EB-VAE for Joint Tumor Growth and Dropout Modeling

Abstract-only summary of an EB-VAE extension for joint tumor growth, dropout hazards, semi-mechanistic decoders, and genomic priors.

TL;DR — The abstract describes an empirical Bayes variational autoencoder extension for jointly modeling tumor-volume trajectories and time to dropout. The framework uses latent individual effects regularized by a covariate-conditioned empirical Bayes prior, augments the decoder with a hazard model for informative dropout, compares fully neural and hybrid semi-mechanistic decoders, and reports improved individual-level prior predictions from genetic conditioning in cutaneous melanoma and breast cancer experiments.

Problem and core idea

The abstract frames the problem as the integration of three complementary sources of oncology information: longitudinal tumor measurements, dropout information, and genetic covariates. It states that combining these sources within a single population modeling framework remains challenging.

The proposed solution is to extend empirical Bayes variational autoencoders to joint longitudinal and time-to-event modeling. The framework is evaluated on tumor growth data. It uses latent individual effects to represent inter-individual variability, regularizes those effects with a covariate-conditioned empirical Bayes prior, and decodes them into tumor-volume trajectories.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.