Reward-Guided Flow Matching for Breast Cancer Digital Twins in Muslim-Majority Populations

We develop a reward-guided conditional flow framework for multimodal breast cancer digital twins using 180 matched TCGA-BRCA patient states. Flow matching achieves the best in-domain alignment (MMD $=0.0605$; Wasserstein $=0.4265$), but shows substantial zero-shot shift on two independent Saudi cohorts (MMD $=0.9013$ and $1.0323$). Target calibration substantially improves flow alignment on held-out external evaluation subsets, reducing MMD to 0.5797, 0.3754, and 0.1468 for GSE36295, GSE29044, and the pooled cohort, respectively. Risk steering produces monotonic control of the model-defined mortality-risk score ($\rho=1.0$; zero violations across seven levels). These results highlight the importance of target-domain calibration and strong statistical baselines for cross-cohort generative modeling.