Deep Learning-Based Survival Analysis and Recurrence Prediction in Breast Cancer Patients Using Clinical and Genomic Data

Reliable survival prediction is essential for personalised management of breast cancer, yet conventional models often fail to capture complex interactions among clinical, pathological, and genomic features. This study applied a deep learning framework (DeepSurv) to a cohort of 2,509 breast cancer patients, integrating clinical, histopathological, and genomic data to predict overall survival (OS) and relapse-free survival (RFS). Exploratory analysis revealed a median age at diagnosis of 60.4 years, a median tumour size of 26 mm, and a median of 0 positive lymph nodes, with a relapse rate of ~40%. DeepSurv demonstrated superior predictive performance compared to classical Cox regression, achieving C-index values of 0.7567 (OS) and 0.6495 (RFS) versus 0.7038 (OS) and 0.6403 (RFS) for Cox models. SHAP analysis identified positive lymph nodes, tumour grade, tumour size, age, and Nottingham Prognostic Index (NPI) as the most influential predictors, while mutation count and treatment variables contributed moderately. Survival curves indicated higher individualised survival probabilities with DeepSurv, reflecting improved sensitivity to patient-specific risk patterns. Classical Cox regression performed adequately but exhibited reduced discriminatory power, particularly for RFS. These findings demonstrate that deep learning models can integrate multi-modal data, enhance predictive accuracy, and maintain interpretability, supporting patient stratification and informed clinical decision-making. Future work should incorporate additional molecular and treatment-response data to improve relapse prediction further.