Interpretability methods of machine learning algorithms with applications in breast cancer diagnosis

Area of Science:

• Medical Informatics
• Machine Learning
• Oncology

Background:

• Early breast cancer detection significantly reduces socioeconomic burden.
• Artificial intelligence (AI) shows promise in breast cancer diagnosis but faces adoption challenges due to its "black box" nature.
• Interpretability methods are crucial for understanding and trusting AI in clinical settings.

Purpose of the Study:

• To develop and evaluate interpretable AI models for breast cancer diagnosis.
• To enhance the performance of AI algorithms through feature selection guided by interpretability techniques.
• To validate AI model decisions against existing medical knowledge and explore new pathophysiological insights.

Main Methods:

• Utilized Random Forests (RF), Neural Networks (NN), and Ensembles of Neural Networks (ENN) on the Wisconsin Diagnostic Breast Cancer (WDBC) dataset.
• Applied interpretability techniques including Global Surrogate (GS), Individual Conditional Expectation (ICE) plots, and Shapley Values (SV).
• Performed feature selection based on feature importance derived from GS and SV methods to optimize model performance.

Main Results:

• The proposed ENN model achieved 96.6% accuracy and 0.96 area under the ROC curve.
• ICE plots confirmed that ENN decisions align with medical knowledge, offering potential for new insights.
• Feature selection improved RF accuracy to 97.18% and NN accuracy to 95.53%, with corresponding ROC AUC improvements.

Conclusions:

• Interpretable AI models, particularly the proposed ENN, demonstrate state-of-the-art performance for breast cancer diagnosis.
• Interpretability techniques enhance AI model understanding, trust, and performance.
• The developed methods offer a pathway for reliable AI integration into clinical breast cancer screening and diagnosis.