HARMONY: reputation for federated survival analysis

There is an awkward problem sitting inside federated learning for hospitals. If you want to know whether another institution's contribution is actually helping your model, you have to look at what they sent you. That is a reasonable thing to want. It is also the exact thing federated learning exists to avoid.

Our new paper in IEEE Journal of Biomedical and Health Informatics is about living with that tension instead of ignoring it. "HARMONY: Hybrid Aggregation and Reputation Mechanism for Cooperative Federated Survival Analysis".

Hospitals do not join a federation as equals. They differ in how complete their records are, how many patients are lost to follow-up, who their patients are, and how carefully the data was collected in the first place. Averaging everything together assumes those differences do not matter. They do.

In HARMONY, sites evaluate each other and build a reputation that moves over time, based on whether a peer's update actually improved local survival prediction. The part I keep coming back to is the separation. Peers only ever see privacy-protected versions of each other's updates, enough to judge whether something helped and not enough to learn anything about a patient. The server still aggregates the real ones.

On SEER (the Surveillance, Epidemiology, and End Results Program of the National Cancer Institute) data across ten states it holds. When a fifth of the clients are adversarial, it stays roughly a third steadier round to round than plain averaging, which is the moment reputation is supposed to earn its keep.

Thank you to my co-author, Satyaki Roy, and my advisor, Sajal K. Das.

The question underneath all of this stays the same one. Whose experience should count when a doctor is predicting what happens to the person in front of them, and how much.