Multimodal Empirical Bayes VAE for joint longitudinal and time-to-event modeling

The authors extend the empirical Bayes variational autoencoder (EB-VAE) framework to jointly model longitudinal tumor measurements and time-to-event dropout data. The approach uses a covariate-conditioned prior to represent inter-individual variability and augments the decoder with a hazard model to account for informative dropout.

• Hybrid semi-mechanistic decoders recovered treatment-effect parameters consistent with nonlinear mixed-effects estimates while maintaining neural performance.
• Genetic conditioning improved individual-level prior predictions in cutaneous melanoma and breast cancer experiments.
• Stability selection identified biologically plausible genetic indicators, including alterations in BRAF, NRAS, NF1, and MDM2.

This work demonstrates that EB-VAE provides a flexible probabilistic framework for combining neural dynamics, mechanistic structure, and high-dimensional covariates in pharmacometric applications.