We built a federated learning platform that trains diagnostic AI inside each hospital's own firewall. Hospitals share model updates, not patient data. Diagnostic accuracy rose 40%, with zero data leaving any site.
Industry: HealthcareService: AI & Automation
Overview
Client
BioMetrics Solutions
Multi-centre clinical research network
The challenge
The client wanted AI that could locate tumours early, but the data it needed sat in separate regional hospital systems. Standard machine learning would have meant copying everything into one cloud store. That carried major HIPAA legal risk, and the hospitals refused on data-control grounds.
Models trained at single sites on small datasets were too weak for clinical use. The network needed a way to learn from every site's data without moving it.
Hospital legal teams blocked data from leaving their sites over patient data liability, which halted diagnostic AI work for 14 months.
Separate models trained on small datasets were not accurate enough for clinical use, wasting millions in R&D spend.
Centralised approaches lacked the detailed audit trail needed for patient consent across several organisations.
The solution
Cloud-based training pods run inside each hospital's private firewall, so patient data never leaves the site.
Only model weights, the patterns a model has learned, are collected to update the global model. Patient data stays where it is.
An encrypted audit trail of every training round proves that no raw patient health information went into the global model updates.
Key features
Diagnostic models train inside each hospital's firewall, and only model weights are combined into one shared model through federated averaging.
Business impact
Zero patient data left hospitals
Differential privacy protects patient identities, and an encrypted audit trail records every training round.
Business impact
Diagnostic accuracy up from 62% to 87%
Mathematical noise is added to weight updates so patient identities cannot be worked out from them.
Business impact
Zero re-identification risk
Pods at each hospital pre-process local data and pick only high-quality samples for training.
Business impact
3.5x faster convergence
Remote nodes recover on their own, so training carries on through hospital network interruptions.
Business impact
24/7 training uptime
Before and after
Before
High-risk data transfer
Moving patient records to a central cloud created a large security target and HIPAA liability.
After
Local control
Data stays where it was created. Only model weights move between sites, which removes the patient data risk.
Before
Consent delays
Hosting data with a third party meant a complex process of asking patients to consent again.
After
Privacy by design
AI developers never see the raw data, which keeps the work aligned with GDPR and HIPAA.
Before
Weak single-site models
Models trained on narrow regional data did not generalise across different patient groups.
After
Learning from every site
Diagnostic accuracy rose 40% by learning from varied datasets at all sites at once.
Results
40% Accuracy Gain
The 40% gain is relative to the earlier model. Learning from every site made the model dependable across all major ethnic and age groups.
Before: 62%After: 87%
100% Risk Reduction
No raw patient health information crossed any hospital firewall during training.
Before: High RiskAfter: Zero
18x Faster Sync
Automated edge processing cut the time needed to sync local updates with the global model.
Before: 72h SyncAfter: 4h Sync
Testimonial
“We needed shared diagnostic models without sharing patient data. Coretus set up training inside each hospital's firewall, and accuracy rose from 62% to 87% with zero patient data leaving any site. Training sync is also 18x faster.”
Key results
Trust and control
Raw patient health information never leaves the hospital. Differential privacy stops patient identities being reverse-engineered from the model.
HIPAA-GATEDEach hospital keeps 100% control over training on its own node and can disconnect from the platform at any time.
VERIFIABLE DATAThe platform runs on AWS, Azure or on-premise hardware, with no dependence on one vendor.
READY TO GROWThe client owns 100% of the federated system design and the global diagnostic model it produced.
100% OWNEDFaster delivery
Reusing tested components saved about 12 weeks of setup work, which went into the privacy controls and the diagnostic model.
Our tested data vault kept each hospital's training data locked inside its own firewall.
A reusable federated platform sent the global model out to each site and collected the weight updates back.
An existing monitoring platform tracked training rounds and model accuracy at every site.
Pre-built cost controls kept compute spend on the edge training pods within budget.
Training happens inside each hospital's own firewall, on that hospital's data. Only the model weights, the patterns the model has learned, are sent out and combined into one global model. Differential privacy adds noise to those updates, so patient identities cannot be worked out from them.
Diagnostic accuracy rose by 40%, from 62% to 87%, because the model learned from varied data at every site instead of one narrow regional dataset. It now performs well across all major ethnic and age groups. Syncing local updates with the global model also fell from 72 hours to 4.
Raw patient health information never leaves hospital custody, so no patient data has to be moved or hosted elsewhere. An encrypted audit trail records every training round and shows that no raw data went into model updates. Each hospital keeps full control of its node and can disconnect at any time.
If privacy rules are blocking your diagnostic AI, we can help you train across sites while each hospital keeps its own data.