Esophagogastric junction (AEG) is located at the junction of stomach and esophagus and its clinical and pathological characteristics are different from those of gastric and esophageal cancers. The early diagnosis rate is low, and most of them have lymph node metastasis at the time of diagnosis16,17. The overall prognosis of AEG is poor due to its special anatomical location, compared with that of traditional distal gastric cancer. Studies have shown that the average surgical cure rate is about 80%, and the 5-year survival rate after radical surgery is only about 30%18,19, and Siewert et al20 reported that the 5-year survival rate after radical surgery for pancreatic cancer is 40%, and the 10-year survival rate is 20%. Marc Ychou et al21 reported an overall 5-year survival rate of 38% after AEG in 113 cases. This study showed that the 3-year survival rate after AEG was 43.8% and the 5-year survival rate was 37.0%, and its 5-year survival rate was slightly lower than the 5-year survival rate of pancreatic cancer reported by Siewert et al. and the 5-year survival rate of AEG reported by Marc Ychou et al. Therefore, this study provided a new reference for predicting the survival rate after AEG.

Accurate prediction of surgical prognosis is important for informing subsequent treatment decisions of AEG patients. Currently, prognostic prediction of postoperative tumors is majorly based on the logistic regression and Cox regression models22,23. The logistic regression model lacks survival time and, in terms of survival prediction, is inferior to the Cox regression model. In this study, the Cox-PH model was used to model and predict the survival rate of AEG, which showed better predictive outcomes. Construction of the Cox-PH model should satisfy the assumption of equal proportionality, therefore, some important prognostic factors should be discarded when constructing this model. The Cox-PH model is a linear regression, and its predictive outcomes should satisfy the linear regression equation, which cannot capture the interactions between features. Machine learning techniques can better capture the complex association between features24, thereby improving the model’s accuracy. Previously, scholars25 used artificial neural network (ANN) method to construct a machine learning model for predicting the prognosis of gastric cancer patients based on data of gastric cancer patients in local databases, but they only explored the predictive efficacy between the Union International Contra Cancrum (UICC) TNM classification system and ANN. The prediction model comprised 14 features, and lacked some important preoperative clinical blood features, which reduced the usefulness and reliability of the model in the clinical practice.

In this study, we used 19 clinical characteristics (gender, age, Borrmann staging, degree of differentiation, depth of infiltration (T stage), number of lymph node metastases (N stage), pathological TNM stage, maximum tumor diameter, postoperative chemotherapy, Fibr, D-dimer, surgical approach, postoperative hospital stay, PNI, NLR, WBR, CEA, AFP, and CA199) to construct Cox-PH models and four machine learning models to predict the 3- and 5-year survival status of patients. Among the 19 factors, the correlation between Fibr and D-dimer preoperative blood indicators with survival outcomes of gastric cancer patients has been reported26. Borrmann staging, degree of differentiation, depth of infiltration (T-stage), and the number of lymph node metastases (N stage), pathological TNM stage, and maximum tumor diameter have been shown to affect the prognosis of AEG patients27–30. The three clinical indices of infiltration depth (T stage), number of lymph node metastases (N stage), and pathological TNM stage have a high degree of overlap. To prevent overfitting of the machine learning model, two indices (infiltration depth and number of lymph node metastases) were excluded from construction of the five models. In construction of the machine learning models, cross-validation was performed in the training set for hyperparameter tuning, and each model showed its predictive efficacy. Combining the AUC values, calibration curves, and DCA curve performance of each model in the training and test set ROC curves, the XGBoost model exhibited the best performance with AUC values ≥ 0.80 in both the training and test sets. The Cox regression model also had a high predictive efficacy, however, limitations of its algorithm and the loss of important clinical features prevented it from being comparable to XGBoost. Therefore, the developed XGBoost model has a high clinical utility and reliability.

Limitations of this study: This was a single-center study with a small sample size. Machine learning models should be validated using large data sets to obtain more stable results31. The MLP in this study is a deep learning model, a subset of the ML model, an artificial neural network (ANN) with a high ability to learn simulation of nonlinear feature data. However, the MLP did not have a good predictive ability, probably because the variable features were not effectively extracted and the amount of data was small. Therefore, in follow-up studies, with large multi-center data should be used for the training and external validation tests to develop a more reliable prediction model. Second, factors that may affect the long-term prognosis of AEG patients, such as family history, smoking, and alcohol consumption, were not included in the 19 clinical observations. More factors that may affect the long-term prognostic outcomes of AEG should be included in subsequent model optimizations to continuously improve the prediction model. Finally, this study was developed and validated using retrospective data, and prospective validation studies should also be conducted to confirm the reliability of the model before formal clinical applications.

In conclusion, we constructed a Cox-PH model and a machine learning model for predicting survival risk after AEG from 19 clinicopathological features commonly observed in clinical work, with the XGBoost model showing the best efficacy. This model provides an important reference for individualized prognostic assessment and postoperative treatment decisions of AEG.