Year: 2026
Type: preprint
Abstract: Artificial intelligence in medicine depends on health data that patients are often reluctant to surrender. Federated learning trains models without centralising raw records, yet existing systems still assume each site holds hundreds or thousands of samples, a scale no single patient can reach. Here, we show that collective learning remains practical when each participant contributes just a single medical record that never leaves their own device. Our framework, PrivateBoost, combines gradient boosting with secret sharing: each device splits its contribution into cryptographic shares held by independent institutions, so that only population-level sums, never individual records, are ever reconstructed. Across four datasets with up to 70,692 participants, PrivateBoost matches the accuracy of centralised training, tolerates the loss of 80% of participants in any round, and needs at most a few megabytes of communication per patient. A pilot deployment on four consumer smartphones trains models unattended. Every consenting patient can now help improve medical artificial intelligence without surrendering any raw data.
Source: medRxiv
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7128740076