Key findings

  • Patient-directed FHIR exports remain far more heterogeneous than their standardized interface suggests.
  • Distributed analytics can support multicenter surveillance without centralizing records, but aggregation alone does not provide formal privacy.
  • Average privacy metrics can conceal patients with near-perfect membership-inference exposure.
  • MEDIATOR improved cross-hospital transfer by mapping heterogeneous clinical concepts into a shared discrete representation.
  • Opportunistic wearable collaboration reduced communication substantially, but testing remained largely benchmark based.
  • Personalized differential-privacy budgets are promising but require stronger attack-based and multisite validation.
  • ClinGuard offers a useful transaction-and-audit architecture for clinical agents, but its evidence remains retrospective.

Evidence

Peer-reviewed — Registry Forge processed 22,686 files and 1,791 FHIR Bundles containing 109,599 resources from 94 ALS participants across 56 health systems. Only 15.0% of files were complete C-CDA XML; HTML and RTF represented 48.4% and 34.3%. The system generated OMOP v5.4, GA4GH Phenopackets v2, dashboards and research extracts.

Peer-reviewed — A 20-month distributed laboratory study across seven hospitals used aggregate hematology data for clustering and surveillance. EWMA alerts decreased by 25%, sensitivity-oriented alerts increased by 31.2%, calibration reached 82.3% and 37 of 52 true positives were not identified by conventional quality control.

Peer-reviewed — A patient-level audit across seven medical datasets found that membership-inference attacks could achieve near-perfect success for individual patients even when aggregate attack performance was close to random. Underrepresented groups defined by disease status, race, insurance, sex and imaging protocol were disproportionately exposed.

Preprint — A follow-on privacy-screening study evaluated 7,789 models across six clinical datasets and four modalities. Stronger weight decay reduced the highly exposed cohort by as much as threefold, but the same patients often remained vulnerable across architectures, training recipes and initializations.

Preprint — MEDIATOR improved mean AUPRC over the strongest prior tokenizer by 4.0 points across 20 MIMIC-III tasks and 8.6 points in zero-shot transfer to 14 eICU tasks. It used all external events rather than the 13% available to a closed MIMIC vocabulary.

Preprint — PULSE reduced peer-assisted wearable exchanges by 80%–94% across activity, fall and stress-recognition tasks. Its device implementation used 149.8 KiB of RAM and 561.8 KiB of flash.

Preprint — ClinGuard represented clinical coordination as patient-scoped, policy-bound transactions validated by an independent sentinel layer, restricted connectors and tamper-evident audit records. Its demonstration used MIMIC-IV-FHIR and skilled-nursing-facility placement rather than a prospective deployment.

Limitations

Registry Forge was demonstrated in one rare-disease registry and did not establish independent replication, complete extraction accuracy or clinical-outcome improvement.

The distributed laboratory study exchanged aggregate information but reported no differential privacy, secure aggregation, reconstruction testing or diagnosis-level validation.

Patient-level privacy screening identifies vulnerability but does not itself provide a mathematical privacy guarantee.

MEDIATOR used retrospective critical-care benchmarks, and semantic transfer accuracy does not prove safe clinical portability.

PULSE relied on benchmark tasks and simulated or recorded encounters rather than a long-term clinical wearable deployment.

Personalized privacy budgets can create uneven protection and require transparent justification, formal accounting and attack-based evaluation.

ClinGuard’s governance design remains a retrospective demonstration rather than evidence that autonomous coordination is safe.

What’s worth reading or building on

Read

1. Registry Forge — strongest practical contribution for patient-directed FHIR ingestion.

2. Disparate privacy risks from medical AI — most important evidence that average privacy scores are inadequate.

3. Automated screening of privacy risks — useful pretraining privacy-triage method.

4. MEDIATOR — promising approach to semantic portability across hospitals.

5. PULSE — technically credible resource-aware wearable collaboration.

6. ClinGuard — useful governance architecture awaiting prospective evaluation.

Build

FHIR ingestion systems that preserve original payloads, transformation lineage and terminology versions.

Patient-level and subgroup-level privacy dashboards alongside average attack metrics.

Secure aggregation and differential privacy layered onto distributed analytic pipelines.

Semantic harmonization tests covering local codes, synonyms, missing concepts and free text.

Wearable federations that measure energy consumption, connectivity failures, drift and clinical usefulness.

Clinical-agent architectures that separate proposing, validating, authorizing, executing and auditing actions.

Sources