Longitudinal Self-supervision to Disentangle Inter-patient Variability from Disease Progression

The problem of building disease progression models with longitudinal data has long been addressed parametric mixed-effect models. They provide interpretable at the cost modeling assumptions on profiles and their variability across subjects. Their deep learning counterparts, other hand, strive flexible data-driven modeling, additional interpretability - or, as far generative are involved, disentanglement latent variables respect to factors comes from constraints. In this work, we propose a model designed disentangle inter-patient an estimated timeline. We do not seek for explicit mapping between age stage, but learn latter solely ordering visits using differentiable ranking loss. Furthermore, encourage be encoded in separate space, where each patient single representation is learned its set visits, constraint invariance under permutation visits. modularity network architecture allows us apply our various types: synthetic image dataset known factors, cognitive assessments neuroimaging data. show that, combined encoder, loss helps exceed supervision, particular terms staging.