Background

Gastric cancer (GC), as the GLOBOCAN 2020 reported, is the fifth most common cancer and the fourth primary cause of tumor-related death globally with over 1.08 million new cases and nearly 0.77 million deaths (about one out of every 13 deaths died of gastric cancer) [1]. The incidence of males is more than twice that of females. Gastric cancer can be anatomically divided into two categories: gastric cardiac cancer (GCC) and other sites of GC (non-gastric cardiac cancer, NGCC). Generally speaking, the midpoint of the GCC is between 1 cm proximal and 2 cm distal from the gastroesophageal junction with endoscopic image acquisition of different parts of stomach [2] (Fig. 1). In the past half century, the incidence of NGCC has decreased, replaced by the high incidence of GCC worldwide [3, 4]. Due to the rapid progression and metastasis, the prognosis of GCC is poor, for 5-year overall survival (OS) rate is only about 9–25% [5,6,7]. Black and white ethnicity with GCC had a higher mortality rate than yellow ethnicity, and the eastern region had a better prognosis than the western region globally [8]. Given the prognosis of GCC was different from that of NGCC, it is necessary to explore this clinical issue further [9]. GCC doesn’t have specific symptoms at its early stages and lacks effective diagnostic techniques either, which might contribute to the increasing mortality rate [6, 10]. Studies have found that some clinical features were related to its poor prognosis. For example, it has been observed that tumor size and its anatomical location might be related to GCC outcome [11]. Therefore, it is encouraging and makes sense to predict the prognosis of GCC patients.

Fig. 1
figure 1

Schematic of gastric cardiac cancer (GCC). A Location of GCC. B White-light endoscopy image of GCC

Full size image

To help predict survival outcomes and make treatment decisions, the American Joint Committee on Cancer (AJCC) staging system has been developed and widely used to classify patients based on tumor, node, and metastasis (TNM) stage [12]. However, the AJCC staging system is still controversial to predict the prognosis of GCC patients who received comprehensive treatment [13]. In order to improve the accuracy of the survival estimations in GCC patients, a nomogram based on the traditional Cox proportional hazards (CPH) has been used to achieve that by some clinical researchers [14,15,16,1.

Area under the receiver operating characteristics curve (AUC) was used to evaluate model performance. A better model usually scores an AUC closer to 1. As mentioned earlier, the model was trained using the train cohort, but evaluated with both the train and the test cohorts.

Finally, neural network-based prognostic predictive model for GCC was packaged into a tool (an executable program in Microsoft Windows 11 64bit).

Statistical analysis

All statistical analyses were completed with R software (https://www.r-project.org). Wilcoxon signed-rank test was used for numerical data of skewed distribution, and Chi-square test was performed on categorical data. The Kaplan–Meier curves and log rank test were used to compare the prognosis of different tumor sites of stomach. P value of two-sided smaller than 0.05 was considered statistically significant.

Results

Prognosis of gastric cancer varies by sites

We used the SEER data and the Chinese data to compare the prognosis of GC at the different sites. GCC might have different prognosis compared with other sites of GC. In SEER data, cancers in overlap** lesion of stomach had the worst prognosis, GCC the second worst and greater curvature of stomach had the best prognosis (N = 31,397, P < 0.0001). In China data, GC in overlap** lesion still had the worst survival, but pylorus cancer had the second worst prognosis and GCC had a moderate prognosis (N = 1049, P < 0.0001) (Fig. 3).

Fig. 3
figure 3

The prognosis of gastric cancer in different sites from: A the Surveillance, Epidemiology, and End Results database (SEER) data, (B) the Chinese data

Full size image

Patients’ demographic information

Following our inclusion and exclusion criteria, 5371 patients were included finally. There were 4414 patients in train cohort and 957 patients in test cohort. As Table 1 showed, patients from both train cohort and test cohort had similar clinical features in age, sex, pathology, TNM stage, tumor size, surgery ratio, chemotherapy ratio and history of malignancy. The median age of train cohort was 67 years, and that of test cohort was the same. Train cohort included 20.80% females and test cohort had 21.11% females. The most common pathology was adenocarcinoma, in both cohorts. Most patients were staged T3, N0, or M0 in both cohorts. In train cohort, most patients were diagnosed as IV, but in test cohort IIIA was the most common stage. The median of tumor size was both 40 mm in two cohorts. In train and test cohorts, most patients got surgery or chemotherapy. But as for radiotherapy, test cohort patients most received it, while train cohort patients most not. And most patients in both two cohorts did not have a history of malignancy.

