Explainable Hybrid Learning Framework for Breast Cancer Diagnosis and Progression Prediction

In this paper, a hybrid model for the detection and prognosis of breast cancer was proposed. The model integrates traditional interpretable models and deep temporal learning. Four static models, namely Logistic Regression, Calibrated SVM, XGBoost, and a custom Artificial Neural Network, using thirtyone engineered morphometric and meta-features, showed nearperfect performance with Area Under the Curve values greater than 0.99 and average precision greater than 0.99. No notable differences were found in the performance of the best models. As determined by the DeLong test, the$\mathbf{p}$-value was found to be 0.97. The SHAP-based explanation and Bootstrap Stability tests emphasized the significance of the features associated with malignant morphology. The proposed model was further extended to a deep temporal model using a bidirectional LSTM on the longitudinal data for the progression patterns, achieving a high internal Area Under the Curve score of$\mathbf{1. 0 0}$. The proposed model was found to be effective in the integration of precision, interpretability, and temporal reasoning for the detection and prognosis of breast cancer.