Tracking how a tumor responds to treatment means juggling three kinds of data that don't naturally sit together: repeated size measurements over time, the fact that patients drop out (often for reasons tied to how they're doing), and their genetics. This paper extends an empirical-Bayes variational autoencoder to model all three at once.

Each patient gets latent individual effects pulled toward a prior shaped by their covariates, which a decoder turns into a tumor-volume trajectory. A hazard model is bolted on so informative dropout is handled rather than ignored, and genetics enter through a prior that adapts to a patient's genomic profile. A semi-mechanistic decoder recovered treatment-effect parameters close to established mixed-effects estimates, and genetic conditioning sharpened individual predictions in melanoma and breast cancer, flagging plausible markers like BRAF, NRAS, NF1, and MDM2.

This is a pharmacometrics modeling paper, so read it for the assumptions and how far the held-out results really generalize.