Admission Mechanisms for Scalable Federated Health Data Collaboration: Design and Proof-of-Concept Study

Many digital health and AI initiatives show promising results in pilot settings but do not progress to sustainable services. Pilots can operate within a predefined collaboration, whereas operational services must accommodate changes in participating organizations, available resources, and analytical capabilities while preserving established governance and technical arrangements.

This study examines the pilot-to-service transition as a problem of governed federation evolution and evaluates whether Open Health can admit new forms of participation while preserving bounded capabilities, organizational sovereignty, and auditable decision evidence.

Open Health applies federated computing (FC) as a governance model separated from the analytical runtime. A governed federation context is established before execution, participants receive holder-bound capabilities, and an admission gateway produces explicit ALLOW or DENY decisions before requested operations proceed. The proof of concept uses the Flower federated learning framework and the Pathology Modified National Institute of Standards and Technology (PathMNIST) dataset from the MedMNIST collection with an initial federation comprising hospitals A and B. Mode 1A introduces hospital C as a sponsored contributor permitted to submit locally computed training updates but not invoke the resulting federated model for inference. Mode 1B examines bounded AI-agent participation, where restricted results may be transformed through policy-authorized operations into derivatives whose consumption is governed independently from the source. Evaluation considers admission conformance, signed decision evidence, analytical effects of the additional contribution, and admission overhead.

The demonstrator admitted valid member operations and hospital C's sponsored contribution (mode 1A) while denying model consumption outside hospital C's capability, scope violations, holder-binding mismatch, and federation-envelope mismatch. Mode 1B extended the same bounded-governance pattern to AI-agent participation, admitting bounded inference and governed transformation but denying unauthorized source consumption and privileged governance access. Under matched training conditions, cancer-related macro recall increased from 50.2% in the A+B control to 67.4% in A+B+C (+17.2 percentage points), while noncancer macro recall changed from 85.5% to 86.3% (+0.8 percentage points). Recall increased from 19.5% to 39.9% for cancer-associated stroma and from 80.9% to 94.9% for colorectal adenocarcinoma epithelium. Median (IQR) full admission latency was 1.210 (1.141-1.333) ms for ALLOW, 1.195 (1.136-1.273) ms for capability-scope DENY, and 0.884 (0.837-0.921) ms for holder-binding DENY. Each evaluated request generated signed decision evidence recording the decision, reason, participant and request bindings, governance context, and related model run.

Open Health demonstrates how a federated health-data collaboration can evolve through differentiated participation while preserving membership boundaries and bounded capability relations. The principal contribution is an executable governance boundary preserving capability relations and producing independently verifiable decision evidence separately from the federated-learning runtime. The analytical experiment showed that governed admission of an additional contributor produced a measurable effect on the intended cancer-related classes while preserving noncancer performance. The proof of concept establishes governance behavior under controlled conditions; clinical effectiveness and production-scale performance require further evaluation.