Key findings

  • FHIR and CQL can encode complex quality indicators reproducibly, but semantic and patient-level validation remain incomplete.
  • Culturally contextual variables improved interpretable perinatal mental-health prediction among Aboriginal mothers.
  • Routine EHR data showed potential for identifying future inflammatory bowel disease before diagnosis.
  • Brain-MRI foundation-model embeddings permitted highly accurate patient linkage and sensitive-attribute inference.
  • Infrastructure-as-Code can make medical-AI deployments more reproducible and auditable, but policy checks do not establish clinical safety.
  • No study validated a complete FHIR-native, formally private federated chronic-care deployment.

Evidence

Peer-reviewed — Researchers implemented 39 Taiwan National Health Insurance quality indicators using 21 data elements and six FHIR R4 resources. All libraries executed, 20 medication indicators produced identical aggregate numerators and denominators against an independent Python implementation, 2,317 resources passed structural validation and medication terminology mapping reached 98.3%.

Peer-reviewed — An interpretable perinatal mental-health model used data from 288 Aboriginal mothers across six Western Australian sites. Explainable Boosting Machines with training-only synthetic augmentation achieved an internally evaluated AUC of 0.934 and F1 of 0.817, with loneliness, self-blame, family pride, sleep and daily-life management among the influential variables.

Peer-reviewed — A two-year observational study of 178 people with bipolar disorder used wearable sleep and activity signals to detect hypomanic symptoms. The effective symptomatic modeling subset contained 43 participants, with reported AUROC of 0.93, sensitivity of 92.0% and negative predictive value of 97.8%.

Peer-reviewed — Explainable Boosting Machines trained on routine data from 81,960 individuals, including 21,695 with inflammatory bowel disease, achieved AUROCs of 0.71–0.75 for Crohn’s disease up to two years before diagnosis and approximately 0.70–0.72 in selected age-based screening settings.

Preprint — Longitudinal EHR modeling for incident inflammatory bowel disease achieved an AUROC of 0.969 at UCSF and 0.965 without recalibration at a demographically different safety-net hospital. Median lead time was 20 months at UCSF and 15 months at the external site, although alert-level positive predictive value remained low.

Preprint — Brain-MRI foundation-model representations produced rank-one patient linkage of 99% on MIRIAD and greater than 90% on OASIS-2. Sex inference reached 90.9% and 95.5%, showing that embeddings can preserve identifying and sensitive information.

Preprint — An Azure Infrastructure-as-Code framework tested 25 deployments across five models. Its automated checker flagged 105 of 1,347 evaluations, with 149 of 150 manually reviewed findings confirmed and reported positive predictive value of 98.2%.

Preprint — MedHarnessBench evaluated more than 16,000 controlled executions across 342 FHIR, EHR, administrative and multimodal tasks, demonstrating that healthcare-specific runtime support can reveal process failures hidden by terminal success metrics.

Limitations

The FHIR/CQL study established aggregate implementation agreement for only 20 of 39 indicators and did not test patient-level agreement, independent CQL engines, large-scale performance or clinical outcomes.

The Aboriginal perinatal model used a small cohort, repeated internal splits and synthetic augmentation without external or prospective validation.

The bipolar-wearable study detected concurrent symptoms rather than forecasting future episodes and relied on a small symptomatic subset.

The inflammatory-bowel-disease models were retrospective; high AUROC did not eliminate low alert-level positive predictive value or prove that earlier alerts improve diagnosis or outcomes.

Embedding-linkage attacks assumed auxiliary imaging or encoder access and were tested on two research datasets, but the results still invalidate claims that embeddings are inherently anonymous.

Infrastructure-as-Code improves reproducibility and governance enforcement but provides neither formal privacy protection nor clinical-effectiveness evidence.

What’s worth reading or building on

Read

1. Brain-MRI representation privacy study — clearest evidence that shared embeddings can reveal patient identity.

2. FHIR/CQL quality-indicator implementation — strongest standards-based infrastructure contribution.

3. Incident inflammatory-bowel-disease prediction — strongest externally tested chronic-disease prediction study.

4. Aboriginal perinatal mental-health modeling — valuable example of culturally contextual and interpretable prediction.

5. Infrastructure-as-Code for medical AI — practical governance architecture for controlled cloud deployment.

6. MedHarnessBench — useful framework for evaluating healthcare agents beyond final-answer accuracy.

Build

Patient-level validation for FHIR/CQL measures across independent servers and terminology implementations.

Privacy testing for clinical embeddings before they are exchanged, published or used in federated systems.

FHIR-native longitudinal prediction pipelines with explicit deployment-burden and false-positive measurements.

Participatory SDoH model development that preserves culturally specific variables and local governance.

Version-controlled cloud infrastructure with model, data, policy and audit provenance.

Benchmarks that test authorization boundaries, workflow continuity, verification and auditability rather than answer accuracy alone.

Sources