Ensemble machine learning for predicting breast cancer recurrence and mortality using clinical and hemogram data.

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Breast cancer prognosis varies significantly by molecular subtype, necessitating personalized prognostic tools.
• Current prognostic methods may lack scalability, affordability, or interpretability.
• Predicting recurrence risk is crucial for guiding treatment intensity and improving patient outcomes.

Purpose of the Study:

• To develop and validate machine learning models for predicting 2- and 10-year breast cancer recurrence or death.
• To integrate readily available hematological indices and clinicopathological data for prognostic assessment.
• To create an interpretable tool for identifying high-risk patients across different molecular subtypes.

Main Methods:

• Retrospective analysis of 4277 women with primary breast cancer (2008-2022) across HR+, HER2+, and TNBC subtypes.
• Development of a stacked ensemble machine learning model using logistic regression, incorporating SMOTE for class imbalance.
• Utilized SHAP analysis to identify key predictors of recurrence risk.

Main Results:

• The ensemble model demonstrated strong discrimination, with Area Under the Curve (AUC) of 0.859 (2-year) and 0.811 (10-year) for the general cohort.
• Robust subtype-specific performance was observed: HR+ AUC 0.862/0.804, HER2+ AUC 0.877/0.831, TNBC AUC 0.826/0.826 (2-year/10-year).
• Key adverse predictors included advanced tumor stage, elevated inflammatory ratios (NLR, PLR, MLR), high red cell distribution width, and age.

Conclusions:

• Machine learning models integrating hematological indices and clinicopathological data offer a scalable, affordable, and interpretable approach to breast cancer prognostication.
• The developed tool effectively predicts recurrence risk across diverse molecular subtypes and time horizons.
• This prognostic tool can aid in stratifying patients for intensified therapy, particularly in resource-limited settings where genomic testing is inaccessible.