Predicting Breast Cancer in Chinese Women Using Machine Learning Techniques: Algorithm Development

Background Risk-based breast cancer screening is a cost-effective intervention for controlling breast cancer in China, but the successful implementation of such intervention requires an accurate breast cancer prediction model for Chinese women. Objective This study aimed to evaluate and compare the performance of four machine learning algorithms on predicting breast cancer among Chinese women using 10 breast cancer risk factors. Methods A dataset consisting of 7127 breast cancer cases and 7127 matched healthy controls was used for model training and testing. We used repeated 5-fold cross-validation and calculated AUC, sensitivity, specificity, and accuracy as the measures of the model performance. Results The three novel machine-learning algorithms (XGBoost, Random Forest and Deep Neural Network) all achieved significantly higher area under the receiver operating characteristic curves (AUCs), sensitivity, and accuracy than logistic regression. Among the three novel machine learning algorithms, XGBoost (AUC 0.742) outperformed deep neural network (AUC 0.728) and random forest (AUC 0.728). Main residence, number of live births, menopause status, age, and age at first birth were considered as top-ranked variables in the three novel machine learning algorithms. Conclusions The novel machine learning algorithms, especially XGBoost, can be used to develop breast cancer prediction models to help identify women at high risk for breast cancer in developing countries.