Table 1 Patients’ demographic information
Full size table

Model performance and usage

Neural network-based prognostic predictive model for GCC owned 0.7431 AUC in train cohort (95%, confidence intervals, CI, 0.7423–0.7439) and 0.7419 in test cohort (95% CI, 0.7411–0.7428) (Table 2). This model had a satisfactory performance. We then packaged it into an EXE file. When clicking the Main.exe file after unzip** Supplement File 1 (the linkage: https://drive.google.com/file/d/11-1k1rkx5fLuwcFAQuSlmqhtmVRTOt3q/view?usp=share_link), we could run the neural network-based prognostic predictive tool for GCC (Fig. 4). After clinician and researcher inputted a patient’s demographic information and clicked Predict button, the tool would calculate the OS possibility of this patient and draw his survival curves automatically. The survival curves were shown in users’ web browser, and could be zoomed in or out interactively to show the OS for a specific month.

Table 2 The performance of neural network-based prognostic predictive model for gastric cardiac cancer
Full size table
Fig. 4
figure 4

The interface of neural network-based prognostic predictive tool for gastric cardiac cancer

Full size image

Discussion

GCC is a special malignant tumor located at the gastroesophageal junction (GEJ). The mucosa of gastric cardia is mainly composed of pure mucous and mixed mucoxyntic glands, with few parietal cells and scattered endocrine cells, but no chief cells [29]. It was reported that there were great differences from GCC with tumors of esophagus or distal stomach in epidemiological and biological behavior [6]. To date, its etiology is still unclear [30]. But some previous studies have shown that the prevalence of GCC was strongly correlated with aging, smoking, young women, Helicobacter pylori infection and Epstein-Barr virus (EBV) infection [31]. Meanwhile, there is no agreement on the accurate staging of GCC patients, though some studies have shown that GCC has a better prognosis than esophageal cancer when treated according to gastric cancer stages [32]. But some researchers did observe that the prognosis of GCC might be far worse than esophagus or other GC [33, 34]. So, to explore and compare the potential prognostic difference between GC and GCC, we used the Chinses data and SEER data to conduct survival analysis. As Kaplan–Meier curves illustrated, the prognosis of GCC patient in SEER database was worse than that of NGCC patients except for cancer from overlap** lesion of stomach, while it was not the second worst in the Chinese data, which was similar to previous studies [8, 35]. Some researchers reported that these differences might be related to the surgical method and the number of lymph node resection [36, 37], but more studies are still required. Thus, as far as we have found so far, it might need to treat GC and GCC differently.

To date, surgical resection has still been the most important treatment for early GCC patients [37]. For patients who were not suitable for gastrectomy, endoscopic submucosal dissection (ESD) resection could also achieve a good prognosis because of the low rate of lymph node metastasis in early GCC [38, 39]. Long et al. [10] found that the increased lymph node removal and chemoradiotherapy (CRT) contributed to improving the survival rate of GCC patients. And various other risk factors affecting the prognosis of these patients had been reported too, including sex, age, smoking, alcohol, histological type and TNM stage [40,41,7, 43,44,45,46], evaluation of tumor invasion depth and lymph node metastasis [47,48,49] and the prediction of treatment efficacy [50, 51]. These deep learning models have shown satisfactory performance in their respective fields. Excitingly, a novel deep learning theory called DeepSurv developed by Katzman et al. [26] in 2018, which combined deep learning with ANN and CPH, has achieved initial success in the survival prediction of some cancers. For example, She et al. [52] found that DeepSurv model was significantly better than the traditional AJCC TNM staging system in non-small-cell lung-cancer-specific survival (C-index = 0.739 vs 0.706). Huang et al. [53] demonstrated that the DeepSurv model was superior to the TNM staging model in predicting esophageal CSS with the internal test dataset (C-index = 0.753 vs 0.638) and external validation dataset (C-index = 0.687 vs 0.643). These suggested that the deep learning neural network model could be more widely used as a potential tool to assist clinicians with prognosis prediction. To our knowledge, there was no study using deep learning models to predict survival in patients with GCC.

In this study, the deep learning algorithm was used to analyze the large-scale GCC clinical data and conduct a neural network tool for the first time. The AUC was 0.7431 (95% CI, 0.7423–0.7439) for the train cohort and 0.7419 (95% CI, 0.7411–0.7428) for the test cohort when applying this prediction model. These results showed that this model might have more advantages than previous models in predicting the OS of GCC patients. Finally, we converted the model into a desktop tool to use conveniently, ho** it could offer some references for clinicians and researchers (Supplement File 1).

Limitations

Some characteristic information of GCC patients from SEER database was incomplete, such as surgical methods, chemotherapy types and tumor markers, which might be important to GC patients’ prognosis. And large-scale prospective multicenter data was needed for further verification.

Conclusions

GCC patients indeed have different survival time compared with non GCC patients. And the neural network-based prognostic predictive tool developed in this study is a novel and promising software for the clinical outcome analysis of GCC patients.