FLARE: Federated Learning and Reinforcement Optimization for Cancer Survival Models

Breast cancer remains a leading cause of mortality among women, with accurate survival prediction posing a major clinical challenge. While computational advances and abundant healthcare data enable data-driven prognostic models, data privacy remains a key barrier to multi-institutional collaboration. This study proposes FLARE, a novel framework integrating Federated Learning (FL) and Reinforcement Learning (RL) to predict breast cancer survival. Our approach jointly trains Random Survival Forests (RSF), Cox Proportional Hazards (CoxPH), and DeepHit across distributed datasets while preserving privacy. RL dynamically optimizes model parameters to enhance predictive accuracy and generalization. Experiments on the METABRIC dataset show that our FL-RL-enhanced RSF model achieves a C-index of 0.88, outperforming traditional survival models. This work underscores the potential of FL and RL in advancing clinical prognostics while safeguarding patient data.