Extracting Knowledge from Machine Learning Models to Diagnose Breast Cancer

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Traditional breast cancer diagnosis often relies on medical imaging and demographic data.
• There is a growing need for trustworthy AI in healthcare, particularly for diagnostic tools that offer explainability.
• Serum biomarkers present an alternative and complementary data source for cancer detection.

Purpose of the Study:

• To apply explainable machine learning models for breast cancer diagnosis using serum biomarkers.
• To identify key predictive biomarkers and understand their role in breast cancer pathogenesis.
• To develop accurate and interpretable AI models for enhanced early breast cancer detection.

Main Methods:

• Evaluation of several explainable classification models: OneR, JRIP, FURIA, J48, ADTree, and Random Forest.
• Utilized a dataset comprising various serum biomarkers including electrolytes, metal ions, proteins, hormones, and BMI.
• Assessed model performance based on accuracy and the interpretability of identified predictive factors.

Main Results:

• The Random Forest model achieved the highest accuracy (99.401%), with other models also showing high performance (98.204%-98.802%).
• Oxytocin was consistently identified as a key predictive biomarker across most models.
• Other significant biomarkers included GnRH, β-endorphin, vasopressin, iron, and progesterone, among others.

Conclusions:

• Explainable machine learning models utilizing serum biomarkers show significant potential for accurate breast cancer diagnosis.
• The identification of key biomarkers like oxytocin aids in understanding breast cancer pathogenesis and offers therapeutic targets.
• This approach supports improved early detection, personalized treatment, and advances in breast cancer management.