Using nomogram, decision tree, and deep learning models to predict lymph node metastasis in patients with early gastric cancer: a multi-cohort study

Abstract

The accurate assessment of lymph node metastasis (LNM) in patients with early gastric cancer is critical to the selection of the most appropriate surgical treatment. This study aims to develop an optimal LNM prediction model using different methods, including nomogram, Decision Tree, Naive Bayes, and deep learning methods. In this study, we included two independent datasets: the gastrectomy set (n=3158) and the endoscopic submucosal dissection (ESD) set (n=323). The nomogram, Decision Tree, Naive Bayes, and fully convolutional neural networks (FCNN) models were established based on logistic regression analysis of the development set. The predictive power of the LNM prediction models was revealed by time-dependent receiver operating characteristic (ROC) curves and calibration plots. We then used the ESD set as an external cohort to evaluate the models' performance. In the gastrectomy set, multivariate analysis showed that gender (P=0.008), year when diagnosed (2006-2010 year, P=0.265; 2011-2015 year, P=0.001; and 2016-2020 year, P<0.001, respectively), tumor size (2-4 cm, P=0.001; and ≥4 cm, P<0.001, respectively), tumor grade (poorly-moderately, P=0.016; moderately, P<0.001; well-moderately, P<0.001; and well, P<0.001, respectively), vascular invasion (P<0.001), and pT stage (P<0.001) were independent risk factors for LNM in early gastric cancer. The area under the curve (AUC) for the validation set using the nomogram, Decision Tree, Naive Bayes, and FCNN models were 0.78, 0.76, 0.77, and 0.79, respectively. In conclusion, our multi-cohort study systematically investigated different LNM prediction methods for patients with early gastric cancer. These models were validated and shown to be reliable with AUC>0.76 for all. Specifically, the FCNN model showed the most accurate prediction of LNM risks in early gastric cancer patients with AUC=0.79. Based on the FCNN model, patients with LNM rates of >4.77% are strong candidates for gastrectomy rather than ESD surgery.

Keywords: early gastric cancer; lymph node metastasis.

Figure 1

Flow diagram of the patient selection process in this study.

Figure 2

LNM nomogram for patients with early gastric cancer. The scores of each variable were combined to obtain the total score, and then a vertical line was subtracted from the row of total-points to estimate the risk of LNM.

Figure 3

Validation of nomogram in predicting LNM for patients with early gastric cancer. A. The ROC curve of the nomogram, with the AUC=0.80 in the development set and 0.78 in the validation set. B. The calibration plot, the reference line represents perfect agreement of the predicted probability and the actual incidence of LNM.

Figure 4

The Decision Tree of LNM prediction for patients with early gastric cancer.

Figure 5

The ROC curve of the Decision Tree model, with the AUC=0.79 in the development set and 0.76 in the validation set.

Figure 6

Comparison between the FCNN and Naive Bayes models using the validation set. The AUC was 0.79 in the FCNN model and 0.77 in the Naive Bayes model.