Key findings

  • Federated clinical language models can approach pooled-data performance without moving raw records, but this does not establish formal privacy.
  • Clinical sentences can be reconstructed from federated model updates under an active malicious-server attack.
  • A population-scale hypertension study found that transparent logistic regression performed comparably to a more complex machine-learning model.
  • Federated diabetes research is beginning to optimize calibration and explanation stability together, but current explanations remain noncausal.
  • FHIR-native wearable infrastructure and reproducible federation tools are technically credible but lack prospective chronic-disease validation.
  • No consequential study combined FHIR ingestion, formal privacy protection, SDoH-aware prediction, explainability and clinical outcomes in one validated system.

Evidence

Peer-reviewed — MEDAL sequentially trained LLaMA-3.1-8B adapters across MIMIC-IV and adult and pediatric UCSF ICU cohorts for sepsis identification and mortality prediction. It approached centralized pooled-data performance and reportedly exceeded GPT-5 and locally fine-tuned models by as much as 16.6 F1 points. Evidence strength is moderate because the evaluation used real institutional datasets but only two retrospective NLP tasks.

Peer-reviewed — A gradient-reconstruction study using 368,751 diagnostic reports, 98,206 discharge summaries and 1,500 MIMIC-CXR reports found that an active malicious server could exactly reconstruct approximately 31%–44% of short clinical sentences. Larger batches reduced but did not eliminate leakage. Evidence strength is moderate to strong for the defined attack model.

Peer-reviewed — An independent NHS evaluation tested hypertension case-finding among 1,802,920 people in northwest London. Sensitivity was 62.7%, specificity 60.7% and positive predictive value approximately 31.5%–42.9%. Performance varied by age, race, sex and neighborhood deprivation, while logistic regression performed comparably to the evaluated machine-learning model.

Peer-reviewed — FedEC-Dia combined client-specific probability calibration with regularization toward a shared explanation prototype. Simulated federated experiments using PIMA and MIMIC-III reported improved weak-client prediction, calibration and attribution stability under multiple forms of distribution shift.

Peer-reviewed — Fedflow provided Python-based orchestration of FeatureCloud clients across virtual machines and cloud instances, with Snakemake integration for reproducible workflows. Its evidence supports research orchestration rather than clinical effectiveness.

Preprint — Adaptive Bayesian Partner Selection matched FedDyn at an AUROC of 0.758 across 230 simulated MIMIC-IV centers while using approximately 9% of FedAvg’s communication cost.

Preprint — A cloud-native wearable platform mapped streaming signals to FHIR and used microservices, stream processing and a medallion lakehouse architecture. It sustained 50 complete ingestion requests per second with median latency below eight milliseconds, but no clinical outcome was evaluated.

Limitations

MEDAL provides data locality but no reported differential-privacy budget, secure aggregation, leakage attack evaluation or malicious-participant defense.

The gradient-reconstruction study used an active malicious server, short sequences and public datasets; its exact leakage rates should not be generalized to every federation.

The hypertension evaluation measured retrospective case finding rather than improved screening, treatment initiation or cardiovascular outcomes.

FedEC-Dia used simulated institutions, narrow datasets and post-hoc attributions that do not establish causal or clinically correct explanations.

The FHIR wearable architecture reported throughput rather than reliability across hospitals, schema changes, outages, regulatory audits or chronic-disease interventions.

No consequential SDoH study combined individual-level social data, multisite federation, external validation and preventive-care outcomes.

What’s worth reading or building on

Read

1. Independent hypertension case-finding evaluation — strongest real-world preventive-care and subgroup-performance evidence.

2. Clinical-text reconstruction study — clearest evidence that federated learning alone does not guarantee confidentiality.

3. MEDAL — strongest operational multicenter clinical-language-model architecture in this edition.

4. FedEC-Dia — useful framework for combining calibration and explanation stability under non-IID data.

5. FHIR-native wearable platform — valuable infrastructure design despite the absence of clinical validation.

6. Adaptive Bayesian Partner Selection — promising approach for reducing communication in large federations.

7. Fedflow — immediately reusable orchestration software for reproducible federated experiments.

Build

FHIR R4 and US Core ingestion pipelines that preserve terminology versions, provenance and data-quality measurements.

Secure aggregation combined with record-level differential privacy and empirical reconstruction testing.

Site-specific calibration and thresholds rather than one global decision threshold.

Transparent baseline models tested alongside complex temporal and foundation models.

Explanation regression tests covering global, patient-level and subgroup-level attribution changes.

Prospective hypertension or diabetes studies measuring screening yield, false-positive burden, treatment initiation and disease control.

SDoH evaluation using individual-level measures rather than relying only on area-level deprivation proxies.

Sources