplan-Meier curves explicitly manifested that the low-risk group has a better prognosis than the moderate-and high-risk group in both the development and the validation sets.
The morbidity of GC was shown to increase with age and older age in GC patients often means a poor prognosis [13]. Accurately predicting the survival of elderly GC patients is still imperative and a grand challenge for clinicians because it could effectively improve elderly patients' prognosis. Chinese average life span in 2019 was 76.1 years showed in the data released by the World Health OrganizationIn (WHO); therefore, we de ne patients older than 75 years as elderly patients which is more representative to predict their short-term survival. In the study, we developed and validated a prognostic nomogram to predict survival of elderly GC patients based on the SEER database. The nomogram involving ve variables of histologic grade, AJCC stage T, N, M and surgery showed excellent discrimination and calibration in both internal and external validation, which manifested the nomogram could assist with the clinical prognostic prediction.
In our study, we discovered that the proportion of elderly GC patients with poor histologic grade, T4, N3, M1 were higher than the reported proportion of early cases [14], and elderly GC patients had a lower OS than early-onset gastric cancer (EOGC) [15]. In addition, we con rmed ve variables associated with the  [17]. Most studies indicated that elderly GC patients had worse survival outcomes than younger GC patients. These studies were consistent with our ndings.
Taking the in uence of the risk factors mentioned above into account, the traditional AJCC staging system may not predict the survival of elderly GC patients well. Due to this reason, we developed a prognostic nomogram incorporating clinicopathological risk factors to predict survival of the elderly GC patients. Notably, Yu et al. [18] published a nomogram to evaluate the prognosis of EOGC, by including tumor size and tumor site as signi cant risk factors of survival of EOGC; however, in our study, they did not show su cient predictive performance in the multivariate Cox model related to the prognosis of elderly GC patients. This may be due to the collinearity with other variables or the potential in uencing factors different from EOGC in elderly GC patients.
As can be found in our nomogram, surgery emerged as the most signi cant predictor which could be expressed by the length of the axis standing for this. Despite several related research have reported that surgical resection could control elderly patients' condition and improve their prognosis [19][20][21], there remains controversial [22,23]. In view of its high surgical risk, poor prognosis, low quality of life and low survival rate, many clinicians are reluctant to offer surgery to advanced aged patients [24,25]. Currently there is no strong evidence to draw robust conclusions in this population. There was a report that age alone should not preclude gastrectomy in elderly patients [26]. They think that elderly patients should undergo operation because the bene t is the same as for young patients from short-term and long-term prognosis. On the other hand, Choo et al. [27] found elderly patients aged 80-85 years could have a large bene t from the surgical resection and for elderly patients aged over 86 years, especially those with cardiovascular and renal system diseases, the risks and bene ts of the procedure should be weighed before surgery. Surgery and chemotherapy should be the recommended treatment strategy in elderly patients which can effectively improve patients' survival compared with conservative treatment [28]. Consistent with previous reports, our study suggested that surgery could improve the prognosis of elderly GC patients. The patients older than 74 years who underwent surgery had a good outcome, 96% of which improve quality of life and resume daily activities [29]. Therefore, we should adapt active treatment strategies including surgery or chemotherapy to improve patients' quality of life. For patients with poor physical conditions or comorbidities, we need to weigh their surgical risk and survival bene t.
The validation of the nomogram including C-index, the ROC curves and calibration curves all showed good results, which means that the nomogram was discriminative and reliable in predicting 3-and 5-year survival of elderly GC patients. However, the discrimination and calibration could not response the clinical consequences of extent of miscalibration that could affect a model's clinical utility [30]. To evaluate its clinical effectiveness, the DCA curves was applied in our study by comparing the net bene ts with the traditional AJCC stage model. The DCA curves showed that using the nomogram in this study for predicting the short-term survival of elderly GC patients is more bene cial than the AJCC staging system. In order to get a better performance in the clinical prediction, we used the nomogram to calculate the patients' total risk score for death. Elderly patients were divided into different classi cation through X-tile analysis, which could facilitate the personal health management of elderly patients. In addition, the clinicopathological variables building the nomogram were relatively accessible, hence, the nomogram can be applied universally to clinical work. This system also can give clinicians and patients an approximate range of survival at the time of diagnosis. Prognostication is not a targeted result during disease development, as it might change with the adjustment of the patients' treatment plan and subsequent treatment response. We believe that our nomogram could assist clinicians quantify the risk factors associated with cancer death, so as to develop a suitable personalized treatment plan for elderly GC patients.
There are some potential limitations that need to be acknowledged in our study. Firstly, this was a retrospective analysis based on the SEER database, among which the recruited patients were predominantly white race. There may be potential racial heterogeneity. Secondly, several factors that are closely related to prognosis are not available in the SEER database, such as tumor markers (CEA, CA 199), surgery details (surgical approach, lymphadenectomy extent, digestive tract reconstruction) and follow-up treatment (radiotherapy, and chemotherapy). Thirdly, we did not build a nomogram to predict cancer-speci c survival (CSS) of elderly GC patients. It is important to note that gastric cancer is not the only factor affecting prognosis in this population group. Deaths due to other comorbidities account for 34%-37% of the total deaths in GC patients over eighty years [31]. Fourthly, some patients with missing data were excluded from the development and the validation sets, which may lead to selection bias.
In conclusion, we developed and validated a nomogram to predict 3-and 5-year survival rate of elderly GC patients based on a large sample cohort, which could improve the performance of the AJCC staging system. This might be extremely bene cial in assisting the prediction of survival and formulating individualized treatment protocols for elderly GC patients.

Declarations Acknowledgement
The authors acknowledge the efforts of the Surveillance, Epidemiology, and End Results (SEER) program tumor registries in the process of creating the SEER database.
Authors' contributions XZ and YD conceived the study. YW and YX collected and analyzed the data; XZ and YZ wrote the manuscript; CD and WZ reviewed and modi ed the paper. All authors approved the nal manuscript.

Funding
The study was supported by Jinan Science&Technology Bureau, No. 201704125. The funding source played no role in study design, data analysis, article writing and manuscript submitting.

Availability of data and materials
The research data is extracted from the Surveillance, Epidemiology, and End Results (SEER) using SEER*Stat software. We have obtained permission to visit them for research only (Reference number: 18942-Nov2018). The data can be used publicly.

Ethics approval and consent to participate
The SEER program database is publicly available and we signed the data usage agreement. So this study did not require the informed consent.

Consent of publication
Not applicable.