A Hybrid Explainable Logistic Regression model with Bayesian Feature Selection, SHAP Explainability, Stability Analysis, and Uncertainty Quantification for Breast Cancer Diagnosis

Abstract

Breast Cancer (BC) diagnosis requires predictive models with uncertainty, stability, and interpretability. Predictive models have a much better performance than traditional BC methods. The proposed framework is based on prediction, feature selection, explanation, stability, and uncertainty. This approach has provided us with both probabilistic risk estimation and an explanation of what is behind it. The present research was developed to predict uncertainty and improve clinical decision-making for further evaluation by linking Bayesian modeling with XAI, stability analysis, and uncertainty quantification. The current study mathematically presents BC classification. The framework applied is a hybrid explainable logistic regression model used to classify malignant and benign breast tumors. Moreover, Bayesian sparse feature selection was used with logistic regression, SHAP-based explanation, bootstrap stability analysis, and uncertainty on a single platform. Useful predictors were selected owing to Bayesian sparse priors, which perform model simplification and exclusion of irrelevant features. The SHAP value method was used to determine what global and local features mean in regard to prediction. Finally, bootstrap resampling was applied to carefully assess the strong and stable predictors and regression coefficients.