An Improved Training Algorithm Based on Ensemble Penalized Cox Regression for Predicting Absolute Cancer Risk.

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning in Healthcare

Background:

• Cancer incidence data often exhibit biases, leading to imbalanced databases in prospective cohort studies.
• Imbalanced datasets negatively impact the performance of traditional algorithms used for training cancer risk prediction models.

Purpose of the Study:

• To enhance the predictive performance of cancer risk assessment tools.
• To evaluate the effectiveness of a novel ensemble penalized Cox regression (EPCR) framework in handling imbalanced data.

Main Methods:

• Introduced a Bagging ensemble framework to an absolute risk model based on ensemble penalized Cox regression (EPCR).
• Conducted six simulation studies with 100 replicates, varying censoring rates to compare EPCR with traditional regression models.
• Assessed model performance using metrics like false discovery rate, false omission rate, true positive rate, true negative rate, and AUC.

Main Results:

• The EPCR procedure demonstrated improved variable screening by reducing the false discovery rate (FDR) at a similar true positive rate (TPR).
• Applied to the Breast Cancer Cohort Study in Chinese Women, EPCR yielded significantly higher AUC values for 3- and 5-year breast cancer risk prediction compared to the Gail model.
• Achieved AUCs of 0.691 (3-year) and 0.642 (5-year), outperforming the Gail model by 0.189 and 0.117, respectively.

Conclusions:

• The EPCR procedure effectively overcomes challenges associated with imbalanced data in cancer risk prediction.
• EPCR enhances the accuracy and performance of cancer risk assessment tools, offering a more reliable approach for clinical applications.