Performance of Statistical and Machine Learning Risk Prediction Models for Advanced Breast Cancers.

Area of Science:

• Biostatistics
• Medical Informatics
• Oncology

Background:

• Machine learning models offer complex risk prediction, but their performance relative to statistical methods is context-specific.
• Advanced breast cancer risk prediction requires robust and reliable models.

Purpose of the Study:

• To compare the performance of statistical and machine learning models in predicting advanced breast cancer risk.
• To evaluate model calibration and discrimination using large-scale screening mammography data.

Main Methods:

• Utilized data from 968,178 women (aged 40-74) from the Breast Cancer Surveillance Consortium (2005-2019).
• Cross-validated logistic regression, regularized regressions (LASSO, Elastic net), random forests, and gradient boosting models.
• Assessed model performance using calibration and area under the receiver operating characteristic curve (AUC).

Main Results:

• All models demonstrated similar discrimination (AUC 0.677-0.690).
• Regularized regressions achieved the most favorable calibration across all groups (AUC 0.689).
• Gradient boosting showed comparable AUC but suboptimal calibration, while conventional logistic regression had slightly lower AUC and calibration.

Conclusions:

• Regularized regression models provide similar discrimination and superior calibration for advanced breast cancer risk prediction.
• Regularized regression offers a practical balance of performance and interpretability, especially for rare outcomes with limited features.