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

This study addresses the challenge of integrating heterogeneous data—longitudinal tumor measurements, informative dropout (time-to-event), and high-dimensional genomic covariates—into a single population modeling framework. Traditional nonlinear mixed-effects (NLME) models often struggle with the computational complexity and model development effort required for such high-dimensional, non-linear data. The authors extend the Empirical Bayes Variational Autoencoder (EB-VAE) framework to handle these multimodal inputs. The approach uses an encoder to infer individual-specific latent effects and a decoder to map these effects to tumor trajectories, augmented with a hazard model to account for informative dropout. The authors compare fully neural decoders with hybrid semi-mechanistic decoders and introduce a two-stage training process to incorporate genomic covariates without destabilizing the model.

The hybrid decoder successfully recovered treatment-effect parameters that were consistent with established literature values while maintaining predictive performance comparable to the fully neural decoder. The joint model effectively captured both tumor-volume distributions and dropout patterns in held-out individuals. Furthermore, conditioning the prior on genomic covariates improved individual-level predictions across both cutaneous melanoma and breast cancer datasets. Stability selection identified biologically plausible genetic indicators, such as alterations in BRAF, NRAS, NF1, and MDM2, demonstrating the framework's utility for hypothesis generation in pharmacometrics.

This research provides a flexible, modular, and scalable alternative to traditional pharmacometric modeling. By combining the interpretability of mechanistic models with the flexibility of neural dynamics and the power of high-dimensional covariate integration, the EB-VAE framework enables more accurate predictions of treatment response and disease progression. This approach is particularly valuable for drug development, where understanding how genetic profiles influence treatment outcomes can inform precision medicine strategies and improve the design of clinical trials.