In the present study, we established and externally validated a nomogram model for predicting DSS after D2 gastrectomy for AGC based on a large Chinese cohort. This nomogram was more significantly predictive than the eighth edition AJCC staging, with a C-index of 0.82 and good calibration. Compared with previous studies [3, 15, 16], this study included a larger number of patients with AGC (n = 6753), and the nomogram had a higher precision. Furthermore, our nomogram considered the unique characteristics of the patient population, thus addressing discrepancies in previous models regarding demographic features, stage distribution, and treatment strategies of US, Korean, or Japanese gastric cancer patients.
In contrast to the previous nomogram based on US [15], Korean [3], or Japanese [16] studies, sex was excluded from the nomogram in our present study. This is interesting because female gastric cancer patients had better prognoses in both the univariate and multivariate analyses in all of the previously mentioned databases. However, in our study cohort, the superior survival rate of female patients was insignificant (P = 0.360). We believe that this finding represents one of the most important differences in demographic features and that it may be due, in part, to disadvantages in terms of socioeconomic status of the female population in China.
In our nomogram, Lauren classification was included because of its independent prognostic effect, which corresponds with the US [15] database but not the Korean [3] and Japanese [16] databases. This variable was excluded from the Korean database due to a large percentage of missing data and was replaced with macroscopic and histologic types in the model from the Japanese database. Previous research demonstrated a relationship between Lauren classification and the effect of trastuzumab in gastric cancer patients [21]. Therefore, because one purpose of the nomogram was to guide clinical practice, we chose to include the Lauren classification as an important predictive variable in our nomogram.
One important improvement of our nomogram is that we chose to include metastatic lymph node ratio, which refers to the number of metastatic lymph nodes resected divided by the number of total lymph nodes resected. In our previous studies, we successfully verified that the metastatic lymph node ratio had theoretical and practical advantages for decreasing the staging migration phenomenon and improving prognostic prediction [22,23,24]. However, based on the independent datasets from the three hospitals included in the present study, the optimal cutoff values of metastatic lymph node ratio for prediction were different. In contrast to models based on the TNM staging system, the predictive model based on the nomogram can directly use a continuous variable without requiring this variable to be converted into a categorical variable. We believe that this model may be considered a general predictive tool for future clinical use. Another important improvement of our nomogram is the use of continuous variables for the establishment of the model. In the present study, we tested various methods of treating or converting continuous variables such as age, tumor size, and metastatic lymph node ratio. We determined that the use of continuous variables resulted in the most precise predictions.
The accuracy of a predictive model refers to the ability of the model to discriminate between patients with or without the outcome of interest, which is demonstrated by the C-index for censored data. In our nomogram model, the C-index values were high: 0.82 in the training set and 0.83 and 0.81 in validation sets. Furthermore, a model’s accuracy indicates not only the overall ability to predict the outcome of interest but also the ability to predict the outcome of interest in specific patient groups or according to risk level. For this purpose, calibration plots should be obtained for both internal and external data. Unlike previous studies, our study investigated all three independent datasets.
We hypothesize that the inclusion of complete and precise data (including all follow-up data), the conversion of continuous data, and the stage distribution contributed to that our study has a higher C-index value compared with previous studies [3, 15, 16]. In our research dataset, we excluded all cases with missing data to improve the precision of the predictive model. For the same reason, we tried various methods to convert the continuous variables (including age, tumor size, and metastatic lymph node ratio) to maximize the predictive ability of the model. The predictive ability increased gradually from 0.76 to 0.82. Furthermore, the stage distribution was distinct from that observed in previous studies [3, 15, 16], for example, in studies using a dataset including early-stage gastric cancers (the US and Korean datasets) or excluding locally AGC (the Japanese dataset). On the other hand, with advanced disease, the demand for precise prediction of survival probability becomes increasingly important for clinical decisions regarding the use of adjuvant chemotherapy or chemoradiotherapy. For this reason, we think that precise prediction tools are more important in AGC patients than in patients with relatively early-stage diseases.
This study has some limitations. The main limitation of the present study was that adjuvant therapy, such as chemotherapy or chemoradiotherapy, was not included as a variable in the predictive model. In the multivariate analysis using the Cox regression model, adjuvant therapy did not appear to be important for prognosis. In clinical practice in China, patients with stage II or III disease usually receive adjuvant chemotherapy with a relatively uniform protocol including 5-fluorouracil- and/or platinum-based therapeutics. In some patients with high-risk recurrent diseases, adjuvant chemoradiotherapy is also recommended. Therefore, adjuvant therapy was not included as a variable in the present study. Another limitation was that our external validation set was based on the Chinese population. This is a large and multicenter database on Chinese patients has been used to establish such a model. In addition, the baseline parameters at the three hospitals were different because the patients from each hospital come from specific regions of China and thus have comparatively different demographics. However, despite these differences, the nomogram demonstrated great accuracy as shown by the C-indexes (SYSUCC: 0.82; CMU: 0.83; and TJMU: 0.81). Whether this predictive model is suitable for Western, US, or other Asian countries is unknown. Therefore, in future studies, it will be necessary to validate this model using data from other countries.
In conclusion, in the present study, we established a nomogram for precisely and individually predicting DSS for Chinese patients with AGC who underwent radical resection with D2 lymphadenectomy based on databases from multiple centers. Its superiority to the traditional TNM staging system may enhance its clinical use in the near future.