Biomarker CA125 Feature Engineering and Class Imbalance Learning Improves Ovarian Cancer Prediction
Xiaoyan Yang, Matloob Khushi, Kamran Shaukat
Abstract
Ovarian cancer is a fatal female reproductive cancer because it has no specific clinical manifestations and effective screening methods in the early stage. When it is discovered, it is already an advanced stage with a low cure rate. Therefore, it is of great significance to improve the diagnostic ability of early screening for ovarian cancer. In this study, we feature engineered CA125 by calculating the rate of change of CA125 in addition to selecting a few top-ranked important features from PLCO ovarian cancer dataset. The dataset was extremely imbalanced; the imbalance ratio (ratio of negative samples in the majority class to positive examples in the minority class) was 143.7. Twenty-three types of class-imbalanced learning methods were used in this study to improve the predictive ability of the model. We identified the decision tree method with the highest AUC value among nine classic classifiers (decision tree, AdaBoost, etc.) to build a model with the class imbalance method, and showed their comparison. We identified the decision tree using SVMSMOTE has the most robust predictive ability for ovarian cancer, with a PPV of 0.9041, AUC of 0.9532, the sensitivity of 0.7792, and specificity of 0.9982. The high PPV shown by the model selected in this study indicates that the true positive samples predicted by the model account for 90.41%. Compared with other studies, the PPV of this study increased by 81.3%. This study helps to improve the accuracy of early screening for ovarian cancer and makes the diagnosis of ovarian cancer more reliable.