Browser-Native Federated Inference on Existing Italian SSN Clinical Workstations: A Peer-to-Peer Sovereignty-Preserving AI Architecture for Italian Regional Health Networks

Abstract

Clinical adoption of large language models (LLMs) in public healthcare faces a structural impasse: the capital expenditure of centralised high-performance computing on one side and the privacy risk of routing patient data through third-party cloud interfaces on the other. Italian local health authorities (Aziende Sanitarie Locali, ASL) operate large fleets of clinical workstations that remain idle outside peak administrative hours. We present OmniMed Federated, a browser-native architecture using the WebGPU application programming interface (API) and the WebLLM framework to distribute LLM inference tasks across these existing workstations. The system federates task allocation rather than model training or partitioned inference: each query executes in full on one node, selected under a data residency constraint. A five-tier escalation model, coordinated by a metadata-only PHP back end, ranks tiers by data exposure rather than capability, with commercial cloud fallback disabled by default. In a pilot three-node testbed (50 queries), federated throughput reached 19.5 versus 8.2 tokens/second standalone, peak per-node memory fell 62%, and node discovery took 140 ms; query content remained within the institutional perimeter throughout. These figures establish infrastructural feasibility at pilot scale. Clinical output quality, security hardening, and scalability remain unevaluated.