Back at CHASE, with a harder question

IEEE/ACM CHASE 2023 was the conference where I presented my first PhD paper. This year I was back at CHASE, in Pittsburgh, with a harder version of a similar question.

The 2023 paper asked whether where a patient lives changes what we can predict about how their disease will go. The 2026 paper takes on the bigger version of that. How do we help a doctor predict how a patient is likely to do, using everything the hospitals and treatment centers around them have already learned, without a single patient record ever leaving the building it was collected in?

The usual answer is to wire every hospital into one central server and treat every contribution as equally valuable. That is convenient, and it is not how institutions actually relate to each other. A few sites hold data that genuinely sharpens your prediction. Most are neutral. And some will quietly make it worse.

So we stopped treating the network between those hospitals as something fixed. In "Evolving Sparse P2P Networks for Federated Survival Prediction with Adaptive Utility," the peer to peer graph is sparse and it changes over time. Every connection has to keep earning its place, and the ones that stop being useful drop away.

Thank you to my co-authors, Satyaki Roy and Sajal K. Das, and to the people in the Trustworthy AI, Explainability and Privacy session who pushed on the parts that needed pushing.

The work sits close to something I care about well outside the lab. Predictions built on the patient in front of you, not on an average of everyone else.