EOX: An Explainable Optimized XGBoost Framework for Early Breast Cancer Prediction Using Electronic Health Records

Abstract

Background:  Breast cancer is a leading cause of health-related death among females globally. To encourage effective treatment and enhance the chances of survival of a patient, early diagnosis is significant and should reach an accurate conclusion on the disease as it greatly affects one's choice of action planning. Recent progress in machine learning (ML) and explainable artificial intelligence (XAI) has opened new ways to develop robust clinical decision-support systems from electronic health record (EHR) data.

Objective:  The objective of this study is to create an Explainable Optimized XGBoost (EOX) system for early breast cancer prediction based on EHR data. Authers proposes a framework that enhances predictive performance through Bayesian hyperparameter optimization and explainable artificial intelligence using SHAP-based machine learning, bringing model transparency to the forefront.

Methods:  The EOX Framework presented in this paper incorporates data preprocessing, exploratory data analysis, feature selection and used a Bayesian hyperparameter optimization with Optuna along with an optimised XGBoost classifier. Models were evaluated using infamously known statistical parameters such as accuracy, precision, recall, specificity, f1-score, ROC-AUC, PR-AUC and Matthews Correlation Coefficient (MCC), Cohen's Kappa and Brier Score. SHAP is used to deliver interpretable explanations of individual feature contributions and model predictions.

Results:  Experimental results evidenced the success of EOX framework in obtaining robust accurate predictions achieving 97% accuracy in early breast cancer prediction ensuring model interpretability. Bayesian optimization generated classifiers with better classification accuracy as it identified more optimized hyperparameter configurations, while the feature analysis using SHAP indicated that cancer-related pathways are important corroborating features contributing to prediction outcomes.

Conclusion: The EOX as constructed confidently estimates and interprets breast cancer outcomes, and it is efficient computationally for electronic health records-based prediction of potentially preventable breast cancer incidence. The combination of optimized machine learning and explainable artificial intelligence offers a practical framework to guide future clinical integration for early diagnose and educated patient treatment decisions.