An Explainable Hybrid GWO-LightGBM Model for Breast Cancer Diagnosis Using SHAP Interpretation

Abstract

The rapid growth of machine learning and data mining has changed the discipline of medical diagnostics. They developed automated systems that can identify complicated disease patterns with high accuracy. Most current models, on the other hand, still include redundant and correlated features that make computations more expensive to understand and sometimes cause overfitting. This research proposes a hybrid diagnostic system that combines Grey Wolf Optimization (GWO) for feature selection alongside the Light Gradient Boosting Machine (LightGBM) classifier to improve both accuracy and interpretability. It examines the suggested GWO-LightGBM model on the Breast Cancer Wisconsin dataset. The framework successfully reduced the number of input features from 30 to 12. In addition, the test accuracy of 97.37%, and the cross-validation accuracy of 98.02% ± 0.02. This was better than the baseline LightGBM that was trained on all features. Furthermore, the model shortened the training time by 25% and showed statistically significant improvement (p < 0.05).  Furthermore, the SHAP analysis exposed that the selected features were biologically important, which contributed to the model's transparency and trustworthiness. The proposed model shows that using feature selection with LightGBM and explainable artificial intelligence may make diagnostic models that are fast, easy to understand in healthcare applications.