Preoperative prediction for lymph node metastasis in early gastric cancer by interpretable machine learning models: A multicenter study

Abstract

BackgroundThe presence of lymph node metastasis plays a decisive role in the selection of treatment options in patients with early gastric cancer. However, there is currently no established protocol to predict the risk of lymph node metastasis before/after endoscopic resection. The aim of this study was to develop and validate several machine learning algorithms for clinical practice. MethodsA total of 2,348 patients with early gastric cancer were selected from 5 major tertiary medical centers. We applied 6 machine learning algorithms to develop lymph node metastasis prediction models for clinical feature variables. The partial dependence plots were used to explain the prediction of the models. The area under the receiver operating characteristic curve and area under the precision recall curve were measured to assess the detection performance. The R shiny interactive web application was used to translate the prediction model in a clinical setting. ResultsThe incidence of lymph node metastasis in patients with early gastric cancer was 13.63% (320/2348) and significantly higher in young women, in the lower third of the stomach, with a size >2 cm, depressed type, poorly/nondifferentiated, lymphovascular invasion, nerve invasion, and submucosal infiltration. In terms of age, there is a nonlinear and younger trend. XGBOOST displayed the best predictive performance at the initial and postendoscopy evaluation. In addition, the machine learning algorithm was converted to a user-friendly web tool for patients and clinicians. ConclusionXGBOOST can predict the risk of lymph node metastasis with best accuracy in patients with early gastric cancer. Our online web application may help determine the optimal best surgical option for patients with early gastric cancer.