Abstract

Background: Breast cancer affects millions of women; with the increasing growth in data collection in recent years, machine learning models are used in the diagnosis phase. While the accuracy of the models plays a significant role in choosing a model, the interpretability of the model for doctors and decision-makers is crucial in understanding and building trust in breast cancer diagnosis. In practice, it is challenging for researchers and practitioners to select the optimal model based on multiple objectives such as accuracy, interpretability, and computational runtime. We proposed a model selection technique unifying various objectives based on K-means clustering. This study's main contribution is the use of interpretable machine learning techniques such as LIME, ELI5, and SHAP and machine learning algorithms to predict the tumor type. Materials and Methods: The data used in this study were collected by Dr. William H. Wolberg, W. Nick Street, Olvi L. Mangasarian from the University of Wisconsin Hospitals, and donated to the UCI machine learning repository by Nick Street. Forty-three models are built using the dataset. The runtime for each model is recorded in seconds, and the Accuracy, balanced Accuracy, the AUC-ROC, the F1-score, and the interpretability tool are compiled. A K-Means clustering algorithm is applied to the resulting outputs. Through the elbow method, three categories of clusters are selected. Result: The proposed method showed high performance, as well as ease in interpreting the model. The k-means clusters' characteristics show that models in cluster number 2 have low and medium interpretability and low computation runtime. Conclusion: AdaBoost and XGBoost Classifiers with ELI5 interpretability are the most performant and most explainable models. They show the highest accuracy and the lowest computation runtime, and each prediction is explained by a linear combination of the top features.

Keywords

Breast cancer diagnosis, Interpretable machine learning, Interpretability, Explainability, Model selection, K-means clustering