Introduction: Biases in cancer incidence characteristics have led to significant imbalances in databases constructed by prospective cohort studies. Since they use imbalanced databases, many traditional algorithms for training cancer risk prediction models perform poorly.
Methods: To improve prediction performance, we introduced a Bagging ensemble framework to an absolute risk model based on ensemble penalized Cox regression (EPCR). We then tested whether the EPCR model outperformed other traditional regression models by varying the censoring rate of the simulated data.
Results: Six different simulation studies were performed with 100 replicates. To assess model performance, we calculated mean false discovery rate, false omission rate, true positive rate, true negative rate, and the areas under the receiver operating characteristic curve (AUC) values. We found that the EPCR procedure could reduce the false discovery rate (FDR) for important variables at the same true positive rate (TPR), thereby achieving more accurate variable screening. In addition, we used the EPCR procedure to build a breast cancer risk prediction model based on the Breast Cancer Cohort Study in Chinese Women database. AUCs for 3- and 5-year predictions were 0.691 and 0.642, representing improvements of 0.189 and 0.117 over the classical Gail model, respectively.
Discussion: We conclude that the EPCR procedure can overcome challenges posed by imbalanced data and improve the performance of cancer risk assessment tools.
Keywords: Absolute risk prediction; Cox Regression; Ensemble Learning; Imbalanced data.
Copyright and License information: Editorial Office of CCDCW, Chinese Center for Disease Control and Prevention 2023.