Cancer Verse: An Explainable Digital-Twin Framework For Breast Cancer Survival Prediction Using Xgboost, SHAP, And Large Language Model

Abstract

Breast Cancer is still one of the main reasons for cancer mortality amongst women, and accurate and interpretable survival prediction can help with clinical decision-making. This paper introduces CancerVerse, an interpretable digital twin framework that combines a gradient-boosted classifier ensemble with post-hoc explanation generation and a generative language model that produces patient-specific prognostic narratives. A XGBoost classifier was trained on SEER Breast Cancer dataset (4,023 patients after pre-processing, 15 patient attributes with values converted into 26 predictor features) to predict binary survival status (Alive/Dead). The classifier demonstrated 90.68% accuracy, 80.77% precision, 51.22% recall, 62.69% F1 score and 0.8461 ROC-AUC on 20% (n=805) held-out test set. SHapley Additive exPlanations (SHAP) technique was used to generate both global feature attributions and local explanations per-patient and combine them in a structured ‘digitaltwin’ state, representing protective and risk factors of the prediction. Finally, a large language model (Gemini 2.5 Flash) was prompted with the digitaltwin state to generate a concise, clinician-readable summary without revealing the algorithmic details to the end user. This work demonstrates the potential of CancerVerse to produce trustworthy, interpretable, and communicable AI-assisted prognoses in oncology through a fully automated pipeline accessible via an interactive CancerVerse dashboard. The performance of the framework was discussed, along with its limitations regarding class imbalance and potential directions for clinical validation.