Machine learning to predict distant metastasis and prognostic analysis of moderately differentiated gastric adenocarcinoma patients: a novel focus on lymph node indicators

Abstract

Background: Moderately differentiated gastric adenocarcinoma (MDGA) has a high risk of metastasis and individual variation, which strongly affects patient prognosis. Using large-scale datasets and machine learning algorithms for prediction can improve individualized treatment. The specific efficacy of several lymph node indicators in predicting distant metastasis (DM) and patient prognosis in MDGA remains obscure.

Methods: We collected data from MDGA patients from the SEER database from 2010 to 2019. Additionally, we collected data from MDGA patients in China. We used nine machine learning algorithms to predict DM. Subsequently, we used Cox regression analysis to determine the risk factors affecting overall survival (OS) and cancer-specific survival (CSS) in DM patients and constructed nomograms. Furthermore, we used logistic regression and Cox regression analyses to assess the specific impact of six lymph node indicators on DM incidence and patient prognosis.

Results: We collected data from 5,377 MDGA patients from the SEER database and 109 MDGC patients from hospitals. T stage, N stage, tumor size, primary site, number of positive lymph nodes, and chemotherapy were identified as independent risk factors for DM. The random forest prediction model had the best overall predictive performance (AUC = 0.919). T stage, primary site, chemotherapy, and the number of regional lymph nodes were identified as prognostic factors for OS. Moreover, T stage, number of regional lymph nodes, primary site, and chemotherapy were also influential factors for CSS. The nomograms showed good predictive value and stability in predicting the 1-, 3-, and 5-year OS and CSS in DM patients. Additionally, the log odds of a metastatic lymph node and the number of negative lymph nodes may be risk factors for DM, while the regional lymph node ratio and the number of regional lymph nodes are prognostic factors for OS.

Conclusion: The random forest prediction model accurately identified high-risk populations, and we established OS and CSS survival prediction models for MDGA patients with DM. Our hospital samples demonstrated different characteristics of lymph node indicators in terms of distant metastasis and prognosis.

Keywords: distant metastasis; lymph node indicators; machine learning; moderately differentiated gastric adenocarcinoma; nomogram; prognosis.

Figure 1

Flowchart of the data screening process. The figure shows the process of filtering eligible patient data from the SEER database.

Figure 2

Data analysis guide. The figure shows the procedure of this study for processing and analyzing the screened data.

Figure 3

Results of Pearson correlation analysis for each variable (A) and ranking of importance of predictive model characteristics (B). The results of Pearson correlation analysis for each variable (A) showed that all variables existed independently of each other. The predictive model characteristics (B) were the number of Reg LNs, N stage, T stage, chemotherapy, age, tumor size, and primary site, in order of importance.

Figure 4

Receiver operating characteristic (ROC) curves for the training set, test set, and external validation set prediction models. (A) Training set; (B) test set; (C) external validation set. The aggregation of the previous ROC curves for the RF model (D). AUC, area under the ROC curve; RF, random forest; LR, logistic regression; LASSO, least absolute shrinkage and selection operator; SVM, support vector machine; KNN, K-nearest neighbor; NBC, naive Bayes classifier; ANN, artificial neural network.

Figure 5

Nomograms for 1-, 3-, and 5-year OS (A) and CSS (B) in MDCA patients with DM.

Figure 6

Evaluation of the ability of the nomogram to predict OS. ROC curves validating the OS prediction nomogram for 1-, 3-, and 5-year OS in the training set (A) and validation set (B). Calibration curves validating the OS prediction nomograms for 1-, 3-, and 5-year OS in the training set (C–E) and validation set (F–H). Decision curve analysis validating the OS prediction nomogram for 1-, 3-, and 5-year OS in the training set (I–K) and validation set (L–N).

Figure 7

Evaluation of the ability of the nomogram to predict CSS. ROC curves validating the CSS prediction nomogram for 1-, 3-, and 5-year RFS in the training set (A) and validation set (B). Calibration curves validating the CSS prediction nomograms for 1-, 3-, and 5-year survival in the training cohort (C–E) and validation cohort (F–H). Decision curve analysis validating the CSS prediction nomogram for 1-, 3-, and 5-year RFS in the training set (I–K) and validation set (L–N).