A unified multi modal transformer framework for breast cancer recurrence prediction and survival analysis

Breast cancer recurrence prediction is an important feature of post-treatment therapy, requiring accurate identification of both recurrence risk and time-to-event outcomes. In this paper, we offer a unified deep learning system that jointly performs survival analysis and multi-class recurrence classification, enabling full risk stratification for breast cancer patients. The proposed model includes a Transformer-based survival module to estimate time-until-recurrence, and an attention-guided classification module to differentiate between second primary cancer, low-risk, and high-risk recurrence instances. A multi-modal dataset comprising clinical, molecular, demographic, and lifestyle data is created from established sources like METABRIC, GSE2034, GSE2990, BCSC, and the Breast Cancer Coimbra dataset. The model uses cross-modal feature fusion, autoencoder-based dimensionality reduction, and attention-based feature attribution for applicability and accessibility. Experimental results show better accuracy, precision, recall, and F1-score of 99.12%, 98.75%, 99.08%, and 98.91%, outperforming standard machine learning and survival models. This unified paradigm gives doctors a powerful, interpretable tool for early intervention and personalized breast cancer treatment.