Federated learning for sensitive health data
Training models without centralising patient data. A primer on horizontal vs vertical federation and what it would take to make federated training viable for Indian diabetes cohorts.
Centralising patient data raises privacy, regulatory, and trust problems. Federated learning trains a shared model while data stays where it was generated.
Horizontal vs vertical
Horizontal federation: multiple hospitals with similar data on different patients train a single model. Vertical: different institutions hold different features about the same patients.
Indian context
Hospitals, glucometers, and CGMs already generate the right substrate. The hard problems are common protocols, secure aggregation infrastructure, and incentives for participation.
DiaCare position
We do not currently train on user data — opt-in or otherwise. If federated infrastructure matures, it becomes a plausible path to improve risk models without compromising the local-first architecture.