Multimodal Empirical Bayes Variational Autoencoders for Joint Longitudinal and Time-to-Event Modeling

TL;DR — Extends the empirical Bayes variational autoencoder (EB-VAE) into a probabilistic framework for jointly modeling longitudinal tumor trajectories and time-to-event (dropout), integrating genetic covariates for pharmacometric applications. It matters as a flexible method for combining neural dynamics, mechanistic structure, and survival modeling in population models.

• Represents inter-individual variability via latent effects regularized by a covariate-conditioned empirical Bayes prior, with a decoder mapping latents to tumor-volume trajectories; augmented with a hazard model to handle informative dropout.
• Compares fully neural vs. hybrid semi-mechanistic decoders — the hybrid recovers treatment-effect parameters consistent with prior nonlinear mixed-effects estimates while matching neural predictive performance.
• Genetics-conditioned prior adaptation improved individual-level prior predictions in cutaneous melanoma and breast cancer experiments; stability selection flagged biologically plausible markers (BRAF, NRAS, NF1, MDM2).
• Note: "multimodal" here means multiple clinical/genomic data sources, not cross-modal vision/audio — hence classed as a methods contribution rather than generative multimodal.