Ask any hospital CIO: "Can we use your patient data to train a cardiovascular risk prediction model?" The answer is almost always "no."

This isn't a matter of attitude — it's compliance. China's Personal Information Protection Law, Data Security Law, and Level 3 security certifications mean every row of medical data potentially contains patient privacy. Any unauthorized data transfer is illegal.

But this creates a fundamental contradiction: AI model performance depends heavily on data volume, yet medical data is one of the hardest types to aggregate.

Federated learning was born to break this contradiction.

Why Medical Data Is Inherently Fragmented and Hard to Share

China has over 36,000 medical institutions, each accumulating patient data in their own HIS systems — scattered across different databases, formats, and standards.

Even setting aside regulatory issues, data sharing faces enormous barriers: hospitals compete with each other and treat data as a core asset; data formats vary wildly; transmission security can't be guaranteed; patients may never have consented to external use of their data.

The traditional workaround — "anonymize data before sharing" — often strips clinical value from the data, and real re-identification risks remain.

Federated Learning's Core Insight: Move the Model, Not the Data

Federated learning proposes a fundamental shift: instead of bringing data to the model, bring the model to the data.

The process works like this: an initialized global model is distributed to participating hospitals. Each hospital independently trains the model locally using its own patient data. After training, hospitals upload only the model parameter updates (gradients) — never any raw data. The central server aggregates parameter updates from all hospitals and updates the global model. The updated global model is redistributed for the next round.

Patient data never leaves the hospital's servers.

ReHealth AI's Implementation: Three Key Design Choices

1. Medical-Grade Bias Correction

Federated learning faces a classic problem: data distributions vary significantly across institutions. Tier-3 hospital patients differ substantially from community clinic patients in age distribution, disease severity, and medication patterns. Naively aggregating gradients can produce a model that performs well at some hospitals but fails at others.

ReHealth AI's aggregation algorithm includes built-in bias correction: when aggregating gradient updates, distribution differences are weighted and corrected to ensure the final model reaches clinically acceptable performance ranges across different types of medical institutions.

2. Differential Privacy Protection

Gradients aren't raw data, but theoretically, some patient information could be reverse-engineered from them. To eliminate this risk entirely, we add calibrated random noise (differential privacy) to gradients before upload. This makes gradient reversal mathematically infeasible without significantly impacting model performance.

3. Asynchronous Aggregation

Hospitals have vastly different compute resources. Tier-3 hospitals may have dedicated GPU servers; community clinics may only have basic PCs. Traditional synchronous federated learning requires all participants to complete a training round simultaneously before aggregation — nearly impossible in hospital environments.

ReHealth AI supports asynchronous aggregation: hospitals train and upload gradients at their own pace, and the central server automatically triggers aggregation once sufficient updates are received, without waiting for all hospitals to synchronize.

Real Results: Data Volume vs. Model Performance

Federated Learning Isn't a Silver Bullet

Honestly, federated learning has real limitations. Communication overhead is genuine — each aggregation round requires model parameter transfers between hospitals and the central server, placing real demands on network bandwidth. Data quality differences are hard to resolve within the federated framework; hospitals with poor data labeling drag down overall model quality. Federated learning is far more complex to debug and optimize than centralized training, with a higher engineering bar.

This is why production-grade federated learning in healthcare requires extensive engineering optimization specifically for medical environments — you can't just transplant academic paper frameworks directly.

Next: Federated Learning + Causal Attribution

Federated learning solves "how to train better models without sharing data." But in preventive medicine, having a good prediction model isn't enough — you need to prove that "this intervention actually reduced risk." That's what causal attribution analysis addresses.

In our next article, we'll examine how Propensity Score Matching becomes the key evidence for preventive medicine settlement.