Abstract
AIM: To optimize gastric cancer screening score and reduce screening costs using machine learning models. METHODS: This study included 228,634 patients from the Taizhou Gastric Cancer Screening Program. We used three machine learning models to optimize Li's gastric cancer screening score: Gradient Boosting Machine (GBM), Distributed Random Forest (DRF), and Deep Learning (DL). The performance of the binary classification models was evaluated using the area under the curve (AUC) and area under the precision-recall curve (AUCPR). RESULTS: In the binary classification model used to distinguish low-risk and moderate- to high-risk patients, the AUC in the GBM, DRF, and DL full models were 0.9994, 0.9982, and 0.9974, respectively, and the AUCPR was 0.9982, 0.9949, and 0.9918, respectively. Excluding Helicobacter pylori IgG antibody, pepsinogen I, and pepsinogen II, the AUC in the GBM, DRF, and DL models were 0.9932, 0.9879, and 0.9900, respectively, and the AUCPR was 0.9835, 0.9716, and 0.9752, respectively. Remodel after removing variables IgG, PGI, PGII, and G-17, the AUC in GBM, DRF, and DL was 0.8524, 0.8482, 0.8477, and AUCPR was 0.6068, 0.6008, and 0.5890, respectively. When constructing a tri-classification model, we discovered that none of the three machine learning models could effectively distinguish between patients at intermediate and high risk for gastric cancer (F1 scores in the GBM model for the low, medium and high risk: 0.9750, 0.9193, 0.5334, respectively; F1 scores in the DRF model for low, medium, and high risks: 0.9888, 0.9479, 0.6694, respectively; F1 scores in the DL model for low, medium, and high risks: 0.9812, 0.9216, 0.6394, respectively). CONCLUSION: We concluded that gastric cancer screening indicators could be optimized when distinguishing low-risk and moderate to high-risk populations, and detecting gastrin-17 alone can achieve a good discriminative effect, thus saving huge expenditures.