Healthcare Deployments

Deployment evidence

Healthcare PET claims often mix research studies, clinical pilots, and vendor platform stories. Do not assume clinical production use unless the source clearly says so.

Published Studies And Pilots

EXAM federated learning for COVID-19 oxygen prediction

Field	Entry
Organization / project	EXAM, a federated learning study across 20 institutions
Domain	Healthcare
Problem	Train a model to predict future oxygen requirements for symptomatic COVID-19 patients using EMR, labs, vital signs, and chest X-rays without centralizing all institutional data.
PETs used	Federated learning
Deployment maturity	Research prototype / demonstration experiment
Source quality	Peer-reviewed / academic
What worked	The study demonstrated multi-institution FL across 20 institutions and reported performance across sites.
Challenges	Research success does not prove clinical workflow integration, privacy from updates, poisoning resistance, or ongoing operations.
Lessons for builders	FL can coordinate real healthcare institutions, but clinical utility, local data quality, update leakage, and governance must be evaluated separately.
Source	Nature Medicine: Federated learning for predicting clinical outcomes in patients with COVID-19

(Evidence: Literature-backed. Source quality: Peer-reviewed / academic. Reviewed 2026-06-17 — the paper reports a 20-institution study and AUC results, but it is evidence for research feasibility rather than routine clinical production.)

MELLODDY cross-pharma federated learning

Field	Entry
Organization / project	MELLODDY consortium
Domain	Healthcare / life sciences
Problem	Let pharmaceutical companies improve QSAR models using sensitive proprietary chemical and bioactivity data without pooling data.
PETs used	Federated learning, privacy-preserving platform controls
Deployment maturity	Completed consortium pilot / demonstration experiment
Source quality	Primary / official plus peer-reviewed / academic
What worked	The project ran a large cross-company FL experiment and reported aggregated model improvements for participating companies.
Challenges	The result is strong evidence for cross-company feasibility, not a generic proof that FL protects all proprietary information or works for every discovery task.
Lessons for builders	Cross-silo FL needs shared task definitions, platform trust, data harmonization, and agreement about what model improvements can reveal.
Source	MELLODDY paper in Journal of Chemical Information and Modeling

(Evidence: Deployment-backed / literature-backed. Source quality: Primary / official plus peer-reviewed / academic. Reviewed 2026-06-17 — MELLODDY is strong feasibility evidence for cross-pharma FL; ongoing production after the consortium remains unverified.)

Common Proposed Use Cases

Use case	Candidate PETs	Why proposed	Caveats
Hospital quality metrics without sharing patient records	Federated analytics, MPC, DP	Hospitals can compute comparable metrics locally	Small cohorts and inconsistent coding can leak or mislead
Multi-hospital model training	FL, secure aggregation, DP	Data cannot centralize across institutions	FL alone does not provide privacy; non-IID data can hurt smaller sites
Private clinical trial matching	PSI, TEEs, clean rooms	Match patients to eligibility criteria without broad data exposure	Match output may be sensitive; governance matters
Synthetic patient datasets for research	DP synthetic data	Researchers need data-like artifacts	Synthetic data can memorize or lose rare-case utility

Lessons Learned

• Healthcare deployments need clinical validation, not only privacy-preserving computation.
• The protected artifact must be named: records, updates, identifiers, model weights, outputs, logs, or all of them.
• Cross-institution projects often fail on data definitions before they fail on PET mechanics.
• Small cohorts, rare diseases, and site-specific patterns need special treatment.