SynapsePath Federated Transfer Learning Protocol: Letting Small Hospitals Reuse Diagnostic Models Trained at Large Hospitals

SynapsePath Federated Transfer Learning Protocol: Letting Small Hospitals Reuse Diagnostic Models Trained at Large Hospitals

A joint team from ETH Zurich and the University Hospital of Zurich has published the full technical specification of the federated transfer learning protocol SynapsePath in Nature Medicine. The protocol allows medical institutions of different sizes to jointly train diagnostic AI models without directly sharing patient data.

SynapsePath's core idea is to treat a large hospital's pre-trained diagnostic model as a "teacher network," then have smaller hospitals fine-tune sub-modules of that network on their local data. Throughout the process, raw patient data stays at the source site; only model parameter gradients are encrypted and uploaded to a coordinator. The coordinator runs secure aggregation over gradients from multiple hospitals and then distributes the updated global model.

Dr. Anya Volkov, who leads the AI Lab at University Hospital Zurich and heads the team, said conventional federated learning requires all participants' data distributions to be statistically close — an assumption that almost never holds in healthcare. Patient populations at tertiary hospitals and community clinics differ sharply in age, disease spectrum, and imaging equipment, and naive aggregation often hurts smaller sites' model performance.

SynapsePath introduces a layered transfer mechanism. The teacher network is split into three tiers: a low-level feature extractor, a mid-level pathology analyzer, and a top-level diagnostic decision module. Smaller hospitals fine-tune only the top-level decision module; the middle and lower tiers remain frozen. The design is inspired by cognitive science's "domain-general features" theory — visual features pre-trained on sufficiently large datasets transfer well across medical scenarios.

The controlled experiment reported in the paper covered 11 hospitals — 3 tertiary, 5 regional, and 3 community. The task was early lung-cancer screening on CT images. Before the experiment, the three community hospitals' standalone models reached only 73 percent accuracy. After applying SynapsePath, those models climbed to 89 percent — close to the 92 percent achieved by tertiary hospitals — with no privacy leakage observed.

On data security, SynapsePath uses a combination of homomorphic encryption and differential privacy. Model gradients are encrypted with the Paillier scheme before upload, and the coordinator runs aggregation entirely in ciphertext. Calibrated Laplace noise is added to each gradient so no single record can be reconstructed. Auditor TÜV SÜD ran a nine-month penetration test on the protocol and found no exploitable privacy vulnerabilities.

On the business model, the team established a non-profit, the SynapsePath Foundation, to maintain the protocol. Any medical institution can use the protocol free of charge, provided it anonymizes and contributes statistics on its local model performance to the foundation for protocol improvement. The foundation's board includes the head of WHO's Digital Health department, a former member of the European Data Protection Board, and representatives from two medical-AI startups.

Anya Volkov said SynapsePath's next step is to partner with medical institutions in low-income countries in Africa and Southeast Asia. In those regions, community hospitals often have only a few hundred labeled images, and without SynapsePath it has been almost impossible to train models of clinical value. The World Health Organization has said it will provide infrastructure support for the project.

Critics note that SynapsePath still cannot fully prevent model-inversion attacks. When an attacker controls both the coordinator and some participants, it may still be possible to infer features of other participants' data. The authors acknowledge the risk in the supplementary material and propose a gradient-hash-based watermarking mechanism as mitigation, while conceding that its effectiveness needs further verification.