Patient-record similarity retrieval over electronic health records can support cohort discovery, case comparison, treatment-history review, and personalized medicine, but most existing methods assume that clinical records can be centralized. This assumption conflicts with hospital data ownership, patient confidentiality, and institutional governance. This paper presents FedPSS, a privacy-governed federated framework for retrieving similar longitudinal clinical records across distributed BioArc hospitals. The retrieval unit is a de-identified, privacy-transformed clinical record rather than an identity-bearing patient profile, and the framework is designed for offline record retrieval, not autonomous diagnosis or prospective decision-making. FedPSS learns unified record embeddings from static structured variables, temporal visit trajectories, diagnoses, procedures, medications, laboratory series, and specialty-specific descriptors without transferring raw identifiable records to a central server. A LoRA-adapted Gemma backbone is used only as a multilingual semantic mapping layer for structured Persian-first and related Arabic clinical descriptors. Its semantic vectors are fused with temporal and categorical encoders through attention-based fusion, while training is performed with federated optimization, secure aggregation, and optional differential privacy. The contribution is systems-oriented, integrating established representation, privacy, and retrieval components into a governed BioArc workflow. Evaluation used a BioArc federation derived from 15,000,000 privacy-transformed real records across 50 hospital nodes. FedPSS achieved Precision@10 of 0.85, NDCG@10 of 0.84, and MRR of 0.82, improving over the local temporal multimodal baseline while remaining 0.02 NDCG@10 below the centralized upper bound. Results suggest that federated multilingual semantic mapping can enable scalable, privacy-governed clinical record retrieval when embedding sharing, retrieval leakage, and operational cost are explicitly controlled carefully.
