Machine Learning Combined with Cox Models to Identify Predictors for Post-Menopausal Breast Cancer in the UK Biobank

Researchers have combined machine learning (ML) with classical Cox models to identify predictors for incident post-menopausal breast cancer in the UK Biobank. The study focused on post-menopausal women due to the heterogeneity of breast cancer by menopause status. The researchers used SHapley Additive exPlanation (SHAP) feature dependence plots to explore potential interactions between Polygenic Risk Score (PRS) and phenotypic features. They also provided necessary statistical considerations before constructing classical Cox models to further investigate the potentially novel features selected by ML methods. The research team used the imputed genetic data from UK Biobank, followed by model-based feature selection using non-linear methods to identify candidate associated inputs, and conventional medical statistical models for maximum interpretability of the results. The tree-based eXtreme Gradient Boosting (XGBoost) machine learning algorithm was used to discover novel features among ≈1.7 k variables. The main pre-processing they performed on training data prior to ML analysis was assigning the following three categories as missing: Prefer not to answer, Do not know, and empty entries. The study provided necessary preparation before constructing classical statistical models and must not be overlooked. Following the necessary preparation, the researchers constructed a Cox proportional hazard model (i.e. the augmented Cox model) using the training data to assess the associations between novel features and incident breast cancer, adjusting for established risk factors. The researchers have shown that ML methods can be used for feature selection to complement classical statistical methods.