Machine Learning Improves the Prediction Rate of Non-Curative Resection of Endoscopic Submucosal Dissection in Patients with Early Gastric Cancer

Abstract

Simple SummaryEndoscopic submucosal dissection (ESD) is accepted as a standard treatment for early gastric cancer (EGC). Non-curative resection (NCR) of EGC after ESD can increase the burden of additional treatment and medical expenses. Thus, we aimed to develop a machine-learning (ML)-based NCR prediction model for EGC prior to ESD. We obtained data from 4927 patients with EGC who underwent ESD between January 2006 and February 2020. Seven ML-based NCR prediction models were developed using ten clinicopathological characteristics. The performance of NCR prediction was highest in the XGBoost model (AUROC, 0.851; 95% confidence interval, 0.837–0.864). Our ML model improved the ability to predict NCR of ESD in patients with EGC. This ML model can provide useful information for decision-making regarding the appropriate treatment of EGC before ESD.Non-curative resection (NCR) of early gastric cancer (EGC) after endoscopic submucosal dissection (ESD) can increase the burden of additional treatment and medical expenses. We aimed to develop a machine-learning (ML)-based NCR prediction model for EGC prior to ESD. We obtained data from 4927 patients with EGC who underwent ESD between January 2006 and February 2020. Ten clinicopathological characteristics were selected using extreme gradient boosting (XGBoost) and were used to develop a ML-based model. Dataset was divided into the training and internal validation sets and verified using an external validation set. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) were evaluated. The performance of each model was compared by using the Delong test. A total of 1100 (22.1%) patients were identified as being treated non-curatively with ESD. Seven ML-based NCR prediction models were developed. The performance of NCR prediction was highest in the XGBoost model (AUROC, 0.851; 95% confidence interval, 0.837–0.864). When we compared the prediction performance by the Delong test, XGBoost (p = 0.02) and support vector machine (p = 0.02) models showed a significantly higher performance among the NCR prediction models. We developed an ML model capable of accurately predicting the NCR of EGC before ESD. This ML model can provide useful information for decision-making regarding the appropriate treatment of EGC before ESD.