When to Use Compute-to-Data vs Federated Learning
Use this decision framework to choose between compute-to-data and federated learning for your clinical AI project.
Decision flowchart
Does any single institution have enough data for independent training?
YES → Use compute-to-data (Rapha Protocol). You don't need FL's coordination overhead.
NO → Proceed to question 2.
Does your regulatory framework allow gradient transmission?
NO (HIPAA-strict, NHS DSPT, GDPR with health data) → Use compute-to-data. Transmitting model gradients of patient data is likely regulated the same as transmitting the data itself. Avoid FL.
YES (non-regulated research, anonymised data) → FL may be acceptable.
Can you deploy edge hardware at each site?
YES → Deploy Rapha edge appliances. Train independently at each site. No coordination. No gradient transmission. Best of both worlds: multi-site data access with single-site security.
NO → FL may be your only option. Accept the gradient leakage risk, coordination overhead, and non-IID convergence challenges.
The hybrid model: Rapha + FL when you truly need both
You can deploy Rapha edge appliances at multiple hospitals and train a model locally at each. Then aggregate the trained weights (not gradients) using federated averaging. This gives you multi-site scale AND hardware-enforced security. Rapha handles the training. You handle the weight aggregation. No gradient transmission means no gradient leakage.