I’ve been working on a project called Biological JEPA, exploring how JEPA-style representation learning could be combined with biological constraints to model disease progression over time.
I tested the approach on real ADNI data from 2,300+ patients and compared it with several other models.
The main idea is to move beyond predicting a patient’s current state and instead learn representations that capture how the disease changes over time.
The project is open source:
I’d really appreciate any feedback, criticism, or ideas for improving the approach. I’m especially interested in feedback on the biological constraints, experimental setup, and whether the results are convincing.
Fascinating direction — moving from “current state” to “how the disease changes over time” is exactly the right question to ask.
The experiment I’d want to see first is the ablation: run JEPA without the biological constraints and compare. If the unconstrained model captures the same progression structure on ADNI, the constraints are decoration; if the representation collapses or drifts without them, they’re doing real work. That’s the single cheapest test that tells you whether you need them.
And on the 2,300+ patients — is the temporal signal true longitudinal follow-up per patient, or time reconstructed from cross-sectional data? If the latter, my first move would be a held-out longitudinal fold, because cross-sectional ordering can manufacture beautiful trajectories that don’t survive contact with real follow-ups. ADNI has the longitudinal data to test it.
Great to see it open. Happy to dig into the constraint design with you.