Sjöberg et al. use EB-VAE to model tumor growth and dropout with genetics

A team of researchers led by Anders Sjöberg at the Fraunhofer-Chalmers Centre has introduced a new AI framework that combines tumor growth trajectories, dropout timing, and genetic data in a single probabilistic model. Their work, published on arXiv, extends the empirical Bayes variational autoencoder (EB-VAE) to handle both longitudinal measurements (like tumor volume changes over time) and time-to-event outcomes (such as patient dropout or progression). The model uses a decoder to map latent individual effects to tumor-volume trajectories, and augments it with a hazard model for dropout prediction. A key innovation is the covariate-conditioned empirical Bayes prior, which regularizes latent variables using genetic information. The team compared a fully neural decoder with a hybrid semi-mechanistic decoder; the hybrid version recovered treatment-effect parameters consistent with traditional nonlinear mixed-effects models while maintaining predictive performance. In experiments on cutaneous melanoma and breast cancer datasets, the joint model accurately reproduced both tumor-volume distributions and dropout patterns in held-out patients. Genetic conditioning improved individual-level predictions, and stability selection identified several biologically relevant genetic markers, including BRAF, NRAS, NF1, and MDM2—genes known to play roles in cancer progression and treatment resistance. This work demonstrates that EB-VAE provides a flexible, scalable framework for integrating neural dynamics, mechanistic structure, time-to-event modeling, and high-dimensional covariates, paving the way for more personalized and data-driven pharmacometric analyses.