Breaking Barriers, Not Privacy: Real-World Split Learning across Healthcare Systems
summary
Three independent HIEs jointly train a deep model using split learning (no raw data leaves sites) and report successful training on five comorbidities including myocardial infarction. Implementation uses FHIR ingestion, VSAC value sets, regex plus a DI transformer to scrub notes and produce 768‑dim clinical BERT embeddings, a nopeek split from MIT with distance‑based decorrelation, TLS1.3 and RAM‑certified infra, websocket RPC sharding across GPUs and OTP registration. Network traffic averages 2 KB per minute batches, meeting MSA privacy limits and yielding robust preliminary performance.
AI-written summary · about
chapters (10)
- 0:00Healthcare data challenges
- 1:12Agenda and problem
- 2:38Interoperability and governance
- 4:06Privacy with split learning
- 6:55Split learning vs alternatives
- 8:41Privacy decisions and security
- 11:52Collaboration challenges
- 12:56Data quality without visibility
- 14:20Implementation hurdles and solutions
- 15:45Results and future
related talks
45mManaged federated learning on AWS: A case study for healthcare
27mCreate Value and Insight Through Privacy-preserving Data Collaboration [...]
26mMachine Learning without the Data #federatedlearning
18mEnabling Medical AI through Federated Learning
01h 12mModelling Privacy Threats of Medical Data and Health Data Spaces
34mTraining without sharing: Is Federated Learning the answer to data privacy in AI?
40mDemystifying Decentralized Fair Models for Federated Machine Learning
41mAccelerating privacy-enhancing data processing