Predicting Breast Cancer Relapse from Histopathological Images with Ensemble Machine Learning Models

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Relapse and metastasis affect 30-40% of breast cancer patients, even with targeted therapies like trastuzumab for HER2-positive cases.
• Accurate individual prognosis is crucial for tailoring adjuvant treatment and enabling early intervention in breast cancer management.

Purpose of the Study:

• To develop and validate an innovative machine learning and ensemble learning framework for enhanced prediction of relapse and metastasis in HER2-positive breast cancer.
• To improve prognostic accuracy for HER2-positive breast cancer patients to guide treatment decisions.

Main Methods:

• A two-stage experimental approach was employed using The Cancer Genome Atlas (TCGA) dataset comprising 123 HER2-positive breast cancer patients.
• Six basic machine learning models (SVM, logistic regression, decision tree, random forest, AdaBoost, XGBoost) were applied, followed by ensemble techniques (weighted averaging, soft voting, hard voting).
• Performance was evaluated using accuracy, precision, sensitivity, specificity, F-Value, Mathew's correlation coefficient, and AUC.

Main Results:

• The weighted averaging ensemble model demonstrated superior performance.
• Achieved 88.46% accuracy, 89.74% precision, 94.59% sensitivity, 73.33% specificity, 92.11% F-Value, 71.07% MCC, and an AUC of 0.903.
• The framework effectively utilizes H&E images and clinical data for prediction.

Conclusions:

• The developed machine learning and ensemble learning framework significantly enhances the prediction of relapse and metastasis in HER2-positive breast cancer.
• This approach aids in more accurate prognostication, supporting better clinical decision-making for adjuvant therapy and patient management.
• The study highlights the potential of integrating imaging and clinical data with advanced computational methods for personalized cancer care.