XS-FedPRS: Explainable and Secure Federated Learning for Privacy-Preserving Genomic Analytics in Federated Healthcare Networks

In federated healthcare networks (FHNs), collaborative learning enables distributed genomic intelligence across hospital sites and clinical institutions without centralizing sensitive patient data. However, federated polygenic risk prediction (PRS) across heterogeneous healthcare networks faces three critical challenges: population-stratified rare variant distributions that diverge across clinical sites, byzantine attacks mimicking rare variant signals, and lack of interpretability for clinical decision support. We present XS-FedPRS, a framework that decouples common variants (global aggregation) from rare variants (population-specific clustering), defends against byzantine attacks through Hardy-Weinberg equilibrium testing and gradient plausibility analysis ,and provides multi-level explainability for clinical decision support. Proof-of-Authority-and-Association (PoA2) ensures auditable governance across participating sites. Evaluation across three genomic cohorts of increasing scale and realism—CINECA synthetic (483 samples), 1000 Genomes Phase 3 (2,504 samples, 5 continental populations), and msprime-simulated (50,000 samples, Out-of-Africa demography)—demonstrates that XS-FedPRS achieves 84.4% AUC on the largest cohort under population-stratified non-IID partitioning, detects gradient poisoning, model poisoning, sybil, and combined byzantine attacks with perfect precision and recall (F1=1.0), maintains privacy (0.026 MIA advantage), and incurs only 3.3% communication overhead versus FedAvg. XS-FedPRS enables trustworthy, interpretable federated genomic analytics across heterogeneous healthcare networks devices for precision medicine.