Development of Breast Cancer Prognosis Prediction Model Based on Clinical Features Including CEA and CA15-3 Serum Levels

Abstract

Objectives: Serum levels of carcinoembryonic antigen and cancer antigen 15-3 tumor markers are used for breast cancer prognosis. This study developed a breast cancer prognosis prediction model.Methods: We retrospectively analyzed data of 639 patients diagnosed between January 2012 and December 2019. We selected 20 independent variables with carcinoembryonic antigen and cancer antigen 15-3 serum levels and employed four machine-learning algorithms for the model: artificial neural network, random forest, support vector machine, and logistic regression.Results: Significant differences in carcinoembryonic antigen and cancer antigen 15-3 serum levels, age, history of other diseases excluding hypertension and diabetes mellitus, chemotherapy, and drug therapy were noted between control (n = 576) and case groups (n = 63). The sensitivity and specificity of the artificial neural network model for prognosis prediction were 26.7% and 92.6%, respectively.Conclusions: Carcinoembryonic antigen and cancer antigen 15-3 serum levels were the most significant variables for developing a breast cancer prognosis prediction model using the Shapley additive explanations model. The proposed machine-learning model and tumor marker serum levels may be useful for breast cancer prognosis.